Merge bb58c85c839d9d0766fc57e217788aa17a7f85cb into 800bf4b6a2e32745e7d0c31dd78d473f1faff509

egs/librispeech/ASR/RESULTS.md

@@ -375,6 +375,38 @@ for m in greedy_search modified_beam_search fast_beam_search; do
 done
 ```
 
+### zipformer ctc streaming
+
+| decoding method      | test-clean | test-other | comment            |
+|----------------------|------------|------------|--------------------|
+| greedy_search        | 4.07       | 10.51      | --epoch 30 --avg 15|
+| greedy_search        | 4.0        | 10.3       | --epoch 30 --avg 9 |
+
+The training command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1"
+python ./zipformer_ctc_streaming/train.py \
+  --world-size 2   \
+  --num-epochs 30  \
+  --start-epoch 1  \
+  --exp-dir ./zipformer_ctc_streaming/exp \
+  --max-duration 100 
+```
+
+The decoding command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0"
+./zipformer_ctc_streaming/decode.py \
+--epoch 30  \
+--avg 15 \
+--exp-dir ./zipformer_ctc_streaming/exp \
+--max-duration 300 \ 
+--decode-chunk-len 32  \
+--method ctc-decoding \
+--lm-dir data/lm \
+--lang-dir  data/lang_bpe_500
+```
+
 ### pruned_transducer_stateless7 (Fine-tune with mux)
 
 See <https://github.com/k2-fsa/icefall/pull/1059> for more details.
@@ -456,6 +488,39 @@ The decoding commands are:
     --max-states 64
 ```
 
+### zipformer ctc streaming
+
+| decoding method      | test-clean | test-other | comment            |
+|----------------------|------------|------------|--------------------|
+| greedy_search        | 4.07       | 10.51      | --epoch 30 --avg 15|
+| greedy_search        | 4.0        | 10.3       | --epoch 30 --avg 9 |
+
+The training command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1"
+python ./zipformer_ctc_streaming/train.py \
+  --world-size 2   \
+  --num-epochs 30  \
+  --start-epoch 1  \
+  --exp-dir ./zipformer_ctc_streaming/exp \
+  --max-duration 100 
+```
+
+The decoding command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0"
+./zipformer_ctc_streaming/decode.py \
+--epoch 30  \
+--avg 15 \
+--exp-dir ./zipformer_ctc_streaming/exp \
+--max-duration 300 \ 
+--decode-chunk-len 32  \
+--method ctc-decoding \
+--lm-dir data/lm \
+--lang-dir  data/lang_bpe_500
+```
+
+
 ### pruned_transducer_stateless7 (zipformer + multidataset(LibriSpeech + GigaSpeech + CommonVoice 13.0))
 
 See <https://github.com/k2-fsa/icefall/pull/1010> for more details.

egs/librispeech/ASR/zipformer_ctc_streaming/asr_datamodule.py

@@ -0,0 +1,475 @@
+# 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="""Used only when --mini-libri is False.When enabled,
+            use 960h LibriSpeech. Otherwise, use 100h subset.""",
+        )
+        group.add_argument(
+            "--mini-libri",
+            type=str2bool,
+            default=False,
+            help="True for mini librispeech",
+        )
+
+        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(
+                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_5_cuts(self) -> CutSet:
+        logging.info("mini_librispeech: About to get train-clean-5 cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_train-clean-5.jsonl.gz"
+        )
+
+    @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_2_cuts(self) -> CutSet:
+        logging.info("mini_librispeech: About to get dev-clean-2 cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_dev-clean-2.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/zipformer_ctc_streaming/decode.py

@@ -0,0 +1,891 @@
+#!/usr/bin/env python3
+# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo, 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 argparse
+import logging
+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 train import add_model_arguments, get_ctc_model, get_params
+from transformer import encoder_padding_mask
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_attention_decoder,
+    rescore_with_n_best_list,
+    rescore_with_rnn_lm,
+    rescore_with_whole_lattice,
+)
+from icefall.lexicon import Lexicon
+from icefall.rnn_lm.model import RnnLmModel
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    load_averaged_model,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=77,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    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=55,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    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(
+        "--method",
+        type=str,
+        default="attention-decoder",
+        help="""Decoding method.
+        Supported values are:
+            - (0) ctc-decoding. Use CTC decoding. It uses a sentence piece
+              model, i.e., lang_dir/bpe.model, to convert word pieces to words.
+              It needs neither a lexicon nor an n-gram LM.
+            - (1) 1best. Extract the best path from the decoding lattice as the
+              decoding result.
+            - (2) nbest. Extract n paths from the decoding lattice; the path
+              with the highest score is the decoding result.
+            - (3) nbest-rescoring. Extract n paths from the decoding lattice,
+              rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+              the highest score is the decoding result.
+            - (4) whole-lattice-rescoring. Rescore the decoding lattice with an
+              n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+              is the decoding result.
+            - (5) attention-decoder. Extract n paths from the LM rescored
+              lattice, the path with the highest score is the decoding result.
+            - (6) rnn-lm. Rescoring with attention-decoder and RNN LM. We assume
+              you have trained an RNN LM using ./rnn_lm/train.py
+            - (7) nbest-oracle. Its WER is the lower bound of any n-best
+              rescoring method can achieve. Useful for debugging n-best
+              rescoring method.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, attention-decoder, rnn-lm, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, attention-decoder, rnn-lm, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer_ctc/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--lm-dir",
+        type=str,
+        default="data/lm",
+        help="""The n-gram LM dir.
+        It should contain either G_3_gram.pt or G_3_gram.fst.txt
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-exp-dir",
+        type=str,
+        default="rnn_lm/exp",
+        help="""Used only when --method is rnn-lm.
+        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 rnn-lm.
+        It specifies the checkpoint to use.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-avg",
+        type=int,
+        default=2,
+        help="""Used only when --method is rnn-lm.
+        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,
+    rnn_lm_model: Optional[nn.Module],
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> 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 no rescoring is used, the key is the string `no_rescore`.
+               If LM rescoring is used, the key is the string `lm_scale_xxx`,
+               where `xxx` is the value of `lm_scale`. An example key is
+               `lm_scale_0.7`
+        - 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`.
+
+        - params.method is "1best", it uses 1best decoding without LM rescoring.
+        - params.method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      rnn_lm_model:
+        The neural model for RNN LM.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      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.
+      sos_id:
+        The token ID of the SOS.
+      eos_id:
+        The token ID of the EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    nnet_output, _ = model.encoder(feature, feature_lens)
+    ctc_output = model.ctc_output(nnet_output)
+    # nnet_output is (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"] // params.subsampling_factor,
+            supervisions["num_frames"] // params.subsampling_factor,
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        assert bpe_model is not None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=ctc_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        key = "ctc-decoding"
+        return {key: hyps}
+
+    if params.method == "nbest-oracle":
+        # Note: You can also pass rescored lattices to it.
+        # We choose the HLG decoded lattice for speed reasons
+        # as HLG decoding is faster and the oracle WER
+        # is only slightly worse than that of rescored lattices.
+        best_path = nbest_oracle(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            ref_texts=supervisions["text"],
+            word_table=word_table,
+            nbest_scale=params.nbest_scale,
+            oov="<UNK>",
+        )
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}"  # noqa
+        return {key: hyps}
+
+    if params.method in ["1best", "nbest"]:
+        if params.method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}"  # noqa
+
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.method in [
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "attention-decoder",
+        "rnn-lm",
+    ]
+
+    lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+
+    nnet_output = nnet_output.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+    mask = encoder_padding_mask(nnet_output.size(0), supervisions)
+    mask = mask.to(nnet_output.device) if mask is not None else None
+    mmodel = model.decoder.module if hasattr(model.decoder, "module") else model.decoder
+
+    if params.method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.method == "whole-lattice-rescoring":
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+    elif params.method == "attention-decoder":
+        # lattice uses a 3-gram Lm. We rescore it with a 4-gram LM.
+        rescored_lattice = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=None,
+        )
+
+        best_path_dict = rescore_with_attention_decoder(
+            lattice=rescored_lattice,
+            num_paths=params.num_paths,
+            model=mmodel,
+            memory=nnet_output,
+            memory_key_padding_mask=mask,
+            sos_id=sos_id,
+            eos_id=eos_id,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.method == "rnn-lm":
+        # lattice uses a 3-gram Lm. We rescore it with a 4-gram LM.
+        rescored_lattice = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=None,
+        )
+
+        best_path_dict = rescore_with_rnn_lm(
+            lattice=rescored_lattice,
+            num_paths=params.num_paths,
+            rnn_lm_model=rnn_lm_model,
+            model=mmodel,
+            memory=nnet_output,
+            memory_key_padding_mask=mask,
+            sos_id=sos_id,
+            eos_id=eos_id,
+            blank_id=0,
+            nbest_scale=params.nbest_scale,
+        )
+    else:
+        assert False, f"Unsupported decoding method: {params.method}"
+
+    ans = dict()
+    if best_path_dict is not None:
+        for lm_scale_str, best_path in best_path_dict.items():
+            hyps = get_texts(best_path)
+            hyps = [[word_table[i] for i in ids] for ids in hyps]
+            ans[lm_scale_str] = hyps
+    else:
+        ans = None
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    rnn_lm_model: Optional[nn.Module],
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> 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.
+      rnn_lm_model:
+        The neural model for RNN LM.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      sos_id:
+        The token ID for SOS.
+      eos_id:
+        The token ID for EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring 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 = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            rnn_lm_model=rnn_lm_model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        if hyps_dict is not None:
+            for lm_scale, 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[lm_scale].extend(this_batch)
+        else:
+            assert len(results) > 0, "It should not decode to empty in the first batch!"
+            this_batch = []
+            hyp_words = []
+            for ref_text in texts:
+                ref_words = ref_text.split()
+                this_batch.append((ref_words, hyp_words))
+
+            for lm_scale in results.keys():
+                results[lm_scale].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 100 == 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[int], List[int]]]],
+):
+    if params.method in ("attention-decoder", "rnn-lm"):
+        # Set it to False since there are too many logs.
+        enable_log = False
+    else:
+        enable_log = True
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.exp_dir / f"recogs-{test_set_name}-{key}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        if enable_log:
+            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.exp_dir / f"errs-{test_set_name}-{key}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=enable_log
+            )
+            test_set_wers[key] = wer
+
+        if enable_log:
+            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.exp_dir / f"wer-summary-{test_set_name}.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)
+    args.lang_dir = Path(args.lang_dir)
+    args.lm_dir = Path(args.lm_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log-{params.method}/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+    params.vocab_size = num_classes
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    graph_compiler = BpeCtcTrainingGraphCompiler(
+        params.lang_dir,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+    sos_id = graph_compiler.sos_id
+    eos_id = graph_compiler.eos_id
+
+    params.num_classes = num_classes
+    params.sos_id = sos_id
+    params.eos_id = eos_id
+
+    if params.method == "ctc-decoding":
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(str(params.lang_dir / "bpe.model"))
+    else:
+        H = None
+        bpe_model = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
+        )
+        assert HLG.requires_grad is False
+
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    if params.method in (
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "attention-decoder",
+        "rnn-lm",
+    ):
+        if not (params.lm_dir / "G_3_gram.pt").is_file():
+            logging.info("Loading G_3_gram.fst.txt")
+            logging.warning("It may take 8 minutes.")
+            with open(params.lm_dir / "G_3_gram.fst.txt") as f:
+            #with open(params.lm_dir / "G_3_gram.fst.txt") as f:
+                first_word_disambig_id = lexicon.word_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_word_disambig_id] = 0
+                # See https://github.com/k2-fsa/k2/issues/874
+                # for why we need to set G.properties to None
+                G.__dict__["_properties"] = None
+                G = k2.Fsa.from_fsas([G]).to(device)
+                G = k2.arc_sort(G)
+                # Save a dummy value so that it can be loaded in C++.
+                # See https://github.com/pytorch/pytorch/issues/67902
+                # for why we need to do this.
+                G.dummy = 1
+
+                torch.save(G.as_dict(), params.lm_dir / "G_3_gram.pt")
+        else:
+            logging.info("Loading pre-compiled G_3_gram.pt")
+            d = torch.load(params.lm_dir / "G_3_gram.pt", map_location=device)
+            G = k2.Fsa.from_dict(d)
+
+        if params.method in [
+            "whole-lattice-rescoring",
+            "attention-decoder",
+            "rnn-lm",
+        ]:
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    logging.info("About to create model")
+    model = get_ctc_model(params)
+    assert model.encoder.decode_chunk_size == params.decode_chunk_len // 2, (
+        model.encoder.decode_chunk_size,
+        params.decode_chunk_len,
+    )
+
+    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()
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    rnn_lm_model = None
+    if params.method == "rnn-lm":
+        rnn_lm_model = RnnLmModel(
+            vocab_size=params.num_classes,
+            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,
+        )
+        if 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)
+        else:
+            rnn_lm_model = load_averaged_model(
+                params.rnn_lm_exp_dir,
+                rnn_lm_model,
+                params.rnn_lm_epoch,
+                params.rnn_lm_avg,
+                device,
+            )
+        rnn_lm_model.eval()
+
+    # 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_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,
+            rnn_lm_model=rnn_lm_model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        save_results(params=params, test_set_name=test_set, results_dict=results_dict)
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()

egs/librispeech/ASR/zipformer_ctc_streaming/decode_G_4gram.py

@@ -0,0 +1,887 @@
+#!/usr/bin/env python3
+# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo, 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 argparse
+import logging
+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 train import add_model_arguments, get_ctc_model, get_params
+from transformer import encoder_padding_mask
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_attention_decoder,
+    rescore_with_n_best_list,
+    rescore_with_rnn_lm,
+    rescore_with_whole_lattice,
+)
+from icefall.lexicon import Lexicon
+from icefall.rnn_lm.model import RnnLmModel
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    load_averaged_model,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=77,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    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=55,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    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(
+        "--method",
+        type=str,
+        default="attention-decoder",
+        help="""Decoding method.
+        Supported values are:
+            - (0) ctc-decoding. Use CTC decoding. It uses a sentence piece
+              model, i.e., lang_dir/bpe.model, to convert word pieces to words.
+              It needs neither a lexicon nor an n-gram LM.
+            - (1) 1best. Extract the best path from the decoding lattice as the
+              decoding result.
+            - (2) nbest. Extract n paths from the decoding lattice; the path
+              with the highest score is the decoding result.
+            - (3) nbest-rescoring. Extract n paths from the decoding lattice,
+              rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+              the highest score is the decoding result.
+            - (4) whole-lattice-rescoring. Rescore the decoding lattice with an
+              n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+              is the decoding result.
+            - (5) attention-decoder. Extract n paths from the LM rescored
+              lattice, the path with the highest score is the decoding result.
+            - (6) rnn-lm. Rescoring with attention-decoder and RNN LM. We assume
+              you have trained an RNN LM using ./rnn_lm/train.py
+            - (7) nbest-oracle. Its WER is the lower bound of any n-best
+              rescoring method can achieve. Useful for debugging n-best
+              rescoring method.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, attention-decoder, rnn-lm, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, attention-decoder, rnn-lm, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer_ctc/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--lm-dir",
+        type=str,
+        default="data/lm",
+        help="""The n-gram LM dir.
+        It should contain either G_4_gram.pt or G_4_gram.fst.txt
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-exp-dir",
+        type=str,
+        default="rnn_lm/exp",
+        help="""Used only when --method is rnn-lm.
+        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 rnn-lm.
+        It specifies the checkpoint to use.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-avg",
+        type=int,
+        default=2,
+        help="""Used only when --method is rnn-lm.
+        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,
+    rnn_lm_model: Optional[nn.Module],
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> 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 no rescoring is used, the key is the string `no_rescore`.
+               If LM rescoring is used, the key is the string `lm_scale_xxx`,
+               where `xxx` is the value of `lm_scale`. An example key is
+               `lm_scale_0.7`
+        - 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`.
+
+        - params.method is "1best", it uses 1best decoding without LM rescoring.
+        - params.method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      rnn_lm_model:
+        The neural model for RNN LM.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      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.
+      sos_id:
+        The token ID of the SOS.
+      eos_id:
+        The token ID of the EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    nnet_output, _ = model.encoder(feature, feature_lens)
+    ctc_output = model.ctc_output(nnet_output)
+    # nnet_output is (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"] // params.subsampling_factor,
+            supervisions["num_frames"] // params.subsampling_factor,
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        assert bpe_model is not None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=ctc_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        key = "ctc-decoding"
+        return {key: hyps}
+
+    if params.method == "nbest-oracle":
+        # Note: You can also pass rescored lattices to it.
+        # We choose the HLG decoded lattice for speed reasons
+        # as HLG decoding is faster and the oracle WER
+        # is only slightly worse than that of rescored lattices.
+        best_path = nbest_oracle(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            ref_texts=supervisions["text"],
+            word_table=word_table,
+            nbest_scale=params.nbest_scale,
+            oov="<UNK>",
+        )
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}"  # noqa
+        return {key: hyps}
+
+    if params.method in ["1best", "nbest"]:
+        if params.method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}"  # noqa
+
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.method in [
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "attention-decoder",
+        "rnn-lm",
+    ]
+
+    lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+
+    nnet_output = nnet_output.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+    mask = encoder_padding_mask(nnet_output.size(0), supervisions)
+    mask = mask.to(nnet_output.device) if mask is not None else None
+    mmodel = model.decoder.module if hasattr(model.decoder, "module") else model.decoder
+
+    if params.method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.method == "whole-lattice-rescoring":
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+    elif params.method == "attention-decoder":
+        # lattice uses a 3-gram Lm. We rescore it with a 4-gram LM.
+        rescored_lattice = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=None,
+        )
+
+        best_path_dict = rescore_with_attention_decoder(
+            lattice=rescored_lattice,
+            num_paths=params.num_paths,
+            model=mmodel,
+            memory=nnet_output,
+            memory_key_padding_mask=mask,
+            sos_id=sos_id,
+            eos_id=eos_id,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.method == "rnn-lm":
+        # lattice uses a 3-gram Lm. We rescore it with a 4-gram LM.
+        rescored_lattice = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=None,
+        )
+
+        best_path_dict = rescore_with_rnn_lm(
+            lattice=rescored_lattice,
+            num_paths=params.num_paths,
+            rnn_lm_model=rnn_lm_model,
+            model=mmodel,
+            memory=nnet_output,
+            memory_key_padding_mask=mask,
+            sos_id=sos_id,
+            eos_id=eos_id,
+            blank_id=0,
+            nbest_scale=params.nbest_scale,
+        )
+    else:
+        assert False, f"Unsupported decoding method: {params.method}"
+
+    ans = dict()
+    if best_path_dict is not None:
+        for lm_scale_str, best_path in best_path_dict.items():
+            hyps = get_texts(best_path)
+            hyps = [[word_table[i] for i in ids] for ids in hyps]
+            ans[lm_scale_str] = hyps
+    else:
+        ans = None
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    rnn_lm_model: Optional[nn.Module],
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> 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.
+      rnn_lm_model:
+        The neural model for RNN LM.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      sos_id:
+        The token ID for SOS.
+      eos_id:
+        The token ID for EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring 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 = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            rnn_lm_model=rnn_lm_model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        if hyps_dict is not None:
+            for lm_scale, 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[lm_scale].extend(this_batch)
+        else:
+            assert len(results) > 0, "It should not decode to empty in the first batch!"
+            this_batch = []
+            hyp_words = []
+            for ref_text in texts:
+                ref_words = ref_text.split()
+                this_batch.append((ref_words, hyp_words))
+
+            for lm_scale in results.keys():
+                results[lm_scale].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 100 == 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[int], List[int]]]],
+):
+    if params.method in ("attention-decoder", "rnn-lm"):
+        # Set it to False since there are too many logs.
+        enable_log = False
+    else:
+        enable_log = True
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.exp_dir / f"recogs-{test_set_name}-{key}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        if enable_log:
+            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.exp_dir / f"errs-{test_set_name}-{key}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=enable_log
+            )
+            test_set_wers[key] = wer
+
+        if enable_log:
+            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.exp_dir / f"wer-summary-{test_set_name}.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)
+    args.lang_dir = Path(args.lang_dir)
+    args.lm_dir = Path(args.lm_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log-{params.method}/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+    params.vocab_size = num_classes
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    graph_compiler = BpeCtcTrainingGraphCompiler(
+        params.lang_dir,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+    sos_id = graph_compiler.sos_id
+    eos_id = graph_compiler.eos_id
+
+    params.num_classes = num_classes
+    params.sos_id = sos_id
+    params.eos_id = eos_id
+
+    if params.method == "ctc-decoding":
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(str(params.lang_dir / "bpe.model"))
+    else:
+        H = None
+        bpe_model = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
+        )
+        assert HLG.requires_grad is False
+
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    if params.method in (
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "attention-decoder",
+        "rnn-lm",
+    ):
+        if not (params.lm_dir / "G_4_gram.pt").is_file():
+            logging.info("Loading G_4_gram.fst.txt")
+            logging.warning("It may take 8 minutes.")
+            with open(params.lm_dir / "G_4_gram.fst.txt") as f:
+            #with open(params.lm_dir / "G_3_gram.fst.txt") as f:
+                first_word_disambig_id = lexicon.word_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_word_disambig_id] = 0
+                # See https://github.com/k2-fsa/k2/issues/874
+                # for why we need to set G.properties to None
+                G.__dict__["_properties"] = None
+                G = k2.Fsa.from_fsas([G]).to(device)
+                G = k2.arc_sort(G)
+                # Save a dummy value so that it can be loaded in C++.
+                # See https://github.com/pytorch/pytorch/issues/67902
+                # for why we need to do this.
+                G.dummy = 1
+
+                torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt")
+        else:
+            logging.info("Loading pre-compiled G_4_gram.pt")
+            d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device)
+            G = k2.Fsa.from_dict(d)
+
+        if params.method in [
+            "whole-lattice-rescoring",
+            "attention-decoder",
+            "rnn-lm",
+        ]:
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    logging.info("About to create model")
+    model = get_ctc_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()
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    rnn_lm_model = None
+    if params.method == "rnn-lm":
+        rnn_lm_model = RnnLmModel(
+            vocab_size=params.num_classes,
+            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,
+        )
+        if 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)
+        else:
+            rnn_lm_model = load_averaged_model(
+                params.rnn_lm_exp_dir,
+                rnn_lm_model,
+                params.rnn_lm_epoch,
+                params.rnn_lm_avg,
+                device,
+            )
+        rnn_lm_model.eval()
+
+    # 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_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,
+            rnn_lm_model=rnn_lm_model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        save_results(params=params, test_set_name=test_set, results_dict=results_dict)
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()

egs/librispeech/ASR/zipformer_ctc_streaming/decoder.py

@@ -0,0 +1,298 @@
+# 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 List
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from label_smoothing import LabelSmoothingLoss
+from torch.nn.utils.rnn import pad_sequence
+from transformer import PositionalEncoding, TransformerDecoderLayer
+
+
+class Decoder(nn.Module):
+    """This class implements Transformer based decoder for an attention-based encoder-decoder
+    model.
+    """
+
+    def __init__(
+        self,
+        num_layers: int,
+        num_classes: int,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        normalize_before: bool = True,
+    ):
+        """
+        Args:
+          num_layers:
+            Number of layers.
+          num_classes:
+            Number of tokens of the modeling unit including blank.
+          d_model:
+            Dimension of the input embedding, and of the decoder output.
+        """
+        super().__init__()
+
+        if num_layers > 0:
+            self.decoder_num_class = num_classes  # bpe model already has sos/eos symbol
+
+            self.decoder_embed = nn.Embedding(
+                num_embeddings=self.decoder_num_class, embedding_dim=d_model
+            )
+            self.decoder_pos = PositionalEncoding(d_model, dropout)
+
+            decoder_layer = TransformerDecoderLayer(
+                d_model=d_model,
+                nhead=nhead,
+                dim_feedforward=dim_feedforward,
+                dropout=dropout,
+                normalize_before=normalize_before,
+            )
+
+            if normalize_before:
+                decoder_norm = nn.LayerNorm(d_model)
+            else:
+                decoder_norm = None
+
+            self.decoder = nn.TransformerDecoder(
+                decoder_layer=decoder_layer,
+                num_layers=num_layers,
+                norm=decoder_norm,
+            )
+
+            self.decoder_output_layer = torch.nn.Linear(d_model, self.decoder_num_class)
+            self.decoder_criterion = LabelSmoothingLoss()
+        else:
+            self.decoder_criterion = None
+
+    @torch.jit.export
+    def forward(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[List[int]],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape (T, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs. Each sublist contains IDs for an utterance.
+            The IDs can be either phone IDs or word piece IDs.
+          sos_id:
+            sos token id
+          eos_id:
+            eos token id
+        Returns:
+            A scalar, the **sum** of label smoothing loss over utterances
+            in the batch without any normalization.
+        """
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id))
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1))
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device)
+        ys_out_pad = ys_out_pad.to(device)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (N, T) -> (N, T, C)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, N, C)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, N, C) -> (N, T, C)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (N, T, C)
+
+        decoder_loss = self.decoder_criterion(pred_pad, ys_out_pad)
+
+        return decoder_loss
+
+    @torch.jit.export
+    def decoder_nll(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[torch.Tensor],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape (T, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs (e.g., word piece IDs).
+            Each sublist represents an utterance.
+          sos_id:
+            The token ID for SOS.
+          eos_id:
+            The token ID for EOS.
+        Returns:
+            A 2-D tensor of shape (len(token_ids), max_token_length)
+            representing the cross entropy loss (i.e., negative log-likelihood).
+        """
+        # The common part between this function and decoder_forward could be
+        # extracted as a separate function.
+        if isinstance(token_ids[0], torch.Tensor):
+            # This branch is executed by torchscript in C++.
+            # See https://github.com/k2-fsa/k2/pull/870
+            # https://github.com/k2-fsa/k2/blob/3c1c18400060415b141ccea0115fd4bf0ad6234e/k2/torch/bin/attention_rescore.cu#L286
+            token_ids = [tolist(t) for t in token_ids]
+
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id))
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1))
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device, dtype=torch.int64)
+        ys_out_pad = ys_out_pad.to(device, dtype=torch.int64)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (B, T) -> (B, T, F)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, B, F)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, B, F) -> (B, T, F)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (B, T, F)
+        # nll: negative log-likelihood
+        nll = torch.nn.functional.cross_entropy(
+            pred_pad.view(-1, self.decoder_num_class),
+            ys_out_pad.view(-1),
+            ignore_index=-1,
+            reduction="none",
+        )
+
+        nll = nll.view(pred_pad.shape[0], -1)
+
+        return nll
+
+
+def add_sos(token_ids: List[List[int]], sos_id: int) -> List[List[int]]:
+    """Prepend sos_id to each utterance.
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      sos_id:
+        The ID of the SOS token.
+    Return:
+      Return a new list-of-list, where each sublist starts
+      with SOS ID.
+    """
+    return [[sos_id] + utt for utt in token_ids]
+
+
+def add_eos(token_ids: List[List[int]], eos_id: int) -> List[List[int]]:
+    """Append eos_id to each utterance.
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      eos_id:
+        The ID of the EOS token.
+    Return:
+      Return a new list-of-list, where each sublist ends
+      with EOS ID.
+    """
+    return [utt + [eos_id] for utt in token_ids]
+
+
+def decoder_padding_mask(ys_pad: torch.Tensor, ignore_id: int = -1) -> torch.Tensor:
+    """Generate a length mask for input.
+    The masked position are filled with True,
+    Unmasked positions are filled with False.
+    Args:
+      ys_pad:
+        padded tensor of dimension (batch_size, input_length).
+      ignore_id:
+        the ignored number (the padding number) in ys_pad
+    Returns:
+      Tensor:
+        a bool tensor of the same shape as the input tensor.
+    """
+    ys_mask = ys_pad == ignore_id
+    return ys_mask
+
+
+def generate_square_subsequent_mask(sz: int) -> torch.Tensor:
+    """Generate a square mask for the sequence. The masked positions are
+    filled with float('-inf'). Unmasked positions are filled with float(0.0).
+    The mask can be used for masked self-attention.
+    For instance, if sz is 3, it returns::
+        tensor([[0., -inf, -inf],
+                [0., 0., -inf],
+                [0., 0., 0]])
+    Args:
+      sz: mask size
+    Returns:
+      A square mask of dimension (sz, sz)
+    """
+    mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)
+    mask = (
+        mask.float()
+        .masked_fill(mask == 0, float("-inf"))
+        .masked_fill(mask == 1, float(0.0))
+    )
+    return mask
+
+
+def tolist(t: torch.Tensor) -> List[int]:
+    """Used by jit"""
+    return torch.jit.annotate(List[int], t.tolist())

egs/librispeech/ASR/zipformer_ctc_streaming/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/zipformer_ctc_streaming/export.py

@@ -0,0 +1,240 @@
+#!/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.
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_ctc_model, get_params
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import str2bool
+
+
+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 0.",
+    )
+
+    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=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    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="zipformer_ctc/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="""It contains language related input files such as "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=True,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+    params.vocab_size = num_classes
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = get_ctc_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:
+        logging.info("Using torch.jit.script")
+        # 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 = 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/zipformer_ctc_streaming/label_smoothing.py

@@ -0,0 +1,109 @@
+# 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
+
+
+class LabelSmoothingLoss(torch.nn.Module):
+    """
+    Implement the LabelSmoothingLoss proposed in the following paper
+    https://arxiv.org/pdf/1512.00567.pdf
+    (Rethinking the Inception Architecture for Computer Vision)
+
+    """
+
+    def __init__(
+        self,
+        ignore_index: int = -1,
+        label_smoothing: float = 0.1,
+        reduction: str = "sum",
+    ) -> None:
+        """
+        Args:
+          ignore_index:
+            ignored class id
+          label_smoothing:
+            smoothing rate (0.0 means the conventional cross entropy loss)
+          reduction:
+            It has the same meaning as the reduction in
+            `torch.nn.CrossEntropyLoss`. It can be one of the following three
+            values: (1) "none": No reduction will be applied. (2) "mean": the
+            mean of the output is taken. (3) "sum": the output will be summed.
+        """
+        super().__init__()
+        assert 0.0 <= label_smoothing < 1.0, f"{label_smoothing}"
+        assert reduction in ("none", "sum", "mean"), reduction
+        self.ignore_index = ignore_index
+        self.label_smoothing = label_smoothing
+        self.reduction = reduction
+
+    def forward(self, x: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
+        """
+        Compute loss between x and target.
+
+        Args:
+          x:
+            prediction of dimension
+            (batch_size, input_length, number_of_classes).
+          target:
+            target masked with self.ignore_index of
+            dimension (batch_size, input_length).
+
+        Returns:
+          A scalar tensor containing the loss without normalization.
+        """
+        assert x.ndim == 3
+        assert target.ndim == 2
+        assert x.shape[:2] == target.shape
+        num_classes = x.size(-1)
+        x = x.reshape(-1, num_classes)
+        # Now x is of shape (N*T, C)
+
+        # We don't want to change target in-place below,
+        # so we make a copy of it here
+        target = target.clone().reshape(-1)
+
+        ignored = target == self.ignore_index
+
+        # See https://github.com/k2-fsa/icefall/issues/240
+        # and https://github.com/k2-fsa/icefall/issues/297
+        # for why we don't use target[ignored] = 0 here
+        target = torch.where(ignored, torch.zeros_like(target), target)
+
+        true_dist = torch.nn.functional.one_hot(target, num_classes=num_classes).to(x)
+
+        true_dist = (
+            true_dist * (1 - self.label_smoothing) + self.label_smoothing / num_classes
+        )
+
+        # Set the value of ignored indexes to 0
+        #
+        # See https://github.com/k2-fsa/icefall/issues/240
+        # and https://github.com/k2-fsa/icefall/issues/297
+        # for why we don't use true_dist[ignored] = 0 here
+        true_dist = torch.where(
+            ignored.unsqueeze(1).repeat(1, true_dist.shape[1]),
+            torch.zeros_like(true_dist),
+            true_dist,
+        )
+
+        loss = -1 * (torch.log_softmax(x, dim=1) * true_dist)
+        if self.reduction == "sum":
+            return loss.sum()
+        elif self.reduction == "mean":
+            return loss.sum() / (~ignored).sum()
+        else:
+            return loss.sum(dim=-1)

egs/librispeech/ASR/zipformer_ctc_streaming/model.py

@@ -0,0 +1,158 @@
+# 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 List
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from transformer import encoder_padding_mask
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.utils import encode_supervisions
+
+
+class CTCModel(nn.Module):
+    """It implements a CTC model with an auxiliary attention head."""
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        encoder_dim: int,
+        vocab_size: int,
+    ):
+        """
+        Args:
+          encoder:
+            An instance of `EncoderInterface`. The shared encoder for the CTC and attention
+            branches
+          decoder:
+            An instance of `nn.Module`. This is the decoder for the attention branch.
+          encoder_dim:
+            Dimension of the encoder output.
+          decoder_dim:
+            Dimension of the decoder output.
+          vocab_size:
+            Number of tokens of the modeling unit including blank.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+
+        self.encoder = encoder
+        self.ctc_output = nn.Sequential(
+            nn.Dropout(p=0.1),
+            nn.Linear(encoder_dim, vocab_size),
+            nn.LogSoftmax(dim=-1),
+        )
+        self.decoder = decoder
+
+    @torch.jit.ignore
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        supervisions: torch.Tensor,
+        graph_compiler: BpeCtcTrainingGraphCompiler,
+        subsampling_factor: int = 1,
+        beam_size: int = 10,
+        reduction: str = "sum",
+        use_double_scores: bool = False,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            Tensor of dimension (N, T, C) where N is the batch size,
+            T is the number of frames, and C is the feature dimension.
+          x_lens:
+            Tensor of dimension (N,) where N is the batch size.
+          supervisions:
+            Supervisions are used in training.
+          graph_compiler:
+            It is used to compile a decoding graph from texts.
+          subsampling_factor:
+            It is used to compute the `supervisions` for the encoder.
+          beam_size:
+            Beam size used in `k2.ctc_loss`.
+          reduction:
+            Reduction method used in `k2.ctc_loss`.
+          use_double_scores:
+            If True, use double precision in `k2.ctc_loss`.
+        Returns:
+          Return the CTC loss, attention loss, and the total number of frames.
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+
+        nnet_output, x_lens = self.encoder(x, x_lens)
+        assert torch.all(x_lens > 0)
+        # compute ctc log-probs
+        ctc_output = self.ctc_output(nnet_output)
+
+        # NOTE: We need `encode_supervisions` to sort sequences with
+        # different duration in decreasing order, required by
+        # `k2.intersect_dense` called in `k2.ctc_loss`
+        supervision_segments, texts = encode_supervisions(
+            supervisions, subsampling_factor=subsampling_factor
+        )
+        num_frames = supervision_segments[:, 2].sum().item()
+
+        # Works with a BPE model
+        token_ids = graph_compiler.texts_to_ids(texts)
+        decoding_graph = graph_compiler.compile(token_ids)
+
+        dense_fsa_vec = k2.DenseFsaVec(
+            ctc_output,
+            supervision_segments.cpu(),
+            allow_truncate=subsampling_factor - 1,
+        )
+
+        ctc_loss = k2.ctc_loss(
+            decoding_graph=decoding_graph,
+            dense_fsa_vec=dense_fsa_vec,
+            output_beam=beam_size,
+            reduction=reduction,
+            use_double_scores=use_double_scores,
+        )
+
+        if self.decoder is not None:
+            nnet_output = nnet_output.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+            mmodel = (
+                self.decoder.module if hasattr(self.decoder, "module") else self.decoder
+            )
+            # Note: We need to generate an unsorted version of token_ids
+            # `encode_supervisions()` called above sorts text, but
+            # encoder_memory and memory_mask are not sorted, so we
+            # use an unsorted version `supervisions["text"]` to regenerate
+            # the token_ids
+            #
+            # See https://github.com/k2-fsa/icefall/issues/97
+            # for more details
+            unsorted_token_ids = graph_compiler.texts_to_ids(supervisions["text"])
+            mask = encoder_padding_mask(nnet_output.size(0), supervisions)
+            mask = mask.to(nnet_output.device) if mask is not None else None
+            att_loss = mmodel.forward(
+                nnet_output,
+                mask,
+                token_ids=unsorted_token_ids,
+                sos_id=graph_compiler.sos_id,
+                eos_id=graph_compiler.eos_id,
+            )
+        else:
+            att_loss = torch.tensor([0])
+
+        return ctc_loss, att_loss, num_frames

egs/librispeech/ASR/zipformer_ctc_streaming/scaling_converter.py

@@ -0,0 +1,114 @@
+# 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 replaces various modules in a model.
+Specifically, ActivationBalancer is replaced with an identity operator;
+Whiten is also replaced with an identity operator;
+BasicNorm is replaced by a module with `exp` removed.
+"""
+
+import copy
+from typing import List
+
+import torch
+import torch.nn as nn
+from scaling import ActivationBalancer, BasicNorm, Whiten
+
+
+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 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
+
+
+# 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,
+):
+    """
+    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.
+    Return:
+      Return a model without scaled layers.
+    """
+    if not inplace:
+        model = copy.deepcopy(model)
+
+    d = {}
+    for name, m in model.named_modules():
+        if isinstance(m, BasicNorm):
+            d[name] = convert_basic_norm(m)
+        elif isinstance(m, (ActivationBalancer, Whiten)):
+            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/zipformer_ctc_streaming/subsampling.py

@@ -0,0 +1,153 @@
+# 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
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' == T//4
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(self, idim: int, odim: int) -> None:
+        """
+        Args:
+          idim:
+            Input dim. The input shape is (N, T, idim).
+            Caution: It requires: T >=7, idim >=7
+          odim:
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, odim)
+        """
+        assert idim >= 7
+        super().__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(in_channels=1, out_channels=odim, kernel_size=3, stride=2),
+            nn.ReLU(),
+            nn.Conv2d(in_channels=odim, out_channels=odim, kernel_size=3, stride=2),
+            nn.ReLU(),
+        )
+        self.out = nn.Linear(odim * (((idim - 1) // 2 - 1) // 2), odim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        x = self.conv(x)
+        # Now x is of shape (N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        # Now x is of shape (N, ((T-1)//2 - 1))//2, odim)
+        return x
+
+
+class VggSubsampling(nn.Module):
+    """Trying to follow the setup described in the following paper:
+    https://arxiv.org/pdf/1910.09799.pdf
+
+    This paper is not 100% explicit so I am guessing to some extent,
+    and trying to compare with other VGG implementations.
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' = T//4
+    """
+
+    def __init__(self, idim: int, odim: int) -> None:
+        """Construct a VggSubsampling object.
+
+        This uses 2 VGG blocks with 2 Conv2d layers each,
+        subsampling its input by a factor of 4 in the time dimensions.
+
+        Args:
+          idim:
+            Input dim. The input shape is (N, T, idim).
+            Caution: It requires: T >=7, idim >=7
+          odim:
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, odim)
+        """
+        super().__init__()
+
+        cur_channels = 1
+        layers = []
+        block_dims = [32, 64]
+
+        # The decision to use padding=1 for the 1st convolution, then padding=0
+        # for the 2nd and for the max-pooling, and ceil_mode=True, was driven by
+        # a back-compatibility concern so that the number of frames at the
+        # output would be equal to:
+        #  (((T-1)//2)-1)//2.
+        # We can consider changing this by using padding=1 on the
+        # 2nd convolution, so the num-frames at the output would be T//4.
+        for block_dim in block_dims:
+            layers.append(
+                torch.nn.Conv2d(
+                    in_channels=cur_channels,
+                    out_channels=block_dim,
+                    kernel_size=3,
+                    padding=1,
+                    stride=1,
+                )
+            )
+            layers.append(torch.nn.ReLU())
+            layers.append(
+                torch.nn.Conv2d(
+                    in_channels=block_dim,
+                    out_channels=block_dim,
+                    kernel_size=3,
+                    padding=0,
+                    stride=1,
+                )
+            )
+            layers.append(
+                torch.nn.MaxPool2d(kernel_size=2, stride=2, padding=0, ceil_mode=True)
+            )
+            cur_channels = block_dim
+
+        self.layers = nn.Sequential(*layers)
+
+        self.out = nn.Linear(block_dims[-1] * (((idim - 1) // 2 - 1) // 2), odim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        x = x.unsqueeze(1)
+        x = self.layers(x)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        return x

egs/librispeech/ASR/zipformer_ctc_streaming/transformer.py

@@ -0,0 +1,928 @@
+# Copyright    2021 University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# 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 Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+from label_smoothing import LabelSmoothingLoss
+from subsampling import Conv2dSubsampling, VggSubsampling
+from torch.nn.utils.rnn import pad_sequence
+
+# Note: TorchScript requires Dict/List/etc. to be fully typed.
+Supervisions = Dict[str, torch.Tensor]
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        num_features: int,
+        num_classes: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        num_decoder_layers: int = 6,
+        dropout: float = 0.1,
+        normalize_before: bool = True,
+        vgg_frontend: bool = False,
+        use_feat_batchnorm: Union[float, bool] = 0.1,
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          num_classes:
+            The output dimension of the model.
+          subsampling_factor:
+            Number of output frames is num_in_frames // subsampling_factor.
+            Currently, subsampling_factor MUST be 4.
+          d_model:
+            Attention dimension.
+          nhead:
+            Number of heads in multi-head attention.
+            Must satisfy d_model // nhead == 0.
+          dim_feedforward:
+            The output dimension of the feedforward layers in encoder/decoder.
+          num_encoder_layers:
+            Number of encoder layers.
+          num_decoder_layers:
+            Number of decoder layers.
+          dropout:
+            Dropout in encoder/decoder.
+          normalize_before:
+            If True, use pre-layer norm; False to use post-layer norm.
+          vgg_frontend:
+            True to use vgg style frontend for subsampling.
+          use_feat_batchnorm:
+            True to use batchnorm for the input layer.
+            Float value to scale the input layer.
+            False to do nothing.
+        """
+        super().__init__()
+        self.use_feat_batchnorm = use_feat_batchnorm
+        assert isinstance(use_feat_batchnorm, (float, bool))
+        if isinstance(use_feat_batchnorm, bool) and use_feat_batchnorm:
+            self.feat_batchnorm = nn.BatchNorm1d(num_features)
+
+        self.num_features = num_features
+        self.num_classes = num_classes
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_classes)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_classes -> d_model
+        if vgg_frontend:
+            self.encoder_embed = VggSubsampling(num_features, d_model)
+        else:
+            self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder_pos = PositionalEncoding(d_model, dropout)
+
+        encoder_layer = TransformerEncoderLayer(
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            dropout=dropout,
+            normalize_before=normalize_before,
+        )
+
+        if normalize_before:
+            encoder_norm = nn.LayerNorm(d_model)
+        else:
+            encoder_norm = None
+
+        self.encoder = nn.TransformerEncoder(
+            encoder_layer=encoder_layer,
+            num_layers=num_encoder_layers,
+            norm=encoder_norm,
+        )
+
+        # TODO(fangjun): remove dropout
+        self.encoder_output_layer = nn.Sequential(
+            nn.Dropout(p=dropout), nn.Linear(d_model, num_classes)
+        )
+
+        if num_decoder_layers > 0:
+            self.decoder_num_class = (
+                self.num_classes
+            )  # bpe model already has sos/eos symbol
+
+            self.decoder_embed = nn.Embedding(
+                num_embeddings=self.decoder_num_class, embedding_dim=d_model
+            )
+            self.decoder_pos = PositionalEncoding(d_model, dropout)
+
+            decoder_layer = TransformerDecoderLayer(
+                d_model=d_model,
+                nhead=nhead,
+                dim_feedforward=dim_feedforward,
+                dropout=dropout,
+                normalize_before=normalize_before,
+            )
+
+            if normalize_before:
+                decoder_norm = nn.LayerNorm(d_model)
+            else:
+                decoder_norm = None
+
+            self.decoder = nn.TransformerDecoder(
+                decoder_layer=decoder_layer,
+                num_layers=num_decoder_layers,
+                norm=decoder_norm,
+            )
+
+            self.decoder_output_layer = torch.nn.Linear(d_model, self.decoder_num_class)
+
+            self.decoder_criterion = LabelSmoothingLoss()
+        else:
+            self.decoder_criterion = None
+
+    def forward(
+        self, x: torch.Tensor, supervision: Optional[Supervisions] = None
+    ) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (N, T, C).
+          supervision:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            (CAUTION: It contains length information, i.e., start and number of
+             frames, before subsampling)
+
+        Returns:
+          Return a tuple containing 3 tensors:
+            - CTC output for ctc decoding. Its shape is (N, T, C)
+            - Encoder output with shape (T, N, C). It can be used as key and
+              value for the decoder.
+            - Encoder output padding mask. It can be used as
+              memory_key_padding_mask for the decoder. Its shape is (N, T).
+              It is None if `supervision` is None.
+        """
+        if isinstance(self.use_feat_batchnorm, bool) and self.use_feat_batchnorm:
+            x = x.permute(0, 2, 1)  # (N, T, C) -> (N, C, T)
+            x = self.feat_batchnorm(x)
+            x = x.permute(0, 2, 1)  # (N, C, T) -> (N, T, C)
+        if isinstance(self.use_feat_batchnorm, float):
+            x *= self.use_feat_batchnorm
+        encoder_memory, memory_key_padding_mask = self.run_encoder(x, supervision)
+        x = self.ctc_output(encoder_memory)
+        return x, encoder_memory, memory_key_padding_mask
+
+    def run_encoder(
+        self, x: torch.Tensor, supervisions: Optional[Supervisions] = None
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """Run the transformer encoder.
+
+        Args:
+          x:
+            The model input. Its shape is (N, T, C).
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute the encoder padding mask, which is used as memory key
+            padding mask for the decoder.
+        Returns:
+          Return a tuple with two tensors:
+            - The encoder output, with shape (T, N, C)
+            - encoder padding mask, with shape (N, T).
+              The mask is None if `supervisions` is None.
+              It is used as memory key padding mask in the decoder.
+        """
+        x = self.encoder_embed(x)
+        x = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        mask = encoder_padding_mask(x.size(0), supervisions)
+        mask = mask.to(x.device) if mask is not None else None
+        x = self.encoder(x, src_key_padding_mask=mask)  # (T, N, C)
+
+        return x, mask
+
+    def ctc_output(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            The output tensor from the transformer encoder.
+            Its shape is (T, N, C)
+
+        Returns:
+          Return a tensor that can be used for CTC decoding.
+          Its shape is (N, T, C)
+        """
+        x = self.encoder_output_layer(x)
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+        x = nn.functional.log_softmax(x, dim=-1)  # (N, T, C)
+        return x
+
+    @torch.jit.export
+    def decoder_forward(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[List[int]],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape (T, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs. Each sublist contains IDs for an utterance.
+            The IDs can be either phone IDs or word piece IDs.
+          sos_id:
+            sos token id
+          eos_id:
+            eos token id
+
+        Returns:
+            A scalar, the **sum** of label smoothing loss over utterances
+            in the batch without any normalization.
+        """
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id))
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1))
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device)
+        ys_out_pad = ys_out_pad.to(device)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (N, T) -> (N, T, C)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, N, C)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, N, C) -> (N, T, C)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (N, T, C)
+
+        decoder_loss = self.decoder_criterion(pred_pad, ys_out_pad)
+
+        return decoder_loss
+
+    @torch.jit.export
+    def decoder_nll(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[torch.Tensor],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape (T, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs (e.g., word piece IDs).
+            Each sublist represents an utterance.
+          sos_id:
+            The token ID for SOS.
+          eos_id:
+            The token ID for EOS.
+        Returns:
+            A 2-D tensor of shape (len(token_ids), max_token_length)
+            representing the cross entropy loss (i.e., negative log-likelihood).
+        """
+        # The common part between this function and decoder_forward could be
+        # extracted as a separate function.
+        if isinstance(token_ids[0], torch.Tensor):
+            # This branch is executed by torchscript in C++.
+            # See https://github.com/k2-fsa/k2/pull/870
+            # https://github.com/k2-fsa/k2/blob/3c1c18400060415b141ccea0115fd4bf0ad6234e/k2/torch/bin/attention_rescore.cu#L286
+            token_ids = [tolist(t) for t in token_ids]
+
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id))
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1))
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device, dtype=torch.int64)
+        ys_out_pad = ys_out_pad.to(device, dtype=torch.int64)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (B, T) -> (B, T, F)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, B, F)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, B, F) -> (B, T, F)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (B, T, F)
+        # nll: negative log-likelihood
+        nll = torch.nn.functional.cross_entropy(
+            pred_pad.view(-1, self.decoder_num_class),
+            ys_out_pad.view(-1),
+            ignore_index=-1,
+            reduction="none",
+        )
+
+        nll = nll.view(pred_pad.shape[0], -1)
+
+        return nll
+
+
+class TransformerEncoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerEncoderLayer.
+    Add support of normalize_before,
+    i.e., use layer_norm before the first block.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+      normalize_before:
+        whether to use layer_norm before the first block.
+
+    Examples::
+        >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = encoder_layer(src)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: str = "relu",
+        normalize_before: bool = True,
+    ) -> None:
+        super(TransformerEncoderLayer, self).__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+
+        self.activation = _get_activation_fn(activation)
+
+        self.normalize_before = normalize_before
+
+    def __setstate__(self, state):
+        if "activation" not in state:
+            state["activation"] = nn.functional.relu
+        super(TransformerEncoderLayer, self).__setstate__(state)
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        src_mask: Optional[torch.Tensor] = None,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional)
+
+        Shape:
+            src: (S, N, E).
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        residual = src
+        if self.normalize_before:
+            src = self.norm1(src)
+        src2 = self.self_attn(
+            src,
+            src,
+            src,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = residual + self.dropout1(src2)
+        if not self.normalize_before:
+            src = self.norm1(src)
+
+        residual = src
+        if self.normalize_before:
+            src = self.norm2(src)
+        src2 = self.linear2(self.dropout(self.activation(self.linear1(src))))
+        src = residual + self.dropout2(src2)
+        if not self.normalize_before:
+            src = self.norm2(src)
+        return src
+
+
+class TransformerDecoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerDecoderLayer.
+    Add support of normalize_before,
+    i.e., use layer_norm before the first block.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+
+    Examples::
+        >>> decoder_layer = nn.TransformerDecoderLayer(d_model=512, nhead=8)
+        >>> memory = torch.rand(10, 32, 512)
+        >>> tgt = torch.rand(20, 32, 512)
+        >>> out = decoder_layer(tgt, memory)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: str = "relu",
+        normalize_before: bool = True,
+    ) -> None:
+        super(TransformerDecoderLayer, self).__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        self.src_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.norm3 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+        self.dropout3 = nn.Dropout(dropout)
+
+        self.activation = _get_activation_fn(activation)
+
+        self.normalize_before = normalize_before
+
+    def __setstate__(self, state):
+        if "activation" not in state:
+            state["activation"] = nn.functional.relu
+        super(TransformerDecoderLayer, self).__setstate__(state)
+
+    def forward(
+        self,
+        tgt: torch.Tensor,
+        memory: torch.Tensor,
+        tgt_mask: Optional[torch.Tensor] = None,
+        memory_mask: Optional[torch.Tensor] = None,
+        tgt_key_padding_mask: Optional[torch.Tensor] = None,
+        memory_key_padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Pass the inputs (and mask) through the decoder layer.
+
+        Args:
+          tgt:
+            the sequence to the decoder layer (required).
+          memory:
+            the sequence from the last layer of the encoder (required).
+          tgt_mask:
+            the mask for the tgt sequence (optional).
+          memory_mask:
+            the mask for the memory sequence (optional).
+          tgt_key_padding_mask:
+            the mask for the tgt keys per batch (optional).
+          memory_key_padding_mask:
+            the mask for the memory keys per batch (optional).
+
+        Shape:
+            tgt: (T, N, E).
+            memory: (S, N, E).
+            tgt_mask: (T, T).
+            memory_mask: (T, S).
+            tgt_key_padding_mask: (N, T).
+            memory_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm1(tgt)
+        tgt2 = self.self_attn(
+            tgt,
+            tgt,
+            tgt,
+            attn_mask=tgt_mask,
+            key_padding_mask=tgt_key_padding_mask,
+        )[0]
+        tgt = residual + self.dropout1(tgt2)
+        if not self.normalize_before:
+            tgt = self.norm1(tgt)
+
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm2(tgt)
+        tgt2 = self.src_attn(
+            tgt,
+            memory,
+            memory,
+            attn_mask=memory_mask,
+            key_padding_mask=memory_key_padding_mask,
+        )[0]
+        tgt = residual + self.dropout2(tgt2)
+        if not self.normalize_before:
+            tgt = self.norm2(tgt)
+
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm3(tgt)
+        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt))))
+        tgt = residual + self.dropout3(tgt2)
+        if not self.normalize_before:
+            tgt = self.norm3(tgt)
+        return tgt
+
+
+def _get_activation_fn(activation: str):
+    if activation == "relu":
+        return nn.functional.relu
+    elif activation == "gelu":
+        return nn.functional.gelu
+
+    raise RuntimeError("activation should be relu/gelu, not {}".format(activation))
+
+
+class PositionalEncoding(nn.Module):
+    """This class implements the positional encoding
+    proposed in the following paper:
+
+    - Attention Is All You Need: https://arxiv.org/pdf/1706.03762.pdf
+
+        PE(pos, 2i) = sin(pos / (10000^(2i/d_modle))
+        PE(pos, 2i+1) = cos(pos / (10000^(2i/d_modle))
+
+    Note::
+
+      1 / (10000^(2i/d_model)) = exp(-log(10000^(2i/d_model)))
+                               = exp(-1* 2i / d_model * log(100000))
+                               = exp(2i * -(log(10000) / d_model))
+    """
+
+    def __init__(self, d_model: int, dropout: float = 0.1) -> None:
+        """
+        Args:
+          d_model:
+            Embedding dimension.
+          dropout:
+            Dropout probability to be applied to the output of this module.
+        """
+        super().__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = nn.Dropout(p=dropout)
+        # not doing: self.pe = None because of errors thrown by torchscript
+        self.pe = torch.zeros(1, 0, self.d_model, dtype=torch.float32)
+
+    def extend_pe(self, x: torch.Tensor) -> None:
+        """Extend the time t in the positional encoding if required.
+
+        The shape of `self.pe` is (1, T1, d_model). The shape of the input x
+        is (N, T, d_model). If T > T1, then we change the shape of self.pe
+        to (N, T, d_model). Otherwise, nothing is done.
+
+        Args:
+          x:
+            It is a tensor of shape (N, T, C).
+        Returns:
+          Return None.
+        """
+        if self.pe is not None:
+            if self.pe.size(1) >= x.size(1):
+                self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        pe = torch.zeros(x.size(1), self.d_model, dtype=torch.float32)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe[:, 0::2] = torch.sin(position * div_term)
+        pe[:, 1::2] = torch.cos(position * div_term)
+        pe = pe.unsqueeze(0)
+        # Now pe is of shape (1, T, d_model), where T is x.size(1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Add positional encoding.
+
+        Args:
+          x:
+            Its shape is (N, T, C)
+
+        Returns:
+          Return a tensor of shape (N, T, C)
+        """
+        self.extend_pe(x)
+        x = x * self.xscale + self.pe[:, : x.size(1), :]
+        return self.dropout(x)
+
+
+class Noam(object):
+    """
+    Implements Noam optimizer.
+
+    Proposed in
+    "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf
+
+    Modified from
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/optimizer.py  # noqa
+
+    Args:
+      params:
+        iterable of parameters to optimize or dicts defining parameter groups
+      model_size:
+        attention dimension of the transformer model
+      factor:
+        learning rate factor
+      warm_step:
+        warmup steps
+    """
+
+    def __init__(
+        self,
+        params,
+        model_size: int = 256,
+        factor: float = 10.0,
+        warm_step: int = 25000,
+        weight_decay=0,
+    ) -> None:
+        """Construct an Noam object."""
+        self.optimizer = torch.optim.Adam(
+            params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=weight_decay
+        )
+        self._step = 0
+        self.warmup = warm_step
+        self.factor = factor
+        self.model_size = model_size
+        self._rate = 0
+
+    @property
+    def param_groups(self):
+        """Return param_groups."""
+        return self.optimizer.param_groups
+
+    def step(self):
+        """Update parameters and rate."""
+        self._step += 1
+        rate = self.rate()
+        for p in self.optimizer.param_groups:
+            p["lr"] = rate
+        self._rate = rate
+        self.optimizer.step()
+
+    def rate(self, step=None):
+        """Implement `lrate` above."""
+        if step is None:
+            step = self._step
+        return (
+            self.factor
+            * self.model_size ** (-0.5)
+            * min(step ** (-0.5), step * self.warmup ** (-1.5))
+        )
+
+    def zero_grad(self):
+        """Reset gradient."""
+        self.optimizer.zero_grad()
+
+    def state_dict(self):
+        """Return state_dict."""
+        return {
+            "_step": self._step,
+            "warmup": self.warmup,
+            "factor": self.factor,
+            "model_size": self.model_size,
+            "_rate": self._rate,
+            "optimizer": self.optimizer.state_dict(),
+        }
+
+    def load_state_dict(self, state_dict):
+        """Load state_dict."""
+        for key, value in state_dict.items():
+            if key == "optimizer":
+                self.optimizer.load_state_dict(state_dict["optimizer"])
+            else:
+                setattr(self, key, value)
+
+
+def encoder_padding_mask(
+    max_len: int, supervisions: Optional[Supervisions] = None
+) -> Optional[torch.Tensor]:
+    """Make mask tensor containing indexes of padded part.
+
+    TODO::
+      This function **assumes** that the model uses
+      a subsampling factor of 4. We should remove that
+      assumption later.
+
+    Args:
+      max_len:
+        Maximum length of input features.
+        CAUTION: It is the length after subsampling.
+      supervisions:
+        Supervision in lhotse format.
+        See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+        (CAUTION: It contains length information, i.e., start and number of
+         frames, before subsampling)
+
+    Returns:
+        Tensor: Mask tensor of dimension (batch_size, input_length),
+        True denote the masked indices.
+    """
+    if supervisions is None:
+        return None
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"],
+            supervisions["num_frames"],
+        ),
+        1,
+    ).to(torch.int32)
+
+    lengths = [0 for _ in range(int(supervision_segments[:, 0].max().item()) + 1)]
+    for idx in range(supervision_segments.size(0)):
+        # Note: TorchScript doesn't allow to unpack tensors as tuples
+        sequence_idx = supervision_segments[idx, 0].item()
+        start_frame = supervision_segments[idx, 1].item()
+        num_frames = supervision_segments[idx, 2].item()
+        lengths[sequence_idx] = start_frame + num_frames
+
+    lengths = [((i - 1) // 2 - 1) // 2 for i in lengths]
+    bs = int(len(lengths))
+    seq_range = torch.arange(0, max_len, dtype=torch.int64)
+    seq_range_expand = seq_range.unsqueeze(0).expand(bs, max_len)
+    # Note: TorchScript doesn't implement Tensor.new()
+    seq_length_expand = torch.tensor(
+        lengths, device=seq_range_expand.device, dtype=seq_range_expand.dtype
+    ).unsqueeze(-1)
+    mask = seq_range_expand >= seq_length_expand
+
+    return mask
+
+
+def decoder_padding_mask(ys_pad: torch.Tensor, ignore_id: int = -1) -> torch.Tensor:
+    """Generate a length mask for input.
+
+    The masked position are filled with True,
+    Unmasked positions are filled with False.
+
+    Args:
+      ys_pad:
+        padded tensor of dimension (batch_size, input_length).
+      ignore_id:
+        the ignored number (the padding number) in ys_pad
+
+    Returns:
+      Tensor:
+        a bool tensor of the same shape as the input tensor.
+    """
+    ys_mask = ys_pad == ignore_id
+    return ys_mask
+
+
+def generate_square_subsequent_mask(sz: int) -> torch.Tensor:
+    """Generate a square mask for the sequence. The masked positions are
+    filled with float('-inf'). Unmasked positions are filled with float(0.0).
+    The mask can be used for masked self-attention.
+
+    For instance, if sz is 3, it returns::
+
+        tensor([[0., -inf, -inf],
+                [0., 0., -inf],
+                [0., 0., 0]])
+
+    Args:
+      sz: mask size
+
+    Returns:
+      A square mask of dimension (sz, sz)
+    """
+    mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)
+    mask = (
+        mask.float()
+        .masked_fill(mask == 0, float("-inf"))
+        .masked_fill(mask == 1, float(0.0))
+    )
+    return mask
+
+
+def add_sos(token_ids: List[List[int]], sos_id: int) -> List[List[int]]:
+    """Prepend sos_id to each utterance.
+
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      sos_id:
+        The ID of the SOS token.
+
+    Return:
+      Return a new list-of-list, where each sublist starts
+      with SOS ID.
+    """
+    return [[sos_id] + utt for utt in token_ids]
+
+
+def add_eos(token_ids: List[List[int]], eos_id: int) -> List[List[int]]:
+    """Append eos_id to each utterance.
+
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      eos_id:
+        The ID of the EOS token.
+
+    Return:
+      Return a new list-of-list, where each sublist ends
+      with EOS ID.
+    """
+    return [utt + [eos_id] for utt in token_ids]
+
+
+def tolist(t: torch.Tensor) -> List[int]:
+    """Used by jit"""
+    return torch.jit.annotate(List[int], t.tolist())