from local

egs/librispeech/ASR/pruned_transducer_stateless_d2v/asr_datamodule.py

@@ -0,0 +1,454 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2022  Xiaomi Corporation     (Author: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SingleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import (  # noqa F401 For AudioSamples
+    AudioSamples,
+    OnTheFlyFeatures,
+)
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class LibriSpeechAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--full-libri",
+            type=str2bool,
+            default=True,
+            help="When enabled, use 960h LibriSpeech. Otherwise, use 100h subset.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            logging.info("About to get Musan cuts")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            input_strategy=eval(self.args.input_strategy)(),
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SingleCutSampler.")
+            train_sampler = SingleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                input_strategy=eval(self.args.input_strategy)(),
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else eval(self.args.input_strategy)(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_clean_100_cuts(self) -> CutSet:
+        logging.info("About to get train-clean-100 cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_train-clean-100.jsonl.gz"
+        )
+
+    @lru_cache()
+    def train_clean_360_cuts(self) -> CutSet:
+        logging.info("About to get train-clean-360 cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_train-clean-360.jsonl.gz"
+        )
+
+    @lru_cache()
+    def train_other_500_cuts(self) -> CutSet:
+        logging.info("About to get train-other-500 cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_train-other-500.jsonl.gz"
+        )
+
+    @lru_cache()
+    def train_all_shuf_cuts(self) -> CutSet:
+        logging.info(
+            "About to get the shuffled train-clean-100, \
+            train-clean-360 and train-other-500 cuts"
+        )
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_train-all-shuf.jsonl.gz"
+        )
+
+    @lru_cache()
+    def dev_clean_cuts(self) -> CutSet:
+        logging.info("About to get dev-clean cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_dev-clean.jsonl.gz"
+        )
+
+    @lru_cache()
+    def dev_other_cuts(self) -> CutSet:
+        logging.info("About to get dev-other cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_dev-other.jsonl.gz"
+        )
+
+    @lru_cache()
+    def test_clean_cuts(self) -> CutSet:
+        logging.info("About to get test-clean cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_test-clean.jsonl.gz"
+        )
+
+    @lru_cache()
+    def test_other_cuts(self) -> CutSet:
+        logging.info("About to get test-other cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_test-other.jsonl.gz"
+        )

egs/librispeech/ASR/pruned_transducer_stateless_d2v/checkpoint.py

@@ -0,0 +1,225 @@
+# Copyright  2021-2022  Xiaomi Corporation  (authors: Fangjun Kuang,
+#                                                     Zengwei Yao)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import glob
+import logging
+import os
+import re
+from pathlib import Path
+from typing import Any, Dict, List, Optional, Union
+
+from lhotse.dataset.sampling.base import CutSampler
+
+import torch
+import torch.nn as nn
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.optim import Optimizer
+
+# use duck typing for LRScheduler since we have different possibilities, see
+# our class LRScheduler.
+LRSchedulerType = object
+
+
+def save_checkpoint(
+    filename: Path,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    params: Optional[Dict[str, Any]] = None,
+    optimizer = None,
+    scheduler = None,
+    scaler: Optional[GradScaler] = None,
+    sampler: Optional[CutSampler] = None,
+    rank: int = 0,
+) -> None:
+    """Save training information to a file.
+
+    Args:
+      filename:
+        The checkpoint filename.
+      model:
+        The model to be saved. We only save its `state_dict()`.
+      model_avg:
+        The stored model averaged from the start of training.
+      params:
+        User defined parameters, e.g., epoch, loss.
+      optimizer:
+        The optimizer to be saved. We only save its `state_dict()`.
+      scheduler:
+        The scheduler to be saved. We only save its `state_dict()`.
+      scalar:
+        The GradScaler to be saved. We only save its `state_dict()`.
+      rank:
+        Used in DDP. We save checkpoint only for the node whose rank is 0.
+    Returns:
+      Return None.
+    """
+    if rank != 0:
+        return
+
+    logging.info(f"Saving checkpoint to {filename}")
+
+    if isinstance(model, DDP):
+        model = model.module
+    
+    if type(optimizer) == list:
+        checkpoint = {
+            "model": model.state_dict(),
+            "optimizer_enc": optimizer[0].state_dict() if optimizer is not None else None,
+            "optimizer_dec": optimizer[1].state_dict() if optimizer is not None else None,
+            "scheduler_enc": scheduler[0].state_dict() if scheduler is not None else None,
+            "scheduler_dec": scheduler[1].state_dict() if scheduler is not None else None,
+            "grad_scaler": scaler.state_dict() if scaler is not None else None,
+            "sampler": sampler.state_dict() if sampler is not None else None,
+        }
+    else:
+        checkpoint = {
+            "model": model.state_dict(),
+            "optimizer": optimizer.state_dict() if optimizer is not None else None,
+            "scheduler": scheduler.state_dict() if scheduler is not None else None,
+            "grad_scaler": scaler.state_dict() if scaler is not None else None,
+            "sampler": sampler.state_dict() if sampler is not None else None,
+        }
+
+
+    if model_avg is not None:
+        checkpoint["model_avg"] = model_avg.to(torch.float32).state_dict()
+
+    if params:
+        for k, v in params.items():
+            assert k not in checkpoint
+            checkpoint[k] = v
+
+    torch.save(checkpoint, filename)
+
+
+def load_checkpoint(
+    filename: Path,
+    model: nn.Module,
+    model_avg: Optional[nn.Module] = None,
+    optimizer = None,
+    scheduler = None,
+    scaler: Optional[GradScaler] = None,
+    sampler: Optional[CutSampler] = None,
+    strict: bool = True,
+) -> Dict[str, Any]:
+    """
+    TODO: document it
+    """
+    logging.info(f"Loading checkpoint from {filename}")
+    checkpoint = torch.load(filename, map_location="cpu")
+
+    if next(iter(checkpoint["model"])).startswith("module."):
+        logging.info("Loading checkpoint saved by DDP")
+
+        dst_state_dict = model.state_dict()
+        src_state_dict = checkpoint["model"]
+        for key in dst_state_dict.keys():
+            src_key = "{}.{}".format("module", key)
+            dst_state_dict[key] = src_state_dict.pop(src_key)
+        assert len(src_state_dict) == 0
+        model.load_state_dict(dst_state_dict, strict=strict)
+    else:
+        model.load_state_dict(checkpoint["model"], strict=strict)
+
+    checkpoint.pop("model")
+
+    if model_avg is not None and "model_avg" in checkpoint:
+        logging.info("Loading averaged model")
+        model_avg.load_state_dict(checkpoint["model_avg"], strict=strict)
+        checkpoint.pop("model_avg")
+
+    def load(name, obj):
+        s = checkpoint.get(name, None)
+        if obj and s:
+            obj.load_state_dict(s)
+            checkpoint.pop(name)
+    
+    if type(optimizer) == list:
+        load("optimizer_enc", optimizer[0])
+        load("optimizer_dec", optimizer[1])
+        load("scheduler_enc", scheduler[0])
+        load("scheduler_dec", scheduler[1])
+    else:
+        load("optimizer", optimizer)
+        load("scheduler", scheduler)
+
+    load("grad_scaler", scaler)
+    load("sampler", sampler)
+
+    return checkpoint
+
+
+def save_checkpoint_with_global_batch_idx(
+    out_dir: Path,
+    global_batch_idx: int,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    params: Optional[Dict[str, Any]] = None,
+    optimizer = None,
+    scheduler = None,
+    scaler: Optional[GradScaler] = None,
+    sampler: Optional[CutSampler] = None,
+    rank: int = 0,
+):
+    """Save training info after processing given number of batches.
+
+    Args:
+      out_dir:
+        The directory to save the checkpoint.
+      global_batch_idx:
+        The number of batches processed so far from the very start of the
+        training. The saved checkpoint will have the following filename:
+
+            f'out_dir / checkpoint-{global_batch_idx}.pt'
+      model:
+        The neural network model whose `state_dict` will be saved in the
+        checkpoint.
+      model_avg:
+        The stored model averaged from the start of training.
+      params:
+        A dict of training configurations to be saved.
+      optimizer:
+        The optimizer used in the training. Its `state_dict` will be saved.
+      scheduler:
+        The learning rate scheduler used in the training. Its `state_dict` will
+        be saved.
+      scaler:
+        The scaler used for mix precision training. Its `state_dict` will
+        be saved.
+      sampler:
+        The sampler used in the training dataset.
+      rank:
+        The rank ID used in DDP training of the current node. Set it to 0
+        if DDP is not used.
+    """
+    out_dir = Path(out_dir)
+    out_dir.mkdir(parents=True, exist_ok=True)
+    filename = out_dir / f"checkpoint-{global_batch_idx}.pt"
+    save_checkpoint(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        scaler=scaler,
+        sampler=sampler,
+        rank=rank,
+    )

egs/librispeech/ASR/pruned_transducer_stateless_d2v/data2vec_encoder.py

@@ -0,0 +1,199 @@
+# Copyright 2021 Xuankai Chang
+#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
+
+"""Encoder definition."""
+import contextlib
+import time
+import copy
+import math 
+import logging
+import os
+from typing import List, Optional, Tuple
+import warnings
+
+import torch
+from filelock import FileLock
+from typeguard import check_argument_types
+
+from nets_utils import make_pad_mask
+from encoder_interface import EncoderInterface
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledConv1d,
+    ScaledConv2d,
+    ScaledLinear,
+)
+from torch import Tensor, nn
+
+from icefall.utils import make_pad_mask, subsequent_chunk_mask
+
+
+class FairSeqData2VecEncoder(EncoderInterface):
+    """FairSeq Wav2Vec2 encoder module.
+
+    Args:
+        input_size: input dim
+        output_size: dimension of attention
+        w2v_url: url to Wav2Vec2.0 pretrained model
+        w2v_dir_path: directory to download the Wav2Vec2.0 pretrained model.
+        normalize_before: whether to use layer_norm before the first block
+        finetune_last_n_layers: last n layers to be finetuned in Wav2Vec2.0
+                                0 means to finetune every layer if freeze_w2v=False.
+    """
+
+    def __init__(
+        self,
+        input_size: int,
+        w2v_url: str,
+        w2v_dir_path: str = "./",
+        output_size: int = 256,
+        freeze_finetune_updates: int = 0,
+        additional_block: bool = False,
+    ):
+        assert check_argument_types()
+        super().__init__()
+
+        if w2v_url != "":
+            try:
+                import fairseq
+                from fairseq.models.wav2vec.wav2vec2 import Wav2Vec2Model
+            except Exception as e:
+                print("Error: FairSeq is not properly installed.")
+                print(
+                    "Please install FairSeq: cd ${MAIN_ROOT}/tools && make fairseq.done"
+                )
+                raise e
+
+        if os.path.exists('/home/work/workspace/models/data2vec_model/audio_base_ls.pt'):
+            self.w2v_model_path = '/home/work/workspace/models/data2vec_model/audio_base_ls.pt'
+        if os.path.exists('./models/audio_base_ls.pt'):
+            self.w2v_model_path = './models/audio_base_ls.pt'
+
+        self._output_size = output_size
+
+        models, _, _ = fairseq.checkpoint_utils.load_model_ensemble_and_task(
+            [self.w2v_model_path],
+            strict=False,
+        )
+        model = models[0]
+        model.feature_grad_mult = 0.0 ## for conv network freeze
+        #model.mask_prob = 0.3 ## for conv network freeze
+        
+        if not isinstance(model, Wav2Vec2Model):
+            try:
+                model = model.w2v_encoder.w2v_model
+            
+            except:
+                print(
+                    "using data2vec ..."
+                )
+
+        self.encoders = model
+        self.pretrained_params = copy.deepcopy(model.state_dict())
+
+        if model.cfg.encoder_embed_dim != output_size or additional_block:
+            # TODO(xkc09): try LSTM
+            self.output_layer = torch.nn.Sequential(
+                torch.nn.Linear(model.cfg.encoder_embed_dim, output_size),
+                torch.nn.LayerNorm(output_size),
+                torch.nn.GELU(),
+            )
+        else:
+            self.output_layer = None
+
+        self.freeze_finetune_updates = freeze_finetune_updates
+        self.num_updates = 0
+
+    def output_size(self) -> int:
+        return self._output_size
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        warmup = None,
+        prev_states: torch.Tensor = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        xs_pad = x
+        ilens = x_lens
+        """Forward FairSeqWav2Vec2 Encoder.
+
+        Args:
+            xs_pad: input tensor (B, L, D)
+            ilens: input length (B)
+            prev_states: Not to be used now.
+        Returns:
+            position embedded tensor and mask
+        """
+        with torch.no_grad():
+            xs_pad = torch.nn.functional.layer_norm(xs_pad, xs_pad.shape)
+
+        masks = make_pad_mask(ilens).to(xs_pad.device)
+
+        ft = (self.freeze_finetune_updates <= self.num_updates) and self.encoders.training
+        if self.num_updates <= self.freeze_finetune_updates:
+            self.num_updates += 1
+        elif ft and self.num_updates == self.freeze_finetune_updates + 1:
+            self.num_updates += 1
+            logging.info("Start fine-tuning wav2vec parameters!")
+        
+        with torch.no_grad() if not ft else contextlib.nullcontext():
+            enc_outputs = self.encoders(
+                xs_pad,
+                masks,
+                mask = ft,
+                features_only=True,
+            )
+
+        xs_pad = enc_outputs["x"]  # (B,T,C),
+        bs = xs_pad.shape[0]
+        if enc_outputs["padding_mask"] is not None:
+            masks = enc_outputs["padding_mask"]  # (B, T)
+            olens = (~masks).sum(dim=1)  # (B)
+        else:
+            olens = torch.IntTensor([xs_pad.shape[1]]).repeat(bs).to(xs_pad.device)
+
+        if self.output_layer is not None:
+            xs_pad = self.output_layer(xs_pad)
+
+        return xs_pad, olens
+
+    def reload_pretrained_parameters(self):
+        self.encoders.load_state_dict(self.pretrained_params)
+        logging.info("Pretrained Wav2Vec model parameters reloaded!")
+
+
+def download_w2v(model_url, dir_path):
+    os.makedirs(dir_path, exist_ok=True)
+
+    model_name = model_url.split("/")[-1]
+    model_path = os.path.join(dir_path, model_name)
+
+    dict_url = "https://dl.fbaipublicfiles.com/fairseq/wav2vec/dict.ltr.txt"
+    dict_path = os.path.join(dir_path, dict_url.split("/")[-1])
+
+    with FileLock(model_path + ".lock"):
+        if not os.path.exists(model_path):
+            torch.hub.download_url_to_file(model_url, model_path)
+            torch.hub.download_url_to_file(dict_url, dict_path)
+            logging.info(f"Wav2Vec model downloaded {model_path}")
+        else:
+            logging.info(f"Wav2Vec model {model_path} already exists.")
+
+    return model_path
+
+
+if __name__ == '__main__':
+    d2v = FairSeqData2VecEncoder(input_size=768, w2v_url='ww', output_size=768)
+    inputs = torch.randn([1, 211564])
+    #a = torch.ones([1000]
+    #b = torch.ones([10000])
+    #c = torch.ones([10000])
+    length = torch.tensor([211564])
+    outputs = d2v(inputs, length)
+    print(outputs[0].size())
+
+    #for n, p in d2v.named_parameters():
+    #    print(n)

egs/librispeech/ASR/pruned_transducer_stateless_d2v/decode.py

@@ -0,0 +1,989 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+(2) beam search (not recommended)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+(3) modified beam search
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+(4) fast beam search (one best)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+(5) fast beam search (nbest)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+(7) fast beam search (with LG)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(8) modified beam search with RNNLM shallow fusion (with LG)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 4 \
+    --max-contexts 4 \
+    --rnn-lm-scale 0.4 \
+    --rnn-lm-exp-dir /path/to/RNNLM/exp \
+    --rnn-lm-epoch 99 \
+    --rnn-lm-avg 1 \
+    --rnn-lm-num-layers 3 \
+    --rnn-lm-tie-weights 1
+
+
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+    modified_beam_search_rnnlm_shallow_fusion,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.rnn_lm.model import RnnLmModel
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_LG
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+          - modified_beam_search_rnnlm_shallow_fusion # for rnn lm shallow fusion
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search, fast_beam_search_LG,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG and fast_beam_search_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is fast_beam_search_LG,
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is fast_beam_search_LG,
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--simulate-streaming",
+        type=str2bool,
+        default=False,
+        help="""Whether to simulate streaming in decoding, this is a good way to
+        test a streaming model.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-scale",
+        type=float,
+        default=0.0,
+        help="""Used only when --method is modified_beam_search_rnnlm_shallow_fusion.
+        It specifies the path to RNN LM exp dir.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-exp-dir",
+        type=str,
+        default="rnn_lm/exp",
+        help="""Used only when --method is modified_beam_search_rnnlm_shallow_fusion.
+        It specifies the path to RNN LM exp dir.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-epoch",
+        type=int,
+        default=7,
+        help="""Used only when --method is modified_beam_search_rnnlm_shallow_fusion.
+        It specifies the checkpoint to use.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-avg",
+        type=int,
+        default=2,
+        help="""Used only when --method is modified_beam_search_rnnlm_shallow_fusion.
+        It specifies the number of checkpoints to average.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-embedding-dim",
+        type=int,
+        default=2048,
+        help="Embedding dim of the model",
+    )
+
+    parser.add_argument(
+        "--rnn-lm-hidden-dim",
+        type=int,
+        default=2048,
+        help="Hidden dim of the model",
+    )
+
+    parser.add_argument(
+        "--rnn-lm-num-layers",
+        type=int,
+        default=4,
+        help="Number of RNN layers the model",
+    )
+    parser.add_argument(
+        "--rnn-lm-tie-weights",
+        type=str2bool,
+        default=False,
+        help="""True to share the weights between the input embedding layer and the
+        last output linear layer
+        """,
+    )
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    rnnlm: Optional[RnnLmModel] = None,
+    rnnlm_scale: float = 1.0,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_LG, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 2, feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+
+    if feature.ndim == 2:
+        feature_lens = [] 
+        for supervision in supervisions['cut']:
+            try: feature_lens.append(supervision.tracks[0].cut.recording.num_samples)
+            except: feature_lens.append(supervision.recording.num_samples)
+        feature_lens = torch.tensor(feature_lens).to(device)
+    
+    elif feature.ndim == 3:
+        feature_lens = supervisions["num_frames"].to(device)
+
+    if params.simulate_streaming:
+        feature_lens += params.left_context
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, params.left_context),
+            value=LOG_EPS,
+        )
+        encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward(
+            x=feature,
+            x_lens=feature_lens,
+            chunk_size=params.decode_chunk_size,
+            left_context=params.left_context,
+            simulate_streaming=True,
+        )
+    else:
+        encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+
+    if (
+        params.decoding_method == "fast_beam_search"
+        or params.decoding_method == "fast_beam_search_LG"
+    ):
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        if params.decoding_method == "fast_beam_search":
+            for hyp in sp.decode(hyp_tokens):
+                hyps.append(hyp.split())
+        else:
+            for hyp in hyp_tokens:
+                hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+
+        texts = supervisions["text"]
+        #for enum, text in enumerate(texts):
+        logging.info(f"ref: {texts[0]}")
+        logging.info(f"hyp: {' '.join(hyps[0])}")
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_rnnlm_shallow_fusion":
+        hyp_tokens = modified_beam_search_rnnlm_shallow_fusion(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            sp=sp,
+            rnnlm=rnnlm,
+            rnnlm_scale=rnnlm_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+    
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+        if "LG" in params.decoding_method:
+            key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    rnnlm: Optional[RnnLmModel] = None,
+    rnnlm_scale: float = 1.0,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_LG, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+        logging.info(f"Decoding {batch_idx}-th batch")
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+            rnnlm=rnnlm,
+            rnnlm_scale=rnnlm_scale,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_LG",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+        "modified_beam_search_rnnlm_shallow_fusion",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+    if params.simulate_streaming:
+        params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
+        params.suffix += f"-left-context-{params.left_context}"
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+        if "LG" in params.decoding_method:
+            params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    params.suffix += f"-rnnlm-lm-scale-{params.rnn_lm_scale}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if params.simulate_streaming:
+        assert (
+            params.causal_convolution
+        ), "Decoding in streaming requires causal convolution"
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    rnn_lm_model = None
+    rnn_lm_scale = params.rnn_lm_scale
+    if params.decoding_method == "modified_beam_search_rnnlm_shallow_fusion":
+        rnn_lm_model = RnnLmModel(
+            vocab_size=params.vocab_size,
+            embedding_dim=params.rnn_lm_embedding_dim,
+            hidden_dim=params.rnn_lm_hidden_dim,
+            num_layers=params.rnn_lm_num_layers,
+            tie_weights=params.rnn_lm_tie_weights,
+        )
+        assert params.rnn_lm_avg == 1
+
+        load_checkpoint(
+            f"{params.rnn_lm_exp_dir}/epoch-{params.rnn_lm_epoch}.pt",
+            rnn_lm_model,
+        )
+        rnn_lm_model.to(device)
+        rnn_lm_model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if "LG" in params.decoding_method:
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+    #test_clean_cuts = librispeech.train_clean_100_cuts()
+    #test_other_cuts = librispeech.train_clean_100_cuts()
+    
+    #test_clean_dl = librispeech.train_dataloaders(test_clean_cuts)
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+            rnnlm=rnn_lm_model,
+            rnnlm_scale=rnn_lm_scale,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()

egs/librispeech/ASR/pruned_transducer_stateless_d2v/decode_stream.py

@@ -0,0 +1,146 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import List, Optional, Tuple
+
+import k2
+import torch
+from beam_search import Hypothesis, HypothesisList
+
+from icefall.utils import AttributeDict
+
+
+class DecodeStream(object):
+    def __init__(
+        self,
+        params: AttributeDict,
+        cut_id: str,
+        initial_states: List[torch.Tensor],
+        decoding_graph: Optional[k2.Fsa] = None,
+        device: torch.device = torch.device("cpu"),
+    ) -> None:
+        """
+        Args:
+          initial_states:
+            Initial decode states of the model, e.g. the return value of
+            `get_init_state` in conformer.py
+          decoding_graph:
+            Decoding graph used for decoding, may be a TrivialGraph or a HLG.
+            Used only when decoding_method is fast_beam_search.
+          device:
+            The device to run this stream.
+        """
+        if params.decoding_method == "fast_beam_search":
+            assert decoding_graph is not None
+            assert device == decoding_graph.device
+
+        self.params = params
+        self.cut_id = cut_id
+        self.LOG_EPS = math.log(1e-10)
+
+        self.states = initial_states
+
+        # It contains a 2-D tensors representing the feature frames.
+        self.features: torch.Tensor = None
+
+        self.num_frames: int = 0
+        # how many frames have been processed. (before subsampling).
+        # we only modify this value in `func:get_feature_frames`.
+        self.num_processed_frames: int = 0
+
+        self._done: bool = False
+
+        # The transcript of current utterance.
+        self.ground_truth: str = ""
+
+        # The decoding result (partial or final) of current utterance.
+        self.hyp: List = []
+
+        # how many frames have been processed, after subsampling (i.e. a
+        # cumulative sum of the second return value of
+        # encoder.streaming_forward
+        self.done_frames: int = 0
+
+        self.pad_length = (params.right_context + 2) * params.subsampling_factor + 3
+
+        if params.decoding_method == "greedy_search":
+            self.hyp = [params.blank_id] * params.context_size
+        elif params.decoding_method == "modified_beam_search":
+            self.hyps = HypothesisList()
+            self.hyps.add(
+                Hypothesis(
+                    ys=[params.blank_id] * params.context_size,
+                    log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+                )
+            )
+        elif params.decoding_method == "fast_beam_search":
+            # The rnnt_decoding_stream for fast_beam_search.
+            self.rnnt_decoding_stream: k2.RnntDecodingStream = k2.RnntDecodingStream(
+                decoding_graph
+            )
+        else:
+            raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    @property
+    def done(self) -> bool:
+        """Return True if all the features are processed."""
+        return self._done
+
+    @property
+    def id(self) -> str:
+        return self.cut_id
+
+    def set_features(
+        self,
+        features: torch.Tensor,
+    ) -> None:
+        """Set features tensor of current utterance."""
+        assert features.dim() == 2, features.dim()
+        self.features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, self.pad_length),
+            mode="constant",
+            value=self.LOG_EPS,
+        )
+        self.num_frames = self.features.size(0)
+
+    def get_feature_frames(self, chunk_size: int) -> Tuple[torch.Tensor, int]:
+        """Consume chunk_size frames of features"""
+        chunk_length = chunk_size + self.pad_length
+
+        ret_length = min(self.num_frames - self.num_processed_frames, chunk_length)
+
+        ret_features = self.features[
+            self.num_processed_frames : self.num_processed_frames + ret_length  # noqa
+        ]
+
+        self.num_processed_frames += chunk_size
+        if self.num_processed_frames >= self.num_frames:
+            self._done = True
+
+        return ret_features, ret_length
+
+    def decoding_result(self) -> List[int]:
+        """Obtain current decoding result."""
+        if self.params.decoding_method == "greedy_search":
+            return self.hyp[self.params.context_size :]  # noqa
+        elif self.params.decoding_method == "modified_beam_search":
+            best_hyp = self.hyps.get_most_probable(length_norm=True)
+            return best_hyp.ys[self.params.context_size :]  # noqa
+        else:
+            assert self.params.decoding_method == "fast_beam_search"
+            return self.hyp

egs/librispeech/ASR/pruned_transducer_stateless_d2v/decoder.py

@@ -0,0 +1,123 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from scaling import ScaledConv1d, ScaledEmbedding
+
+from icefall.utils import is_jit_tracing
+
+
+class Decoder(nn.Module):
+    """This class modifies the stateless decoder from the following paper:
+
+        RNN-transducer with stateless prediction network
+        https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+
+    It removes the recurrent connection from the decoder, i.e., the prediction
+    network. Different from the above paper, it adds an extra Conv1d
+    right after the embedding layer.
+
+    TODO: Implement https://arxiv.org/pdf/2109.07513.pdf
+    """
+
+    def __init__(
+        self,
+        vocab_size: int,
+        decoder_dim: int,
+        blank_id: int,
+        context_size: int,
+    ):
+        """
+        Args:
+          vocab_size:
+            Number of tokens of the modeling unit including blank.
+          decoder_dim:
+            Dimension of the input embedding, and of the decoder output.
+          blank_id:
+            The ID of the blank symbol.
+          context_size:
+            Number of previous words to use to predict the next word.
+            1 means bigram; 2 means trigram. n means (n+1)-gram.
+        """
+        super().__init__()
+
+        self.embedding = ScaledEmbedding(
+            num_embeddings=vocab_size,
+            embedding_dim=decoder_dim,
+            padding_idx=blank_id,
+        )
+        self.blank_id = blank_id
+
+        assert context_size >= 1, context_size
+        self.context_size = context_size
+        self.vocab_size = vocab_size
+        if context_size > 1:
+            self.conv = ScaledConv1d(
+                in_channels=decoder_dim,
+                out_channels=decoder_dim,
+                kernel_size=context_size,
+                padding=0,
+                groups=decoder_dim,
+                bias=False,
+            )
+        else:
+            # It is to support torch script
+            self.conv = nn.Identity()
+
+    def forward(
+        self,
+        y: torch.Tensor,
+        need_pad: bool = True  # Annotation should be Union[bool, torch.Tensor]
+        # but, torch.jit.script does not support Union.
+    ) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, U).
+          need_pad:
+            True to left pad the input. Should be True during training.
+            False to not pad the input. Should be False during inference.
+        Returns:
+          Return a tensor of shape (N, U, decoder_dim).
+        """
+        if isinstance(need_pad, torch.Tensor):
+            # This is for torch.jit.trace(), which cannot handle the case
+            # when the input argument is not a tensor.
+            need_pad = bool(need_pad)
+
+        y = y.to(torch.int64)
+        # this stuff about clamp() is a temporary fix for a mismatch
+        # at utterance start, we use negative ids in beam_search.py
+        if torch.jit.is_tracing():
+            # This is for exporting to PNNX via ONNX
+            embedding_out = self.embedding(y)
+        else:
+            embedding_out = self.embedding(y.clamp(min=0)) * (y >= 0).unsqueeze(-1)
+        if self.context_size > 1:
+            embedding_out = embedding_out.permute(0, 2, 1)
+            if need_pad:
+                embedding_out = F.pad(embedding_out, pad=(self.context_size - 1, 0))
+            else:
+                # During inference time, there is no need to do extra padding
+                # as we only need one output
+                if not is_jit_tracing():
+                    assert embedding_out.size(-1) == self.context_size
+            embedding_out = self.conv(embedding_out)
+            embedding_out = embedding_out.permute(0, 2, 1)
+        embedding_out = F.relu(embedding_out)
+        return embedding_out

egs/librispeech/ASR/pruned_transducer_stateless_d2v/encoder_interface.py

@@ -0,0 +1,43 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+
+
+class EncoderInterface(nn.Module):
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A tensor of shape (batch_size, input_seq_len, num_features)
+            containing the input features.
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames
+            in `x` before padding.
+        Returns:
+          Return a tuple containing two tensors:
+            - encoder_out, a tensor of (batch_size, out_seq_len, output_dim)
+              containing unnormalized probabilities, i.e., the output of a
+              linear layer.
+            - encoder_out_lens, a tensor of shape (batch_size,) containing
+              the number of frames in `encoder_out` before padding.
+        """
+        raise NotImplementedError("Please implement it in a subclass")

egs/librispeech/ASR/pruned_transducer_stateless_d2v/export.py

@@ -0,0 +1,287 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./pruned_transducer_stateless5/export.py \
+  --exp-dir ./pruned_transducer_stateless5/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `pruned_transducer_stateless5/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./pruned_transducer_stateless5/decode.py \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--streaming-model",
+        type=str2bool,
+        default=False,
+        help="""Whether to export a streaming model, if the models in exp-dir
+        are streaming model, this should be True.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if params.streaming_model:
+        assert params.causal_convolution
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        convert_scaled_to_non_scaled(model, inplace=True)
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()

egs/librispeech/ASR/pruned_transducer_stateless_d2v/joiner.py

@@ -0,0 +1,69 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+from scaling import ScaledLinear
+
+from icefall.utils import is_jit_tracing
+
+
+class Joiner(nn.Module):
+    def __init__(
+        self,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+    ):
+        super().__init__()
+
+        self.encoder_proj = ScaledLinear(encoder_dim, joiner_dim)
+        self.decoder_proj = ScaledLinear(decoder_dim, joiner_dim)
+        self.output_linear = ScaledLinear(joiner_dim, vocab_size)
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+        project_input: bool = True,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            Output from the encoder. Its shape is (N, T, s_range, C).
+          decoder_out:
+            Output from the decoder. Its shape is (N, T, s_range, C).
+           project_input:
+            If true, apply input projections encoder_proj and decoder_proj.
+            If this is false, it is the user's responsibility to do this
+            manually.
+        Returns:
+          Return a tensor of shape (N, T, s_range, C).
+        """
+        if not is_jit_tracing():
+            assert encoder_out.ndim == decoder_out.ndim
+            assert encoder_out.ndim in (2, 4)
+            assert encoder_out.shape == decoder_out.shape
+
+        if project_input:
+            logit = self.encoder_proj(encoder_out) + self.decoder_proj(decoder_out)
+        else:
+            logit = encoder_out + decoder_out
+
+        logit = self.output_linear(torch.tanh(logit))
+
+        return logit

egs/librispeech/ASR/pruned_transducer_stateless_d2v/lstmp.py

@@ -0,0 +1,102 @@
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class LSTMP(nn.Module):
+    """LSTM with projection.
+
+    PyTorch does not support exporting LSTM with projection to ONNX.
+    This class reimplements LSTM with projection using basic matrix-matrix
+    and matrix-vector operations. It is not intended for training.
+    """
+
+    def __init__(self, lstm: nn.LSTM):
+        """
+        Args:
+          lstm:
+            LSTM with proj_size. We support only uni-directional,
+            1-layer LSTM with projection at present.
+        """
+        super().__init__()
+        assert lstm.bidirectional is False, lstm.bidirectional
+        assert lstm.num_layers == 1, lstm.num_layers
+        assert 0 < lstm.proj_size < lstm.hidden_size, (
+            lstm.proj_size,
+            lstm.hidden_size,
+        )
+
+        assert lstm.batch_first is False, lstm.batch_first
+
+        state_dict = lstm.state_dict()
+
+        w_ih = state_dict["weight_ih_l0"]
+        w_hh = state_dict["weight_hh_l0"]
+
+        b_ih = state_dict["bias_ih_l0"]
+        b_hh = state_dict["bias_hh_l0"]
+
+        w_hr = state_dict["weight_hr_l0"]
+        self.input_size = lstm.input_size
+        self.proj_size = lstm.proj_size
+        self.hidden_size = lstm.hidden_size
+
+        self.w_ih = w_ih
+        self.w_hh = w_hh
+        self.b = b_ih + b_hh
+        self.w_hr = w_hr
+
+    def forward(
+        self,
+        input: torch.Tensor,
+        hx: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    ) -> Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Args:
+          input:
+            A tensor of shape [T, N, hidden_size]
+          hx:
+            A tuple containing:
+              - h0: a tensor of shape (1, N, proj_size)
+              - c0: a tensor of shape (1, N, hidden_size)
+        Returns:
+          Return a tuple containing:
+            - output: a tensor of shape (T, N, proj_size).
+            - A tuple containing:
+               - h: a tensor of shape (1, N, proj_size)
+               - c: a tensor of shape (1, N, hidden_size)
+
+        """
+        x_list = input.unbind(dim=0)  # We use batch_first=False
+
+        if hx is not None:
+            h0, c0 = hx
+        else:
+            h0 = torch.zeros(1, input.size(1), self.proj_size)
+            c0 = torch.zeros(1, input.size(1), self.hidden_size)
+        h0 = h0.squeeze(0)
+        c0 = c0.squeeze(0)
+        y_list = []
+        for x in x_list:
+            gates = F.linear(x, self.w_ih, self.b) + F.linear(h0, self.w_hh)
+            i, f, g, o = gates.chunk(4, dim=1)
+
+            i = i.sigmoid()
+            f = f.sigmoid()
+            g = g.tanh()
+            o = o.sigmoid()
+
+            c = f * c0 + i * g
+            h = o * c.tanh()
+
+            h = F.linear(h, self.w_hr)
+            y_list.append(h)
+
+            c0 = c
+            h0 = h
+
+        y = torch.stack(y_list, dim=0)
+
+        return y, (h0.unsqueeze(0), c0.unsqueeze(0))

egs/librispeech/ASR/pruned_transducer_stateless_d2v/model.py

@@ -0,0 +1,280 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from scaling import ScaledLinear
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dm) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, decoder_dim).
+            It should contain one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, encoder_dim) and
+            (N, U, decoder_dim).
+            Its output shape is (N, T, U, vocab_size). Note that its output
+            contains unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+        self.simple_am_proj = ScaledLinear(encoder_dim, vocab_size, initial_speed=0.5)
+        self.simple_lm_proj = ScaledLinear(decoder_dim, vocab_size)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+        warmup: float = 1.0,
+        reduction: str = "sum",
+        delay_penalty: float = 0.0,
+        print_tensor: bool = False,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+          warmup:
+            A value warmup >= 0 that determines which modules are active, values
+            warmup > 1 "are fully warmed up" and all modules will be active.
+          reduction:
+            "sum" to sum the losses over all utterances in the batch.
+            "none" to return the loss in a 1-D tensor for each utterance
+            in the batch.
+          delay_penalty:
+            A constant value used to penalize symbol delay, to encourage
+            streaming models to emit symbols earlier.
+            See https://github.com/k2-fsa/k2/issues/955 and
+            https://arxiv.org/pdf/2211.00490.pdf for more details.
+        Returns:
+        Returns:
+          Return the transducer loss.
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        assert reduction in ("sum", "none"), reduction
+        assert x.ndim == 2 or x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, x_lens = self.encoder(x, x_lens, warmup=warmup)
+
+        #encoder_out *= 100
+
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, decoder_dim]
+        decoder_out = self.decoder(sos_y_padded)
+        
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros((x.size(0), 4), dtype=torch.int64, device=x.device)
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = x_lens
+
+        lm = self.simple_lm_proj(decoder_out)
+        am = self.simple_am_proj(encoder_out)
+        
+        if encoder_out.get_device() == 0 and 0:
+            print('encoder out = ', encoder_out)
+            print('encoder size = ', encoder_out.size())
+            
+            print('am = ', am)
+            print('am size = ', am.size())
+
+            print('sos y padded = ', sos_y_padded)
+            print('sos y padded size = ', sos_y_padded.size())
+
+            print('decoder out = ', decoder_out)
+            print('decoder size = ', decoder_out.size())
+            
+            print('lm = ', lm)
+            print('lm size = ', lm.size())
+
+            print('\n\n\n\n\n')
+
+        with torch.cuda.amp.autocast(enabled=False):
+            simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+                lm=lm.float(),
+                am=am.float(),
+                symbols=y_padded,
+                termination_symbol=blank_id,
+                lm_only_scale=lm_scale,
+                am_only_scale=am_scale,
+                boundary=boundary,
+                reduction=reduction,
+                delay_penalty=delay_penalty,
+                return_grad=True,
+            )
+            #print('1. simple loss = ', simple_loss)
+            #print('2. px_grad = ', px_grad)
+            #print('3. py_grad = ', py_grad)
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, encoder_dim]
+        # lm_pruned : [B, T, prune_range, decoder_dim]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=self.joiner.encoder_proj(encoder_out),
+            lm=self.joiner.decoder_proj(decoder_out),
+            ranges=ranges,
+        )
+
+        # logits : [B, T, prune_range, vocab_size]
+
+        # project_input=False since we applied the decoder's input projections
+        # prior to do_rnnt_pruning (this is an optimization for speed).
+        logits = self.joiner(am_pruned, lm_pruned, project_input=False)
+        
+        with torch.cuda.amp.autocast(enabled=False):
+            pruned_loss = k2.rnnt_loss_pruned(
+                logits=logits.float(),
+                symbols=y_padded,
+                ranges=ranges,
+                termination_symbol=blank_id,
+                boundary=boundary,
+                delay_penalty=delay_penalty,
+                reduction=reduction,
+            )
+
+        return (simple_loss, pruned_loss)
+    
+    def decode(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        sp,
+    ): 
+        from beam_search import greedy_search_batch, greedy_search_batch_target_input
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, x_lens = self.encoder(x, x_lens)
+
+        assert torch.all(x_lens > 0)
+        
+        '''
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, decoder_dim]
+        decoder_out = self.decoder(sos_y_padded)
+        
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros((x.size(0), 4), dtype=torch.int64, device=x.device)
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = x_lens
+        '''
+
+        hyps = []
+        #hyp_tokens = greedy_search_batch_target_input(self, encoder_out, x_lens, decoder_out)
+        hyp_tokens = greedy_search_batch(self, encoder_out, x_lens)#, decoder_out)
+
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+        
+        return hyps

egs/librispeech/ASR/pruned_transducer_stateless_d2v/nets_utils.py

@@ -0,0 +1,503 @@
+# -*- coding: utf-8 -*-
+
+"""Network related utility tools."""
+
+import logging
+from typing import Dict
+
+import numpy as np
+import torch
+
+
+def to_device(m, x):
+    """Send tensor into the device of the module.
+
+    Args:
+        m (torch.nn.Module): Torch module.
+        x (Tensor): Torch tensor.
+
+    Returns:
+        Tensor: Torch tensor located in the same place as torch module.
+
+    """
+    if isinstance(m, torch.nn.Module):
+        device = next(m.parameters()).device
+    elif isinstance(m, torch.Tensor):
+        device = m.device
+    else:
+        raise TypeError(
+            "Expected torch.nn.Module or torch.tensor, " f"bot got: {type(m)}"
+        )
+    return x.to(device)
+
+
+def pad_list(xs, pad_value):
+    """Perform padding for the list of tensors.
+
+    Args:
+        xs (List): List of Tensors [(T_1, `*`), (T_2, `*`), ..., (T_B, `*`)].
+        pad_value (float): Value for padding.
+
+    Returns:
+        Tensor: Padded tensor (B, Tmax, `*`).
+
+    Examples:
+        >>> x = [torch.ones(4), torch.ones(2), torch.ones(1)]
+        >>> x
+        [tensor([1., 1., 1., 1.]), tensor([1., 1.]), tensor([1.])]
+        >>> pad_list(x, 0)
+        tensor([[1., 1., 1., 1.],
+                [1., 1., 0., 0.],
+                [1., 0., 0., 0.]])
+
+    """
+    n_batch = len(xs)
+    max_len = max(x.size(0) for x in xs)
+    pad = xs[0].new(n_batch, max_len, *xs[0].size()[1:]).fill_(pad_value)
+
+    for i in range(n_batch):
+        pad[i, : xs[i].size(0)] = xs[i]
+
+    return pad
+
+
+def make_pad_mask(lengths, xs=None, length_dim=-1, maxlen=None):
+    """Make mask tensor containing indices of padded part.
+
+    Args:
+        lengths (LongTensor or List): Batch of lengths (B,).
+        xs (Tensor, optional): The reference tensor.
+            If set, masks will be the same shape as this tensor.
+        length_dim (int, optional): Dimension indicator of the above tensor.
+            See the example.
+
+    Returns:
+        Tensor: Mask tensor containing indices of padded part.
+                dtype=torch.uint8 in PyTorch 1.2-
+                dtype=torch.bool in PyTorch 1.2+ (including 1.2)
+
+    Examples:
+        With only lengths.
+
+        >>> lengths = [5, 3, 2]
+        >>> make_pad_mask(lengths)
+        masks = [[0, 0, 0, 0 ,0],
+                 [0, 0, 0, 1, 1],
+                 [0, 0, 1, 1, 1]]
+
+        With the reference tensor.
+
+        >>> xs = torch.zeros((3, 2, 4))
+        >>> make_pad_mask(lengths, xs)
+        tensor([[[0, 0, 0, 0],
+                 [0, 0, 0, 0]],
+                [[0, 0, 0, 1],
+                 [0, 0, 0, 1]],
+                [[0, 0, 1, 1],
+                 [0, 0, 1, 1]]], dtype=torch.uint8)
+        >>> xs = torch.zeros((3, 2, 6))
+        >>> make_pad_mask(lengths, xs)
+        tensor([[[0, 0, 0, 0, 0, 1],
+                 [0, 0, 0, 0, 0, 1]],
+                [[0, 0, 0, 1, 1, 1],
+                 [0, 0, 0, 1, 1, 1]],
+                [[0, 0, 1, 1, 1, 1],
+                 [0, 0, 1, 1, 1, 1]]], dtype=torch.uint8)
+
+        With the reference tensor and dimension indicator.
+
+        >>> xs = torch.zeros((3, 6, 6))
+        >>> make_pad_mask(lengths, xs, 1)
+        tensor([[[0, 0, 0, 0, 0, 0],
+                 [0, 0, 0, 0, 0, 0],
+                 [0, 0, 0, 0, 0, 0],
+                 [0, 0, 0, 0, 0, 0],
+                 [0, 0, 0, 0, 0, 0],
+                 [1, 1, 1, 1, 1, 1]],
+                [[0, 0, 0, 0, 0, 0],
+                 [0, 0, 0, 0, 0, 0],
+                 [0, 0, 0, 0, 0, 0],
+                 [1, 1, 1, 1, 1, 1],
+                 [1, 1, 1, 1, 1, 1],
+                 [1, 1, 1, 1, 1, 1]],
+                [[0, 0, 0, 0, 0, 0],
+                 [0, 0, 0, 0, 0, 0],
+                 [1, 1, 1, 1, 1, 1],
+                 [1, 1, 1, 1, 1, 1],
+                 [1, 1, 1, 1, 1, 1],
+                 [1, 1, 1, 1, 1, 1]]], dtype=torch.uint8)
+        >>> make_pad_mask(lengths, xs, 2)
+        tensor([[[0, 0, 0, 0, 0, 1],
+                 [0, 0, 0, 0, 0, 1],
+                 [0, 0, 0, 0, 0, 1],
+                 [0, 0, 0, 0, 0, 1],
+                 [0, 0, 0, 0, 0, 1],
+                 [0, 0, 0, 0, 0, 1]],
+                [[0, 0, 0, 1, 1, 1],
+                 [0, 0, 0, 1, 1, 1],
+                 [0, 0, 0, 1, 1, 1],
+                 [0, 0, 0, 1, 1, 1],
+                 [0, 0, 0, 1, 1, 1],
+                 [0, 0, 0, 1, 1, 1]],
+                [[0, 0, 1, 1, 1, 1],
+                 [0, 0, 1, 1, 1, 1],
+                 [0, 0, 1, 1, 1, 1],
+                 [0, 0, 1, 1, 1, 1],
+                 [0, 0, 1, 1, 1, 1],
+                 [0, 0, 1, 1, 1, 1]]], dtype=torch.uint8)
+
+    """
+    if length_dim == 0:
+        raise ValueError("length_dim cannot be 0: {}".format(length_dim))
+
+    if not isinstance(lengths, list):
+        lengths = lengths.long().tolist()
+
+    bs = int(len(lengths))
+    if maxlen is None:
+        if xs is None:
+            maxlen = int(max(lengths))
+        else:
+            maxlen = xs.size(length_dim)
+    else:
+        assert xs is None
+        assert maxlen >= int(max(lengths))
+
+    seq_range = torch.arange(0, maxlen, dtype=torch.int64)
+    seq_range_expand = seq_range.unsqueeze(0).expand(bs, maxlen)
+    seq_length_expand = seq_range_expand.new(lengths).unsqueeze(-1)
+    mask = seq_range_expand >= seq_length_expand
+
+    if xs is not None:
+        assert xs.size(0) == bs, (xs.size(0), bs)
+
+        if length_dim < 0:
+            length_dim = xs.dim() + length_dim
+        # ind = (:, None, ..., None, :, , None, ..., None)
+        ind = tuple(
+            slice(None) if i in (0, length_dim) else None for i in range(xs.dim())
+        )
+        mask = mask[ind].expand_as(xs).to(xs.device)
+    return mask
+
+
+def make_non_pad_mask(lengths, xs=None, length_dim=-1):
+    """Make mask tensor containing indices of non-padded part.
+
+    Args:
+        lengths (LongTensor or List): Batch of lengths (B,).
+        xs (Tensor, optional): The reference tensor.
+            If set, masks will be the same shape as this tensor.
+        length_dim (int, optional): Dimension indicator of the above tensor.
+            See the example.
+
+    Returns:
+        ByteTensor: mask tensor containing indices of padded part.
+                    dtype=torch.uint8 in PyTorch 1.2-
+                    dtype=torch.bool in PyTorch 1.2+ (including 1.2)
+
+    Examples:
+        With only lengths.
+
+        >>> lengths = [5, 3, 2]
+        >>> make_non_pad_mask(lengths)
+        masks = [[1, 1, 1, 1 ,1],
+                 [1, 1, 1, 0, 0],
+                 [1, 1, 0, 0, 0]]
+
+        With the reference tensor.
+
+        >>> xs = torch.zeros((3, 2, 4))
+        >>> make_non_pad_mask(lengths, xs)
+        tensor([[[1, 1, 1, 1],
+                 [1, 1, 1, 1]],
+                [[1, 1, 1, 0],
+                 [1, 1, 1, 0]],
+                [[1, 1, 0, 0],
+                 [1, 1, 0, 0]]], dtype=torch.uint8)
+        >>> xs = torch.zeros((3, 2, 6))
+        >>> make_non_pad_mask(lengths, xs)
+        tensor([[[1, 1, 1, 1, 1, 0],
+                 [1, 1, 1, 1, 1, 0]],
+                [[1, 1, 1, 0, 0, 0],
+                 [1, 1, 1, 0, 0, 0]],
+                [[1, 1, 0, 0, 0, 0],
+                 [1, 1, 0, 0, 0, 0]]], dtype=torch.uint8)
+
+        With the reference tensor and dimension indicator.
+
+        >>> xs = torch.zeros((3, 6, 6))
+        >>> make_non_pad_mask(lengths, xs, 1)
+        tensor([[[1, 1, 1, 1, 1, 1],
+                 [1, 1, 1, 1, 1, 1],
+                 [1, 1, 1, 1, 1, 1],
+                 [1, 1, 1, 1, 1, 1],
+                 [1, 1, 1, 1, 1, 1],
+                 [0, 0, 0, 0, 0, 0]],
+                [[1, 1, 1, 1, 1, 1],
+                 [1, 1, 1, 1, 1, 1],
+                 [1, 1, 1, 1, 1, 1],
+                 [0, 0, 0, 0, 0, 0],
+                 [0, 0, 0, 0, 0, 0],
+                 [0, 0, 0, 0, 0, 0]],
+                [[1, 1, 1, 1, 1, 1],
+                 [1, 1, 1, 1, 1, 1],
+                 [0, 0, 0, 0, 0, 0],
+                 [0, 0, 0, 0, 0, 0],
+                 [0, 0, 0, 0, 0, 0],
+                 [0, 0, 0, 0, 0, 0]]], dtype=torch.uint8)
+        >>> make_non_pad_mask(lengths, xs, 2)
+        tensor([[[1, 1, 1, 1, 1, 0],
+                 [1, 1, 1, 1, 1, 0],
+                 [1, 1, 1, 1, 1, 0],
+                 [1, 1, 1, 1, 1, 0],
+                 [1, 1, 1, 1, 1, 0],
+                 [1, 1, 1, 1, 1, 0]],
+                [[1, 1, 1, 0, 0, 0],
+                 [1, 1, 1, 0, 0, 0],
+                 [1, 1, 1, 0, 0, 0],
+                 [1, 1, 1, 0, 0, 0],
+                 [1, 1, 1, 0, 0, 0],
+                 [1, 1, 1, 0, 0, 0]],
+                [[1, 1, 0, 0, 0, 0],
+                 [1, 1, 0, 0, 0, 0],
+                 [1, 1, 0, 0, 0, 0],
+                 [1, 1, 0, 0, 0, 0],
+                 [1, 1, 0, 0, 0, 0],
+                 [1, 1, 0, 0, 0, 0]]], dtype=torch.uint8)
+
+    """
+    return ~make_pad_mask(lengths, xs, length_dim)
+
+
+def mask_by_length(xs, lengths, fill=0):
+    """Mask tensor according to length.
+
+    Args:
+        xs (Tensor): Batch of input tensor (B, `*`).
+        lengths (LongTensor or List): Batch of lengths (B,).
+        fill (int or float): Value to fill masked part.
+
+    Returns:
+        Tensor: Batch of masked input tensor (B, `*`).
+
+    Examples:
+        >>> x = torch.arange(5).repeat(3, 1) + 1
+        >>> x
+        tensor([[1, 2, 3, 4, 5],
+                [1, 2, 3, 4, 5],
+                [1, 2, 3, 4, 5]])
+        >>> lengths = [5, 3, 2]
+        >>> mask_by_length(x, lengths)
+        tensor([[1, 2, 3, 4, 5],
+                [1, 2, 3, 0, 0],
+                [1, 2, 0, 0, 0]])
+
+    """
+    assert xs.size(0) == len(lengths)
+    ret = xs.data.new(*xs.size()).fill_(fill)
+    for i, l in enumerate(lengths):
+        ret[i, :l] = xs[i, :l]
+    return ret
+
+
+def th_accuracy(pad_outputs, pad_targets, ignore_label):
+    """Calculate accuracy.
+
+    Args:
+        pad_outputs (Tensor): Prediction tensors (B * Lmax, D).
+        pad_targets (LongTensor): Target label tensors (B, Lmax, D).
+        ignore_label (int): Ignore label id.
+
+    Returns:
+        float: Accuracy value (0.0 - 1.0).
+
+    """
+    pad_pred = pad_outputs.view(
+        pad_targets.size(0), pad_targets.size(1), pad_outputs.size(1)
+    ).argmax(2)
+    mask = pad_targets != ignore_label
+    numerator = torch.sum(
+        pad_pred.masked_select(mask) == pad_targets.masked_select(mask)
+    )
+    denominator = torch.sum(mask)
+    return float(numerator) / float(denominator)
+
+
+def to_torch_tensor(x):
+    """Change to torch.Tensor or ComplexTensor from numpy.ndarray.
+
+    Args:
+        x: Inputs. It should be one of numpy.ndarray, Tensor, ComplexTensor, and dict.
+
+    Returns:
+        Tensor or ComplexTensor: Type converted inputs.
+
+    Examples:
+        >>> xs = np.ones(3, dtype=np.float32)
+        >>> xs = to_torch_tensor(xs)
+        tensor([1., 1., 1.])
+        >>> xs = torch.ones(3, 4, 5)
+        >>> assert to_torch_tensor(xs) is xs
+        >>> xs = {'real': xs, 'imag': xs}
+        >>> to_torch_tensor(xs)
+        ComplexTensor(
+        Real:
+        tensor([1., 1., 1.])
+        Imag;
+        tensor([1., 1., 1.])
+        )
+
+    """
+    # If numpy, change to torch tensor
+    if isinstance(x, np.ndarray):
+        if x.dtype.kind == "c":
+            # Dynamically importing because torch_complex requires python3
+            from torch_complex.tensor import ComplexTensor
+
+            return ComplexTensor(x)
+        else:
+            return torch.from_numpy(x)
+
+    # If {'real': ..., 'imag': ...}, convert to ComplexTensor
+    elif isinstance(x, dict):
+        # Dynamically importing because torch_complex requires python3
+        from torch_complex.tensor import ComplexTensor
+
+        if "real" not in x or "imag" not in x:
+            raise ValueError("has 'real' and 'imag' keys: {}".format(list(x)))
+        # Relative importing because of using python3 syntax
+        return ComplexTensor(x["real"], x["imag"])
+
+    # If torch.Tensor, as it is
+    elif isinstance(x, torch.Tensor):
+        return x
+
+    else:
+        error = (
+            "x must be numpy.ndarray, torch.Tensor or a dict like "
+            "{{'real': torch.Tensor, 'imag': torch.Tensor}}, "
+            "but got {}".format(type(x))
+        )
+        try:
+            from torch_complex.tensor import ComplexTensor
+        except Exception:
+            # If PY2
+            raise ValueError(error)
+        else:
+            # If PY3
+            if isinstance(x, ComplexTensor):
+                return x
+            else:
+                raise ValueError(error)
+
+
+def get_subsample(train_args, mode, arch):
+    """Parse the subsampling factors from the args for the specified `mode` and `arch`.
+
+    Args:
+        train_args: argument Namespace containing options.
+        mode: one of ('asr', 'mt', 'st')
+        arch: one of ('rnn', 'rnn-t', 'rnn_mix', 'rnn_mulenc', 'transformer')
+
+    Returns:
+        np.ndarray / List[np.ndarray]: subsampling factors.
+    """
+    if arch == "transformer":
+        return np.array([1])
+
+    elif mode == "mt" and arch == "rnn":
+        # +1 means input (+1) and layers outputs (train_args.elayer)
+        subsample = np.ones(train_args.elayers + 1, dtype=np.int64)
+        logging.warning("Subsampling is not performed for machine translation.")
+        logging.info("subsample: " + " ".join([str(x) for x in subsample]))
+        return subsample
+
+    elif (
+        (mode == "asr" and arch in ("rnn", "rnn-t"))
+        or (mode == "mt" and arch == "rnn")
+        or (mode == "st" and arch == "rnn")
+    ):
+        subsample = np.ones(train_args.elayers + 1, dtype=np.int64)
+        if train_args.etype.endswith("p") and not train_args.etype.startswith("vgg"):
+            ss = train_args.subsample.split("_")
+            for j in range(min(train_args.elayers + 1, len(ss))):
+                subsample[j] = int(ss[j])
+        else:
+            logging.warning(
+                "Subsampling is not performed for vgg*. "
+                "It is performed in max pooling layers at CNN."
+            )
+        logging.info("subsample: " + " ".join([str(x) for x in subsample]))
+        return subsample
+
+    elif mode == "asr" and arch == "rnn_mix":
+        subsample = np.ones(
+            train_args.elayers_sd + train_args.elayers + 1, dtype=np.int64
+        )
+        if train_args.etype.endswith("p") and not train_args.etype.startswith("vgg"):
+            ss = train_args.subsample.split("_")
+            for j in range(
+                min(train_args.elayers_sd + train_args.elayers + 1, len(ss))
+            ):
+                subsample[j] = int(ss[j])
+        else:
+            logging.warning(
+                "Subsampling is not performed for vgg*. "
+                "It is performed in max pooling layers at CNN."
+            )
+        logging.info("subsample: " + " ".join([str(x) for x in subsample]))
+        return subsample
+
+    elif mode == "asr" and arch == "rnn_mulenc":
+        subsample_list = []
+        for idx in range(train_args.num_encs):
+            subsample = np.ones(train_args.elayers[idx] + 1, dtype=np.int64)
+            if train_args.etype[idx].endswith("p") and not train_args.etype[
+                idx
+            ].startswith("vgg"):
+                ss = train_args.subsample[idx].split("_")
+                for j in range(min(train_args.elayers[idx] + 1, len(ss))):
+                    subsample[j] = int(ss[j])
+            else:
+                logging.warning(
+                    "Encoder %d: Subsampling is not performed for vgg*. "
+                    "It is performed in max pooling layers at CNN.",
+                    idx + 1,
+                )
+            logging.info("subsample: " + " ".join([str(x) for x in subsample]))
+            subsample_list.append(subsample)
+        return subsample_list
+
+    else:
+        raise ValueError("Invalid options: mode={}, arch={}".format(mode, arch))
+
+
+def rename_state_dict(
+    old_prefix: str, new_prefix: str, state_dict: Dict[str, torch.Tensor]
+):
+    """Replace keys of old prefix with new prefix in state dict."""
+    # need this list not to break the dict iterator
+    old_keys = [k for k in state_dict if k.startswith(old_prefix)]
+    if len(old_keys) > 0:
+        logging.warning(f"Rename: {old_prefix} -> {new_prefix}")
+    for k in old_keys:
+        v = state_dict.pop(k)
+        new_k = k.replace(old_prefix, new_prefix)
+        state_dict[new_k] = v
+
+
+def get_activation(act):
+    """Return activation function."""
+    # Lazy load to avoid unused import
+    from espnet.nets.pytorch_backend.conformer.swish import Swish
+
+    activation_funcs = {
+        "hardtanh": torch.nn.Hardtanh,
+        "tanh": torch.nn.Tanh,
+        "relu": torch.nn.ReLU,
+        "selu": torch.nn.SELU,
+        "swish": Swish,
+    }
+
+    return activation_funcs[act]()

egs/librispeech/ASR/pruned_transducer_stateless_d2v/optim.py

@@ -0,0 +1,424 @@
+# Copyright      2022  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import math
+from typing import List, Optional, Union
+
+import torch
+from torch.optim import Optimizer
+
+from torch.optim.lr_scheduler import _LRScheduler
+
+
+class Eve(Optimizer):
+    r"""
+    Implements Eve algorithm.  This is a modified version of AdamW with a special
+    way of setting the weight-decay / shrinkage-factor, which is designed to make the
+    rms of the parameters approach a particular target_rms (default: 0.1).  This is
+    for use with networks with 'scaled' versions of modules (see scaling.py), which
+    will be close to invariant to the absolute scale on the parameter matrix.
+
+    The original Adam algorithm was proposed in `Adam: A Method for Stochastic Optimization`_.
+    The AdamW variant was proposed in `Decoupled Weight Decay Regularization`_.
+    Eve is unpublished so far.
+
+    Arguments:
+        params (iterable): iterable of parameters to optimize or dicts defining
+            parameter groups
+        lr (float, optional): learning rate (default: 1e-3)
+        betas (Tuple[float, float], optional): coefficients used for computing
+            running averages of gradient and its square (default: (0.9, 0.999))
+        eps (float, optional): term added to the denominator to improve
+            numerical stability (default: 1e-8)
+        weight_decay (float, optional): weight decay coefficient (default: 3e-4;
+            this value means that the weight would decay significantly after
+            about 3k minibatches.  Is not multiplied by learning rate, but
+            is conditional on RMS-value of parameter being > target_rms.
+        target_rms (float, optional): target root-mean-square value of
+           parameters, if they fall below this we will stop applying weight decay.
+
+
+    .. _Adam\: A Method for Stochastic Optimization:
+        https://arxiv.org/abs/1412.6980
+    .. _Decoupled Weight Decay Regularization:
+        https://arxiv.org/abs/1711.05101
+    .. _On the Convergence of Adam and Beyond:
+        https://openreview.net/forum?id=ryQu7f-RZ
+    """
+
+    def __init__(
+        self,
+        params,
+        lr=1e-3,
+        betas=(0.9, 0.98),
+        eps=1e-8,
+        weight_decay=1e-3,
+        target_rms=0.1,
+    ):
+
+        if not 0.0 <= lr:
+            raise ValueError("Invalid learning rate: {}".format(lr))
+        if not 0.0 <= eps:
+            raise ValueError("Invalid epsilon value: {}".format(eps))
+        if not 0.0 <= betas[0] < 1.0:
+            raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
+        if not 0.0 <= betas[1] < 1.0:
+            raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
+        if not 0 <= weight_decay <= 0.1:
+            raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
+        if not 0 < target_rms <= 10.0:
+            raise ValueError("Invalid target_rms value: {}".format(target_rms))
+        defaults = dict(
+            lr=lr,
+            betas=betas,
+            eps=eps,
+            weight_decay=weight_decay,
+            target_rms=target_rms,
+        )
+        super(Eve, self).__init__(params, defaults)
+
+    def __setstate__(self, state):
+        super(Eve, self).__setstate__(state)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """Performs a single optimization step.
+
+        Arguments:
+            closure (callable, optional): A closure that reevaluates the model
+                and returns the loss.
+        """
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        for group in self.param_groups:
+            for p in group["params"]:
+                if p.grad is None:
+                    continue
+
+                # Perform optimization step
+                grad = p.grad
+                if grad.is_sparse:
+                    raise RuntimeError("AdamW does not support sparse gradients")
+
+                state = self.state[p]
+
+                # State initialization
+                if len(state) == 0:
+                    state["step"] = 0
+                    # Exponential moving average of gradient values
+                    state["exp_avg"] = torch.zeros_like(
+                        p, memory_format=torch.preserve_format
+                    )
+                    # Exponential moving average of squared gradient values
+                    state["exp_avg_sq"] = torch.zeros_like(
+                        p, memory_format=torch.preserve_format
+                    )
+
+                exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"]
+
+                beta1, beta2 = group["betas"]
+
+                state["step"] += 1
+                bias_correction1 = 1 - beta1 ** state["step"]
+                bias_correction2 = 1 - beta2 ** state["step"]
+
+                # Decay the first and second moment running average coefficient
+                exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
+                exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
+                denom = (exp_avg_sq.sqrt() * (bias_correction2**-0.5)).add_(
+                    group["eps"]
+                )
+
+                step_size = group["lr"] / bias_correction1
+                target_rms = group["target_rms"]
+                weight_decay = group["weight_decay"]
+
+                if p.numel() > 1:
+                    # avoid applying this weight-decay on "scaling factors"
+                    # (which are scalar).
+                    is_above_target_rms = p.norm() > (target_rms * (p.numel() ** 0.5))
+                    p.mul_(1 - (weight_decay * is_above_target_rms))
+                p.addcdiv_(exp_avg, denom, value=-step_size)
+
+                # Constrain the range of scalar weights
+                if p.numel() == 1:
+                    p.clamp_(min=-10, max=2)
+
+        return loss
+
+
+class LRScheduler(object):
+    """
+    Base-class for learning rate schedulers where the learning-rate depends on both the
+    batch and the epoch.
+    """
+
+    def __init__(self, optimizer: Optimizer, verbose: bool = False):
+        # Attach optimizer
+        if not isinstance(optimizer, Optimizer):
+            raise TypeError("{} is not an Optimizer".format(type(optimizer).__name__))
+        self.optimizer = optimizer
+        self.verbose = verbose
+
+        for group in optimizer.param_groups:
+            group.setdefault("initial_lr", group["lr"])
+
+        self.base_lrs = [group["initial_lr"] for group in optimizer.param_groups]
+
+        self.epoch = 0
+        self.batch = 0
+
+    def state_dict(self):
+        """Returns the state of the scheduler as a :class:`dict`.
+
+        It contains an entry for every variable in self.__dict__ which
+        is not the optimizer.
+        """
+        return {
+            "base_lrs": self.base_lrs,
+            "epoch": self.epoch,
+            "batch": self.batch,
+        }
+
+    def load_state_dict(self, state_dict):
+        """Loads the schedulers state.
+
+        Args:
+            state_dict (dict): scheduler state. Should be an object returned
+                from a call to :meth:`state_dict`.
+        """
+        self.__dict__.update(state_dict)
+
+    def get_last_lr(self) -> List[float]:
+        """Return last computed learning rate by current scheduler.  Will be a list of float."""
+        return self._last_lr
+
+    def get_lr(self):
+        # Compute list of learning rates from self.epoch and self.batch and
+        # self.base_lrs; this must be overloaded by the user.
+        # e.g. return [some_formula(self.batch, self.epoch, base_lr) for base_lr in self.base_lrs ]
+        raise NotImplementedError
+
+    def step_batch(self, batch: Optional[int] = None) -> None:
+        # Step the batch index, or just set it.  If `batch` is specified, it
+        # must be the batch index from the start of training, i.e. summed over
+        # all epochs.
+        # You can call this in any order; if you don't provide 'batch', it should
+        # of course be called once per batch.
+        if batch is not None:
+            self.batch = batch
+        else:
+            self.batch = self.batch + 1
+        self._set_lrs()
+
+    def step_epoch(self, epoch: Optional[int] = None):
+        # Step the epoch index, or just set it.  If you provide the 'epoch' arg,
+        # you should call this at the start of the epoch; if you don't provide the 'epoch'
+        # arg, you should call it at the end of the epoch.
+        if epoch is not None:
+            self.epoch = epoch
+        else:
+            self.epoch = self.epoch + 1
+        self._set_lrs()
+
+    def _set_lrs(self):
+        values = self.get_lr()
+        assert len(values) == len(self.optimizer.param_groups)
+
+        for i, data in enumerate(zip(self.optimizer.param_groups, values)):
+            param_group, lr = data
+            param_group["lr"] = lr
+            self.print_lr(self.verbose, i, lr)
+        self._last_lr = [group["lr"] for group in self.optimizer.param_groups]
+
+    def print_lr(self, is_verbose, group, lr):
+        """Display the current learning rate."""
+        if is_verbose:
+            print(
+                f"Epoch={self.epoch}, batch={self.batch}: adjusting learning rate"
+                f" of group {group} to {lr:.4e}."
+            )
+
+
+class Eden(LRScheduler):
+    """
+    Eden scheduler.
+     lr = initial_lr * (((batch**2 + lr_batches**2) / lr_batches**2) ** -0.25 *
+                       (((epoch**2 + lr_epochs**2) / lr_epochs**2) ** -0.25))
+
+     E.g. suggest initial-lr = 0.003 (passed to optimizer).
+
+    Args:
+        optimizer: the optimizer to change the learning rates on
+        lr_batches: the number of batches after which we start significantly
+              decreasing the learning rate, suggest 5000.
+        lr_epochs: the number of epochs after which we start significantly
+              decreasing the learning rate, suggest 6 if you plan to do e.g.
+              20 to 40 epochs, but may need smaller number if dataset is huge
+              and you will do few epochs.
+    """
+
+    def __init__(
+        self,
+        optimizer: Optimizer,
+        lr_batches: Union[int, float],
+        lr_epochs: Union[int, float],
+        verbose: bool = False,
+    ):
+        super(Eden, self).__init__(optimizer, verbose)
+        self.lr_batches = lr_batches
+        self.lr_epochs = lr_epochs
+
+    def get_lr(self):
+        factor = (
+            (self.batch**2 + self.lr_batches**2) / self.lr_batches**2
+        ) ** -0.25 * (
+            ((self.epoch**2 + self.lr_epochs**2) / self.lr_epochs**2) ** -0.25
+        )
+
+        return [x * factor for x in self.base_lrs]
+
+
+class TriStageLRScheduler(LRScheduler):
+    r"""
+    Tri-Stage Learning Rate Scheduler. Implement the learning rate scheduler in "SpecAugment"
+
+    Args:
+        optimizer (Optimizer): Optimizer.
+        init_lr (float): Initial learning rate.
+        peak_lr (float): Maximum learning rate.
+        final_lr (float): Final learning rate.
+        init_lr_scale (float): Initial learning rate scale.
+        final_lr_scale (float): Final learning rate scale.
+        warmup_steps (int): Warmup the learning rate linearly for the first N updates.
+        hold_steps (int): Hold the learning rate for the N updates.
+        decay_steps (int): Decay the learning rate linearly for the first N updates.
+        total_steps (int): Total steps in training.
+    """
+    def __init__(
+            self,
+            optimizer: Optimizer,
+            init_lr: float,
+            peak_lr: float,
+            final_lr: float,
+            init_lr_scale: float,
+            final_lr_scale: float,
+            warmup_steps: int,
+            hold_steps: int,
+            decay_steps: int,
+            total_steps: int,
+            verbose:bool = False
+    ):
+        assert isinstance(warmup_steps, int), "warmup_steps should be inteager type"
+        assert isinstance(total_steps, int), "total_steps should be inteager type"
+        
+        super(TriStageLRScheduler, self).__init__(optimizer, verbose)
+
+        self.init_lr = init_lr
+        self.init_lr *= init_lr_scale
+        self.final_lr = final_lr
+        self.peak_lr = peak_lr
+        self.warmup_steps = warmup_steps
+        self.hold_steps = hold_steps
+        self.decay_steps = decay_steps
+
+        self.warmup_rate = (self.peak_lr - self.init_lr) / self.warmup_steps if self.warmup_steps != 0 else 0
+        self.decay_factor = -math.log(final_lr_scale) / self.decay_steps
+
+        self.lr = self.init_lr
+        self.update_steps = 0
+
+    def _decide_stage(self):
+        if self.update_steps < self.warmup_steps:
+            return 0, self.update_steps
+
+        offset = self.warmup_steps
+
+        if self.update_steps < offset + self.hold_steps:
+            return 1, self.update_steps - offset
+
+        offset += self.hold_steps
+
+        if self.update_steps <= offset + self.decay_steps:
+            # decay stage
+            return 2, self.update_steps - offset
+
+        offset += self.decay_steps
+
+        return 3, self.update_steps - offset
+
+    def get_lr(self, val_loss: Optional[torch.FloatTensor] = None):
+        stage, steps_in_stage = self._decide_stage()
+
+        if stage == 0:
+            self.lr = self.init_lr + self.warmup_rate * steps_in_stage
+        elif stage == 1:
+            self.lr = self.peak_lr
+        elif stage == 2:
+            self.lr = self.peak_lr * math.exp(-self.decay_factor * steps_in_stage)
+        elif stage == 3:
+            self.lr = self.final_lr
+        else:
+            raise ValueError("Undefined stage")
+
+        #self.set_lr(self.optimizer, self.lr)
+        self.update_steps += 1
+
+        return [self.lr]
+
+
+def _test_eden():
+    m = torch.nn.Linear(100, 100)
+    optim = Eve(m.parameters(), lr=0.003)
+
+    #scheduler = Eden(optim, lr_batches=30, lr_epochs=2, verbose=True)
+    scheduler = TriStageLRScheduler(optim, 
+                                    init_lr=1e-05,
+                                    peak_lr=3e-04,
+                                    final_lr=1e-05,
+                                    init_lr_scale=0.1,
+                                    final_lr_scale=0.1,
+                                    warmup_steps=200,
+                                    hold_steps=300,
+                                    decay_steps=200,
+                                    total_steps=800,
+                                    verbose=True,
+                                )
+
+    for epoch in range(10):
+        #scheduler.step_epoch(epoch)  # sets epoch to `epoch`
+
+        for step in range(80):
+            x = torch.randn(200, 100).detach()
+            x.requires_grad = True
+            y = m(x)
+            dy = torch.randn(200, 100).detach()
+            f = (y * dy).sum()
+            f.backward()
+
+            optim.step()
+            scheduler.step_batch()
+            optim.zero_grad()
+    print("last lr = ", scheduler.get_last_lr())
+    print("state dict = ", scheduler.state_dict())
+
+
+if __name__ == "__main__":
+    _test_eden()

egs/librispeech/ASR/pruned_transducer_stateless_d2v/pretrained.py

@@ -0,0 +1,344 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+(1) greedy search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./pruned_transducer_stateless5/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless5/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless5/exp/pretrained.pt is generated by
+./pruned_transducer_stateless5/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lengths)
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append(sp.decode(hyp).split())
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()

egs/librispeech/ASR/pruned_transducer_stateless_d2v/scaling_converter.py

@@ -0,0 +1,320 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file provides functions to convert `ScaledLinear`, `ScaledConv1d`,
+`ScaledConv2d`, and `ScaledEmbedding` to their non-scaled counterparts:
+`nn.Linear`, `nn.Conv1d`, `nn.Conv2d`, and `nn.Embedding`.
+
+The scaled version are required only in the training time. It simplifies our
+life by converting them to their non-scaled version during inference.
+"""
+
+import copy
+import re
+from typing import List
+
+import torch
+import torch.nn as nn
+from lstmp import LSTMP
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    ScaledConv1d,
+    ScaledConv2d,
+    ScaledEmbedding,
+    ScaledLinear,
+    ScaledLSTM,
+)
+
+
+class NonScaledNorm(nn.Module):
+    """See BasicNorm for doc"""
+
+    def __init__(
+        self,
+        num_channels: int,
+        eps_exp: float,
+        channel_dim: int = -1,  # CAUTION: see documentation.
+    ):
+        super().__init__()
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        self.eps_exp = eps_exp
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if not torch.jit.is_tracing():
+            assert x.shape[self.channel_dim] == self.num_channels
+        scales = (
+            torch.mean(x * x, dim=self.channel_dim, keepdim=True) + self.eps_exp
+        ).pow(-0.5)
+        return x * scales
+
+
+def scaled_linear_to_linear(scaled_linear: ScaledLinear) -> nn.Linear:
+    """Convert an instance of ScaledLinear to nn.Linear.
+
+    Args:
+      scaled_linear:
+        The layer to be converted.
+    Returns:
+      Return a linear layer. It satisfies:
+
+        scaled_linear(x) == linear(x)
+
+      for any given input tensor `x`.
+    """
+    assert isinstance(scaled_linear, ScaledLinear), type(scaled_linear)
+
+    weight = scaled_linear.get_weight()
+    bias = scaled_linear.get_bias()
+    has_bias = bias is not None
+
+    linear = torch.nn.Linear(
+        in_features=scaled_linear.in_features,
+        out_features=scaled_linear.out_features,
+        bias=True,  # otherwise, it throws errors when converting to PNNX format
+        # device=weight.device,  # Pytorch version before v1.9.0 does not have
+        # this argument. Comment out for now, we will
+        # see if it will raise error for versions
+        # after v1.9.0
+    )
+    linear.weight.data.copy_(weight)
+
+    if has_bias:
+        linear.bias.data.copy_(bias)
+    else:
+        linear.bias.data.zero_()
+
+    return linear
+
+
+def scaled_conv1d_to_conv1d(scaled_conv1d: ScaledConv1d) -> nn.Conv1d:
+    """Convert an instance of ScaledConv1d to nn.Conv1d.
+
+    Args:
+      scaled_conv1d:
+        The layer to be converted.
+    Returns:
+      Return an instance of nn.Conv1d that has the same `forward()` behavior
+      of the given `scaled_conv1d`.
+    """
+    assert isinstance(scaled_conv1d, ScaledConv1d), type(scaled_conv1d)
+
+    weight = scaled_conv1d.get_weight()
+    bias = scaled_conv1d.get_bias()
+    has_bias = bias is not None
+
+    conv1d = nn.Conv1d(
+        in_channels=scaled_conv1d.in_channels,
+        out_channels=scaled_conv1d.out_channels,
+        kernel_size=scaled_conv1d.kernel_size,
+        stride=scaled_conv1d.stride,
+        padding=scaled_conv1d.padding,
+        dilation=scaled_conv1d.dilation,
+        groups=scaled_conv1d.groups,
+        bias=scaled_conv1d.bias is not None,
+        padding_mode=scaled_conv1d.padding_mode,
+    )
+
+    conv1d.weight.data.copy_(weight)
+    if has_bias:
+        conv1d.bias.data.copy_(bias)
+
+    return conv1d
+
+
+def scaled_conv2d_to_conv2d(scaled_conv2d: ScaledConv2d) -> nn.Conv2d:
+    """Convert an instance of ScaledConv2d to nn.Conv2d.
+
+    Args:
+      scaled_conv2d:
+        The layer to be converted.
+    Returns:
+      Return an instance of nn.Conv2d that has the same `forward()` behavior
+      of the given `scaled_conv2d`.
+    """
+    assert isinstance(scaled_conv2d, ScaledConv2d), type(scaled_conv2d)
+
+    weight = scaled_conv2d.get_weight()
+    bias = scaled_conv2d.get_bias()
+    has_bias = bias is not None
+
+    conv2d = nn.Conv2d(
+        in_channels=scaled_conv2d.in_channels,
+        out_channels=scaled_conv2d.out_channels,
+        kernel_size=scaled_conv2d.kernel_size,
+        stride=scaled_conv2d.stride,
+        padding=scaled_conv2d.padding,
+        dilation=scaled_conv2d.dilation,
+        groups=scaled_conv2d.groups,
+        bias=scaled_conv2d.bias is not None,
+        padding_mode=scaled_conv2d.padding_mode,
+    )
+
+    conv2d.weight.data.copy_(weight)
+    if has_bias:
+        conv2d.bias.data.copy_(bias)
+
+    return conv2d
+
+
+def scaled_embedding_to_embedding(
+    scaled_embedding: ScaledEmbedding,
+) -> nn.Embedding:
+    """Convert an instance of ScaledEmbedding to nn.Embedding.
+
+    Args:
+      scaled_embedding:
+        The layer to be converted.
+    Returns:
+      Return an instance of nn.Embedding that has the same `forward()` behavior
+      of the given `scaled_embedding`.
+    """
+    assert isinstance(scaled_embedding, ScaledEmbedding), type(scaled_embedding)
+    embedding = nn.Embedding(
+        num_embeddings=scaled_embedding.num_embeddings,
+        embedding_dim=scaled_embedding.embedding_dim,
+        padding_idx=scaled_embedding.padding_idx,
+        scale_grad_by_freq=scaled_embedding.scale_grad_by_freq,
+        sparse=scaled_embedding.sparse,
+    )
+    weight = scaled_embedding.weight
+    scale = scaled_embedding.scale
+
+    embedding.weight.data.copy_(weight * scale.exp())
+
+    return embedding
+
+
+def convert_basic_norm(basic_norm: BasicNorm) -> NonScaledNorm:
+    assert isinstance(basic_norm, BasicNorm), type(BasicNorm)
+    norm = NonScaledNorm(
+        num_channels=basic_norm.num_channels,
+        eps_exp=basic_norm.eps.data.exp().item(),
+        channel_dim=basic_norm.channel_dim,
+    )
+    return norm
+
+
+def scaled_lstm_to_lstm(scaled_lstm: ScaledLSTM) -> nn.LSTM:
+    """Convert an instance of ScaledLSTM to nn.LSTM.
+
+    Args:
+      scaled_lstm:
+        The layer to be converted.
+    Returns:
+      Return an instance of nn.LSTM that has the same `forward()` behavior
+      of the given `scaled_lstm`.
+    """
+    assert isinstance(scaled_lstm, ScaledLSTM), type(scaled_lstm)
+    lstm = nn.LSTM(
+        input_size=scaled_lstm.input_size,
+        hidden_size=scaled_lstm.hidden_size,
+        num_layers=scaled_lstm.num_layers,
+        bias=scaled_lstm.bias,
+        batch_first=scaled_lstm.batch_first,
+        dropout=scaled_lstm.dropout,
+        bidirectional=scaled_lstm.bidirectional,
+        proj_size=scaled_lstm.proj_size,
+    )
+
+    assert lstm._flat_weights_names == scaled_lstm._flat_weights_names
+    for idx in range(len(scaled_lstm._flat_weights_names)):
+        scaled_weight = scaled_lstm._flat_weights[idx] * scaled_lstm._scales[idx].exp()
+        lstm._flat_weights[idx].data.copy_(scaled_weight)
+
+    return lstm
+
+
+# Copied from https://pytorch.org/docs/1.9.0/_modules/torch/nn/modules/module.html#Module.get_submodule  # noqa
+# get_submodule was added to nn.Module at v1.9.0
+def get_submodule(model, target):
+    if target == "":
+        return model
+    atoms: List[str] = target.split(".")
+    mod: torch.nn.Module = model
+    for item in atoms:
+        if not hasattr(mod, item):
+            raise AttributeError(
+                mod._get_name() + " has no " "attribute `" + item + "`"
+            )
+        mod = getattr(mod, item)
+        if not isinstance(mod, torch.nn.Module):
+            raise AttributeError("`" + item + "` is not " "an nn.Module")
+    return mod
+
+
+def convert_scaled_to_non_scaled(
+    model: nn.Module,
+    inplace: bool = False,
+    is_onnx: bool = False,
+):
+    """Convert `ScaledLinear`, `ScaledConv1d`, and `ScaledConv2d`
+    in the given modle to their unscaled version `nn.Linear`, `nn.Conv1d`,
+    and `nn.Conv2d`.
+
+    Args:
+      model:
+        The model to be converted.
+      inplace:
+        If True, the input model is modified inplace.
+        If False, the input model is copied and we modify the copied version.
+      is_onnx:
+        If True, we are going to export the model to ONNX. In this case,
+        we will convert nn.LSTM with proj_size to LSTMP.
+    Return:
+      Return a model without scaled layers.
+    """
+    if not inplace:
+        model = copy.deepcopy(model)
+
+    excluded_patterns = r"self_attn\.(in|out)_proj"
+    p = re.compile(excluded_patterns)
+
+    d = {}
+    for name, m in model.named_modules():
+        if isinstance(m, ScaledLinear):
+            if p.search(name) is not None:
+                continue
+            d[name] = scaled_linear_to_linear(m)
+        elif isinstance(m, ScaledConv1d):
+            d[name] = scaled_conv1d_to_conv1d(m)
+        elif isinstance(m, ScaledConv2d):
+            d[name] = scaled_conv2d_to_conv2d(m)
+        elif isinstance(m, ScaledEmbedding):
+            d[name] = scaled_embedding_to_embedding(m)
+        elif isinstance(m, BasicNorm):
+            d[name] = convert_basic_norm(m)
+        elif isinstance(m, ScaledLSTM):
+            if is_onnx:
+                d[name] = LSTMP(scaled_lstm_to_lstm(m))
+                # See
+                # https://github.com/pytorch/pytorch/issues/47887
+                #  d[name] = torch.jit.script(LSTMP(scaled_lstm_to_lstm(m)))
+            else:
+                d[name] = scaled_lstm_to_lstm(m)
+        elif isinstance(m, ActivationBalancer):
+            d[name] = nn.Identity()
+
+    for k, v in d.items():
+        if "." in k:
+            parent, child = k.rsplit(".", maxsplit=1)
+            setattr(get_submodule(model, parent), child, v)
+        else:
+            setattr(model, k, v)
+
+    return model

egs/librispeech/ASR/pruned_transducer_stateless_d2v/streaming_beam_search.py

@@ -0,0 +1,282 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import warnings
+from typing import List
+
+import k2
+import torch
+import torch.nn as nn
+from beam_search import Hypothesis, HypothesisList, get_hyps_shape
+from decode_stream import DecodeStream
+
+from icefall.decode import one_best_decoding
+from icefall.utils import get_texts
+
+
+def greedy_search(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    streams: List[DecodeStream],
+) -> None:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C), where N >= 1.
+      streams:
+        A list of Stream objects.
+    """
+    assert len(streams) == encoder_out.size(0)
+    assert encoder_out.ndim == 3
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    device = model.device
+    T = encoder_out.size(1)
+
+    decoder_input = torch.tensor(
+        [stream.hyp[-context_size:] for stream in streams],
+        device=device,
+        dtype=torch.int64,
+    )
+    # decoder_out is of shape (N, 1, decoder_out_dim)
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    decoder_out = model.joiner.decoder_proj(decoder_out)
+
+    for t in range(T):
+        # current_encoder_out's shape: (batch_size, 1, encoder_out_dim)
+        current_encoder_out = encoder_out[:, t : t + 1, :]  # noqa
+
+        logits = model.joiner(
+            current_encoder_out.unsqueeze(2),
+            decoder_out.unsqueeze(1),
+            project_input=False,
+        )
+        # logits'shape (batch_size,  vocab_size)
+        logits = logits.squeeze(1).squeeze(1)
+
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                streams[i].hyp.append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = torch.tensor(
+                [stream.hyp[-context_size:] for stream in streams],
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(
+                decoder_input,
+                need_pad=False,
+            )
+            decoder_out = model.joiner.decoder_proj(decoder_out)
+
+
+def modified_beam_search(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    streams: List[DecodeStream],
+    num_active_paths: int = 4,
+) -> None:
+    """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded.
+
+    Args:
+      model:
+        The RNN-T model.
+      encoder_out:
+        A 3-D tensor of shape (N, T, encoder_out_dim) containing the output of
+        the encoder model.
+      streams:
+        A list of stream objects.
+      num_active_paths:
+        Number of active paths during the beam search.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert len(streams) == encoder_out.size(0)
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+    batch_size = len(streams)
+    T = encoder_out.size(1)
+
+    B = [stream.hyps for stream in streams]
+
+    for t in range(T):
+        current_encoder_out = encoder_out[:, t].unsqueeze(1).unsqueeze(1)
+        # current_encoder_out's shape: (batch_size, 1, 1, encoder_out_dim)
+
+        hyps_shape = get_hyps_shape(B).to(device)
+
+        A = [list(b) for b in B]
+        B = [HypothesisList() for _ in range(batch_size)]
+
+        ys_log_probs = torch.stack(
+            [hyp.log_prob.reshape(1) for hyps in A for hyp in hyps], dim=0
+        )  # (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyps in A for hyp in hyps],
+            device=device,
+            dtype=torch.int64,
+        )  # (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # decoder_out is of shape (num_hyps, 1, 1, decoder_output_dim)
+
+        # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor
+        # as index, so we use `to(torch.int64)` below.
+        current_encoder_out = torch.index_select(
+            current_encoder_out,
+            dim=0,
+            index=hyps_shape.row_ids(1).to(torch.int64),
+        )  # (num_hyps, encoder_out_dim)
+
+        logits = model.joiner(current_encoder_out, decoder_out, project_input=False)
+        # logits is of shape (num_hyps, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)
+
+        log_probs = logits.log_softmax(dim=-1)  # (num_hyps, vocab_size)
+
+        log_probs.add_(ys_log_probs)
+
+        vocab_size = log_probs.size(-1)
+
+        log_probs = log_probs.reshape(-1)
+
+        row_splits = hyps_shape.row_splits(1) * vocab_size
+        log_probs_shape = k2.ragged.create_ragged_shape2(
+            row_splits=row_splits, cached_tot_size=log_probs.numel()
+        )
+        ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs)
+
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(num_active_paths)
+
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+                topk_token_indexes = (topk_indexes % vocab_size).tolist()
+
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+
+                new_ys = hyp.ys[:]
+                new_token = topk_token_indexes[k]
+                if new_token != blank_id:
+                    new_ys.append(new_token)
+
+                new_log_prob = topk_log_probs[k]
+                new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob)
+                B[i].add(new_hyp)
+
+    for i in range(batch_size):
+        streams[i].hyps = B[i]
+
+
+def fast_beam_search_one_best(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    processed_lens: torch.Tensor,
+    streams: List[DecodeStream],
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+) -> None:
+    """It limits the maximum number of symbols per frame to 1.
+
+    A lattice is first generated by Fsa-based beam search, then we get the
+    recognition by applying shortest path on the lattice.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      processed_lens:
+        A tensor of shape (N,) containing the number of processed frames
+        in `encoder_out` before padding.
+      streams:
+        A list of stream objects.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+    """
+    assert encoder_out.ndim == 3
+    B, T, C = encoder_out.shape
+    assert B == len(streams)
+
+    context_size = model.decoder.context_size
+    vocab_size = model.decoder.vocab_size
+
+    config = k2.RnntDecodingConfig(
+        vocab_size=vocab_size,
+        decoder_history_len=context_size,
+        beam=beam,
+        max_contexts=max_contexts,
+        max_states=max_states,
+    )
+    individual_streams = []
+    for i in range(B):
+        individual_streams.append(streams[i].rnnt_decoding_stream)
+    decoding_streams = k2.RnntDecodingStreams(individual_streams, config)
+
+    for t in range(T):
+        # shape is a RaggedShape of shape (B, context)
+        # contexts is a Tensor of shape (shape.NumElements(), context_size)
+        shape, contexts = decoding_streams.get_contexts()
+        # `nn.Embedding()` in torch below v1.7.1 supports only torch.int64
+        contexts = contexts.to(torch.int64)
+        # decoder_out is of shape (shape.NumElements(), 1, decoder_out_dim)
+        decoder_out = model.decoder(contexts, need_pad=False)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # current_encoder_out is of shape
+        # (shape.NumElements(), 1, joiner_dim)
+        # fmt: off
+        current_encoder_out = torch.index_select(
+            encoder_out[:, t:t + 1, :], 0, shape.row_ids(1).to(torch.int64)
+        )
+        # fmt: on
+        logits = model.joiner(
+            current_encoder_out.unsqueeze(2),
+            decoder_out.unsqueeze(1),
+            project_input=False,
+        )
+        logits = logits.squeeze(1).squeeze(1)
+        log_probs = logits.log_softmax(dim=-1)
+        decoding_streams.advance(log_probs)
+
+    decoding_streams.terminate_and_flush_to_streams()
+
+    lattice = decoding_streams.format_output(processed_lens.tolist())
+    best_path = one_best_decoding(lattice)
+    hyp_tokens = get_texts(best_path)
+
+    for i in range(B):
+        streams[i].hyp = hyp_tokens[i]

egs/librispeech/ASR/pruned_transducer_stateless_d2v/streaming_decode.py

@@ -0,0 +1,653 @@
+#!/usr/bin/env python3
+# Copyright 2022 Xiaomi Corporation (Authors: Wei Kang, Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+./pruned_transducer_stateless5/streaming_decode.py \
+        --epoch 28 \
+        --avg 15 \
+        --left-context 32 \
+        --decode-chunk-size 8 \
+        --right-context 0 \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --decoding_method greedy_search \
+        --num-decode-streams 200
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from decode_stream import DecodeStream
+from kaldifeat import Fbank, FbankOptions
+from lhotse import CutSet
+from streaming_beam_search import (
+    fast_beam_search_one_best,
+    greedy_search,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Supported decoding methods are:
+        greedy_search
+        modified_beam_search
+        fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--num_active_paths",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=32,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-size",
+        type=int,
+        default=16,
+        help="The chunk size for decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--left-context",
+        type=int,
+        default=64,
+        help="left context can be seen during decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--right-context",
+        type=int,
+        default=0,
+        help="right context can be seen during decoding (in frames after subsampling)",
+    )
+
+    parser.add_argument(
+        "--num-decode-streams",
+        type=int,
+        default=2000,
+        help="The number of streams that can be decoded parallel.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_chunk(
+    params: AttributeDict,
+    model: nn.Module,
+    decode_streams: List[DecodeStream],
+) -> List[int]:
+    """Decode one chunk frames of features for each decode_streams and
+    return the indexes of finished streams in a List.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      decode_streams:
+        A List of DecodeStream, each belonging to a utterance.
+    Returns:
+      Return a List containing which DecodeStreams are finished.
+    """
+    device = model.device
+
+    features = []
+    feature_lens = []
+    states = []
+
+    processed_lens = []
+
+    for stream in decode_streams:
+        feat, feat_len = stream.get_feature_frames(
+            params.decode_chunk_size * params.subsampling_factor
+        )
+        features.append(feat)
+        feature_lens.append(feat_len)
+        states.append(stream.states)
+        processed_lens.append(stream.done_frames)
+
+    feature_lens = torch.tensor(feature_lens, device=device)
+    features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS)
+
+    # if T is less than 7 there will be an error in time reduction layer,
+    # because we subsample features with ((x_len - 1) // 2 - 1) // 2
+    # we plus 2 here because we will cut off one frame on each size of
+    # encoder_embed output as they see invalid paddings. so we need extra 2
+    # frames.
+    tail_length = 7 + (2 + params.right_context) * params.subsampling_factor
+    if features.size(1) < tail_length:
+        pad_length = tail_length - features.size(1)
+        feature_lens += pad_length
+        features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, pad_length),
+            mode="constant",
+            value=LOG_EPS,
+        )
+
+    states = [
+        torch.stack([x[0] for x in states], dim=2),
+        torch.stack([x[1] for x in states], dim=2),
+    ]
+    processed_lens = torch.tensor(processed_lens, device=device)
+
+    encoder_out, encoder_out_lens, states = model.encoder.streaming_forward(
+        x=features,
+        x_lens=feature_lens,
+        states=states,
+        left_context=params.left_context,
+        right_context=params.right_context,
+        processed_lens=processed_lens,
+    )
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    if params.decoding_method == "greedy_search":
+        greedy_search(model=model, encoder_out=encoder_out, streams=decode_streams)
+    elif params.decoding_method == "fast_beam_search":
+        processed_lens = processed_lens + encoder_out_lens
+        fast_beam_search_one_best(
+            model=model,
+            encoder_out=encoder_out,
+            processed_lens=processed_lens,
+            streams=decode_streams,
+            beam=params.beam,
+            max_states=params.max_states,
+            max_contexts=params.max_contexts,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        modified_beam_search(
+            model=model,
+            streams=decode_streams,
+            encoder_out=encoder_out,
+            num_active_paths=params.num_active_paths,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    states = [torch.unbind(states[0], dim=2), torch.unbind(states[1], dim=2)]
+
+    finished_streams = []
+    for i in range(len(decode_streams)):
+        decode_streams[i].states = [states[0][i], states[1][i]]
+        decode_streams[i].done_frames += encoder_out_lens[i]
+        if decode_streams[i].done:
+            finished_streams.append(i)
+
+    return finished_streams
+
+
+def decode_dataset(
+    cuts: CutSet,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      cuts:
+        Lhotse Cutset containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    device = model.device
+
+    opts = FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    log_interval = 50
+
+    decode_results = []
+    # Contain decode streams currently running.
+    decode_streams = []
+    initial_states = model.encoder.get_init_state(params.left_context, device=device)
+    for num, cut in enumerate(cuts):
+        # each utterance has a DecodeStream.
+        decode_stream = DecodeStream(
+            params=params,
+            cut_id=cut.id,
+            initial_states=initial_states,
+            decoding_graph=decoding_graph,
+            device=device,
+        )
+
+        audio: np.ndarray = cut.load_audio()
+        # audio.shape: (1, num_samples)
+        assert len(audio.shape) == 2
+        assert audio.shape[0] == 1, "Should be single channel"
+        assert audio.dtype == np.float32, audio.dtype
+
+        # The trained model is using normalized samples
+        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+
+        samples = torch.from_numpy(audio).squeeze(0)
+
+        fbank = Fbank(opts)
+        feature = fbank(samples.to(device))
+        decode_stream.set_features(feature)
+        decode_stream.ground_truth = cut.supervisions[0].text
+
+        decode_streams.append(decode_stream)
+
+        while len(decode_streams) >= params.num_decode_streams:
+            finished_streams = decode_one_chunk(
+                params=params, model=model, decode_streams=decode_streams
+            )
+            for i in sorted(finished_streams, reverse=True):
+                decode_results.append(
+                    (
+                        decode_streams[i].id,
+                        decode_streams[i].ground_truth.split(),
+                        sp.decode(decode_streams[i].decoding_result()).split(),
+                    )
+                )
+                del decode_streams[i]
+
+        if num % log_interval == 0:
+            logging.info(f"Cuts processed until now is {num}.")
+
+    # decode final chunks of last sequences
+    while len(decode_streams):
+        finished_streams = decode_one_chunk(
+            params=params, model=model, decode_streams=decode_streams
+        )
+        for i in sorted(finished_streams, reverse=True):
+            decode_results.append(
+                (
+                    decode_streams[i].id,
+                    decode_streams[i].ground_truth.split(),
+                    sp.decode(decode_streams[i].decoding_result()).split(),
+                )
+            )
+            del decode_streams[i]
+
+    if params.decoding_method == "greedy_search":
+        key = "greedy_search"
+    elif params.decoding_method == "fast_beam_search":
+        key = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}"
+        )
+    elif params.decoding_method == "modified_beam_search":
+        key = f"num_active_paths_{params.num_active_paths}"
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+    return {key: decode_results}
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    # for streaming
+    params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
+    params.suffix += f"-left-context-{params.left_context}"
+    params.suffix += f"-right-context-{params.right_context}"
+
+    # for fast_beam_search
+    if params.decoding_method == "fast_beam_search":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    # Decoding in streaming requires causal convolution
+    params.causal_convolution = True
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if start >= 0:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    decoding_graph = None
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_sets = ["test-clean", "test-other"]
+    test_cuts = [test_clean_cuts, test_other_cuts]
+
+    for test_set, test_cut in zip(test_sets, test_cuts):
+        results_dict = decode_dataset(
+            cuts=test_cut,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()

egs/librispeech/ASR/pruned_transducer_stateless_d2v/test_model.py

@@ -0,0 +1,65 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./pruned_transducer_stateless4/test_model.py
+"""
+
+from train import get_params, get_transducer_model
+
+
+def test_model_1():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.num_encoder_layers = 24
+    params.dim_feedforward = 1536  # 384 * 4
+    params.encoder_dim = 384
+    model = get_transducer_model(params)
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+
+# See Table 1 from https://arxiv.org/pdf/2005.08100.pdf
+def test_model_M():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.num_encoder_layers = 18
+    params.dim_feedforward = 1024
+    params.encoder_dim = 256
+    params.nhead = 4
+    params.decoder_dim = 512
+    params.joiner_dim = 512
+    model = get_transducer_model(params)
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+
+def main():
+    #  test_model_1()
+    test_model_M()
+
+
+if __name__ == "__main__":
+    main()

egs/librispeech/ASR/pruned_transducer_stateless_d2v/wav2vec2_encoder.py

@@ -0,0 +1,168 @@
+# Copyright 2021 Xuankai Chang
+#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
+
+"""Encoder definition."""
+import contextlib
+import copy
+import logging
+import os
+from typing import Optional, Tuple
+
+import torch
+from filelock import FileLock
+from typeguard import check_argument_types
+
+from espnet2.asr.encoder.abs_encoder import AbsEncoder
+from espnet.nets.pytorch_backend.nets_utils import make_pad_mask
+from espnet.nets.pytorch_backend.transformer.layer_norm import LayerNorm
+
+
+class FairSeqWav2Vec2Encoder(AbsEncoder):
+    """FairSeq Wav2Vec2 encoder module.
+
+    Args:
+        input_size: input dim
+        output_size: dimension of attention
+        w2v_url: url to Wav2Vec2.0 pretrained model
+        w2v_dir_path: directory to download the Wav2Vec2.0 pretrained model.
+        normalize_before: whether to use layer_norm before the first block
+        finetune_last_n_layers: last n layers to be finetuned in Wav2Vec2.0
+                                0 means to finetune every layer if freeze_w2v=False.
+    """
+
+    def __init__(
+        self,
+        input_size: int,
+        w2v_url: str,
+        w2v_dir_path: str = "./",
+        output_size: int = 256,
+        normalize_before: bool = False,
+        freeze_finetune_updates: int = 0,
+    ):
+        assert check_argument_types()
+        super().__init__()
+
+        if w2v_url != "":
+            try:
+                import fairseq
+                from fairseq.models.wav2vec.wav2vec2 import Wav2Vec2Model
+            except Exception as e:
+                print("Error: FairSeq is not properly installed.")
+                print(
+                    "Please install FairSeq: cd ${MAIN_ROOT}/tools && make fairseq.done"
+                )
+                raise e
+
+        self.w2v_model_path = download_w2v(w2v_url, w2v_dir_path)
+
+        self._output_size = output_size
+
+        models, _, _ = fairseq.checkpoint_utils.load_model_ensemble_and_task(
+            [self.w2v_model_path],
+            arg_overrides={"data": w2v_dir_path},
+        )
+        model = models[0]
+
+        if not isinstance(model, Wav2Vec2Model):
+            try:
+                model = model.w2v_encoder.w2v_model
+            except Exception as e:
+                print(
+                    "Error: pretrained models should be within: "
+                    "'Wav2Vec2Model, Wav2VecCTC' classes, etc."
+                )
+                raise e
+
+        self.encoders = model
+
+        self.pretrained_params = copy.deepcopy(model.state_dict())
+
+        self.normalize_before = normalize_before
+        if self.normalize_before:
+            self.after_norm = LayerNorm(output_size)
+
+        if model.cfg.encoder_embed_dim != output_size:
+            # TODO(xkc09): try LSTM
+            self.output_layer = torch.nn.Sequential(
+                torch.nn.Linear(model.cfg.encoder_embed_dim, output_size),
+            )
+        else:
+            self.output_layer = None
+
+        self.freeze_finetune_updates = freeze_finetune_updates
+        self.register_buffer("num_updates", torch.LongTensor([0]))
+
+    def output_size(self) -> int:
+        return self._output_size
+
+    def forward(
+        self,
+        xs_pad: torch.Tensor,
+        ilens: torch.Tensor,
+        prev_states: torch.Tensor = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
+        """Forward FairSeqWav2Vec2 Encoder.
+
+        Args:
+            xs_pad: input tensor (B, L, D)
+            ilens: input length (B)
+            prev_states: Not to be used now.
+        Returns:
+            position embedded tensor and mask
+        """
+        masks = make_pad_mask(ilens).to(xs_pad.device)
+
+        ft = self.freeze_finetune_updates <= self.num_updates
+        if self.num_updates <= self.freeze_finetune_updates:
+            self.num_updates += 1
+        elif ft and self.num_updates == self.freeze_finetune_updates + 1:
+            self.num_updates += 1
+            logging.info("Start fine-tuning wav2vec parameters!")
+
+        with torch.no_grad() if not ft else contextlib.nullcontext():
+            enc_outputs = self.encoders(
+                xs_pad,
+                masks,
+                mask=self.training,
+                features_only=True,
+            )
+
+        xs_pad = enc_outputs["x"]  # (B,T,C),
+        bs = xs_pad.shape[0]
+        if enc_outputs["padding_mask"] is not None:
+            masks = enc_outputs["padding_mask"]  # (B, T)
+            olens = (~masks).sum(dim=1)  # (B)
+        else:
+            olens = torch.IntTensor([xs_pad.shape[1]]).repeat(bs).to(xs_pad.device)
+
+        if self.output_layer is not None:
+            xs_pad = self.output_layer(xs_pad)
+
+        if self.normalize_before:
+            xs_pad = self.after_norm(xs_pad)
+
+        return xs_pad, olens, None
+
+    def reload_pretrained_parameters(self):
+        self.encoders.load_state_dict(self.pretrained_params)
+        logging.info("Pretrained Wav2Vec model parameters reloaded!")
+
+
+def download_w2v(model_url, dir_path):
+    os.makedirs(dir_path, exist_ok=True)
+
+    model_name = model_url.split("/")[-1]
+    model_path = os.path.join(dir_path, model_name)
+
+    dict_url = "https://dl.fbaipublicfiles.com/fairseq/wav2vec/dict.ltr.txt"
+    dict_path = os.path.join(dir_path, dict_url.split("/")[-1])
+
+    with FileLock(model_path + ".lock"):
+        if not os.path.exists(model_path):
+            torch.hub.download_url_to_file(model_url, model_path)
+            torch.hub.download_url_to_file(dict_url, dict_path)
+            logging.info(f"Wav2Vec model downloaded {model_path}")
+        else:
+            logging.info(f"Wav2Vec model {model_path} already exists.")
+
+    return model_path

git.sh

@@ -0,0 +1,6 @@
+git add *
+git commit -m "from local"
+git push
+#git push --set-upstream origin summit
+
+echo "git push done"

watchdog_v2.py

@@ -0,0 +1,21 @@
+import os 
+import time 
+
+lastmod = []
+dirToWatchList = []
+path = './'
+for dirpath in os.walk(path):
+    if '.git' in dirpath[0]:
+        continue
+    dirToWatch = dirpath[0]+'/'
+    lastmod.append(int(os.path.getmtime(dirToWatch)))
+    dirToWatchList.append(dirToWatch)
+#lastmod = int(os.path.getmtime(dirToWatch))
+
+while True:
+    for enum, dirToWatch in enumerate(dirToWatchList):
+        if lastmod[enum] != int(os.path.getmtime(dirToWatch)): 
+            #print('Warning: Modify Detected.') 
+            os.system('./git.sh')
+            lastmod[enum] = int(os.path.getmtime(dirToWatch)) 
+    time.sleep(1.0)