Add wenetspeech kws recipe

egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py

@@ -15,6 +15,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import logging
+import math
 import warnings
 from dataclasses import dataclass, field
 from typing import Dict, List, Optional, Tuple, Union
@@ -31,6 +33,7 @@ from icefall.rnn_lm.model import RnnLmModel
 from icefall.transformer_lm.model import TransformerLM
 from icefall.utils import (
     DecodingResults,
+    KeywordResult,
     add_eos,
     add_sos,
     get_texts,
@@ -789,6 +792,8 @@ class Hypothesis:
     # It contains only one entry.
     log_prob: torch.Tensor
 
+    ac_probs: Optional[List[float]] = None
+
     # timestamp[i] is the frame index after subsampling
     # on which ys[i] is decoded
     timestamp: List[int] = field(default_factory=list)
@@ -805,6 +810,8 @@ class Hypothesis:
     # Context graph state
     context_state: Optional[ContextState] = None
 
+    num_tailing_blanks: int = 0
+
     @property
     def key(self) -> str:
         """Return a string representation of self.ys"""
@@ -953,6 +960,232 @@ def get_hyps_shape(hyps: List[HypothesisList]) -> k2.RaggedShape:
     return ans
 
 
+def keywords_search(
+    model: nn.Module,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    context_graph: ContextGraph,
+    beam: int = 4,
+    ac_threshold: float = 0.15,
+    num_tailing_blanks: int = 8,
+    blank_penalty: float = 0,
+) -> List[List[KeywordResult]]:
+    """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded.
+
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C).
+      encoder_out_lens:
+        A 1-D tensor of shape (N,), containing number of valid frames in
+        encoder_out before padding.
+      beam:
+        Number of active paths during the beam search.
+    Returns:
+      Return a list of list of KeywordResult.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+    assert context_graph is not None
+
+    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
+        input=encoder_out,
+        lengths=encoder_out_lens.cpu(),
+        batch_first=True,
+        enforce_sorted=False,
+    )
+
+    blank_id = model.decoder.blank_id
+    unk_id = getattr(model, "unk_id", blank_id)
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+
+    batch_size_list = packed_encoder_out.batch_sizes.tolist()
+    N = encoder_out.size(0)
+    assert torch.all(encoder_out_lens > 0), encoder_out_lens
+    assert N == batch_size_list[0], (N, batch_size_list)
+
+    B = [HypothesisList() for _ in range(N)]
+    for i in range(N):
+        B[i].add(
+            Hypothesis(
+                ys=[-1] * (context_size - 1) + [blank_id],
+                log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+                context_state=context_graph.root,
+                timestamp=[],
+                ac_probs=[],
+            )
+        )
+
+    encoder_out = model.joiner.encoder_proj(packed_encoder_out.data)
+
+    offset = 0
+    finalized_B = []
+    sorted_ans = [[] for _ in range(N)]
+    for t, batch_size in enumerate(batch_size_list):
+        start = offset
+        end = offset + batch_size
+        current_encoder_out = encoder_out.data[start:end]
+        current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1)
+        # current_encoder_out's shape is (batch_size, 1, 1, encoder_out_dim)
+        offset = end
+
+        finalized_B = B[batch_size:] + finalized_B
+        B = B[:batch_size]
+
+        hyps_shape = get_hyps_shape(B).to(device)
+
+        A = [list(b) for b in B]
+
+        B = [HypothesisList() for _ in range(batch_size)]
+
+        ys_log_probs = torch.cat(
+            [hyp.log_prob.reshape(1, 1) for hyps in A for hyp in hyps]
+        )  # (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyps in A for hyp in hyps],
+            device=device,
+            dtype=torch.int64,
+        )  # (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # decoder_out is of shape (num_hyps, 1, 1, joiner_dim)
+
+        # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor
+        # as index, so we use `to(torch.int64)` below.
+        current_encoder_out = torch.index_select(
+            current_encoder_out,
+            dim=0,
+            index=hyps_shape.row_ids(1).to(torch.int64),
+        )  # (num_hyps, 1, 1, encoder_out_dim)
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+            project_input=False,
+        )  # (num_hyps, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)  # (num_hyps, vocab_size)
+
+        if blank_penalty != 0:
+            logits[:, 0] -= blank_penalty
+
+        probs = logits.softmax(dim=-1)  # (num_hyps, vocab_size)
+
+        log_probs = probs.log()
+
+        log_probs.add_(ys_log_probs)
+
+        vocab_size = log_probs.size(-1)
+
+        log_probs = log_probs.reshape(-1)
+
+        row_splits = hyps_shape.row_splits(1) * vocab_size
+        log_probs_shape = k2.ragged.create_ragged_shape2(
+            row_splits=row_splits, cached_tot_size=log_probs.numel()
+        )
+        ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs)
+        ragged_probs = k2.RaggedTensor(shape=log_probs_shape, value=probs)
+
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam)
+            hyp_probs = ragged_probs[i].tolist()
+
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+                topk_token_indexes = (topk_indexes % vocab_size).tolist()
+
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+                new_ys = hyp.ys[:]
+                new_token = topk_token_indexes[k]
+                new_timestamp = hyp.timestamp[:]
+                new_ac_probs = hyp.ac_probs[:]
+                context_score = 0
+                new_context_state = hyp.context_state
+                new_num_tailing_blanks = hyp.num_tailing_blanks + 1
+                if new_token not in (blank_id, unk_id):
+                    new_ys.append(new_token)
+                    new_timestamp.append(t)
+                    new_ac_probs.append(math.exp(hyp_probs[topk_indexes[k]]))
+                    (
+                        context_score,
+                        new_context_state,
+                        _,
+                    ) = context_graph.forward_one_step(hyp.context_state, new_token)
+                    new_num_tailing_blanks = 0
+                    if new_context_state.token == -1:  # root
+                        new_ys[-context_size:] = [-1] * (context_size - 1) + [blank_id]
+
+                new_log_prob = topk_log_probs[k] + context_score
+
+                new_hyp = Hypothesis(
+                    ys=new_ys,
+                    log_prob=new_log_prob,
+                    timestamp=new_timestamp,
+                    ac_probs=new_ac_probs,
+                    context_state=new_context_state,
+                    num_tailing_blanks=new_num_tailing_blanks,
+                )
+                B[i].add(new_hyp)
+
+            top_hyp = B[i].get_most_probable(length_norm=True)
+            matched, matched_state = context_graph.is_matched(top_hyp.context_state)
+            if matched:
+                ac_prob = (
+                    sum(top_hyp.ac_probs[-matched_state.level :]) / matched_state.level
+                )
+            if (
+                matched
+                and top_hyp.num_tailing_blanks > num_tailing_blanks
+                and ac_prob >= ac_threshold
+            ):
+                keyword = KeywordResult(
+                    hyps=top_hyp.ys[-matched_state.level :],
+                    timestamps=top_hyp.timestamp[-matched_state.level :],
+                    phrase=matched_state.phrase,
+                )
+                sorted_ans[i].append(keyword)
+                B[i] = HypothesisList()
+                B[i].add(
+                    Hypothesis(
+                        ys=[-1] * (context_size - 1) + [blank_id],
+                        log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+                        context_state=context_graph.root,
+                        timestamp=[],
+                        ac_probs=[],
+                    )
+                )
+
+    B = B + finalized_B
+
+    for i, hyps in enumerate(B):
+        top_hyp = hyps.get_most_probable(length_norm=True)
+        matched, matched_state = context_graph.is_matched(top_hyp.context_state)
+        if matched:
+            ac_prob = (
+                sum(top_hyp.ac_probs[-matched_state.level :]) / matched_state.level
+            )
+        if matched and ac_prob >= ac_threshold:
+            keyword = KeywordResult(
+                hyps=top_hyp.ys[-matched_state.level :],
+                timestamps=top_hyp.timestamp[-matched_state.level :],
+                phrase=matched_state.phrase,
+            )
+            sorted_ans[i].append(keyword)
+
+    ans = []
+    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
+    for i in range(N):
+        ans.append(sorted_ans[unsorted_indices[i]])
+    return ans
+
+
 def modified_beam_search(
     model: nn.Module,
     encoder_out: torch.Tensor,

egs/wenetspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py

@@ -407,3 +407,10 @@ class WenetSpeechAsrDataModule:
     def test_meeting_cuts(self) -> List[CutSet]:
         logging.info("About to get TEST_MEETING cuts")
         return load_manifest_lazy(self.args.manifest_dir / "cuts_TEST_MEETING.jsonl.gz")
+
+    @lru_cache()
+    def test_open_commands_cuts(self) -> CutSet:
+        logging.info("About to get open commands cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "open-commands-cn_cuts_test.jsonl.gz"
+        )

egs/wenetspeech/KWS/zipformer/asr_datamodule.py

@@ -0,0 +1,416 @@
+# Copyright      2021  Piotr Żelasko
+#
+# 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, List, Optional
+
+import torch
+from lhotse import (
+    CutSet,
+    Fbank,
+    FbankConfig,
+    load_manifest,
+    load_manifest_lazy,
+    set_caching_enabled,
+)
+from lhotse.dataset import (
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import 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 WenetSpeechAsrDataModule:
+    """
+    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(
+            "--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(
+            "--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(
+            "--training-subset",
+            type=str,
+            default="L",
+            help="The training subset for using",
+        )
+
+    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.
+        """
+        logging.info("About to get Musan cuts")
+        cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=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(
+            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,
+                buffer_size=300000,
+                drop_last=True,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        # '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,
+        )
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_dl.sampler.load_state_dict(sampler_state_dict)
+
+        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,
+            batch_size=None,
+            sampler=valid_sampler,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.info("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get train cuts")
+        cuts_train = load_manifest_lazy(
+            self.args.manifest_dir / f"cuts_{self.args.training_subset}.jsonl.gz"
+        )
+        return cuts_train
+
+    @lru_cache()
+    def valid_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "cuts_DEV.jsonl.gz")
+
+    @lru_cache()
+    def test_net_cuts(self) -> List[CutSet]:
+        logging.info("About to get TEST_NET cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "cuts_TEST_NET.jsonl.gz")
+
+    @lru_cache()
+    def test_meeting_cuts(self) -> List[CutSet]:
+        logging.info("About to get TEST_MEETING cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "cuts_TEST_MEETING.jsonl.gz")
+
+    @lru_cache()
+    def test_open_commands_cuts(self) -> CutSet:
+        logging.info("About to get open commands cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "open-commands-cn_cuts_test.jsonl.gz"
+        )

egs/wenetspeech/KWS/zipformer/beam_search.py

@@ -0,0 +1 @@
+../../ASR/pruned_transducer_stateless2/beam_search.py

egs/wenetspeech/KWS/zipformer/decode_pinyin.py

@@ -0,0 +1,637 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao
+#                                                 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.
+"""
+Usage:
+(2) modified beam search
+./zipformer/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Dict, List, Optional, Set, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import WenetSpeechAsrDataModule
+from beam_search import (
+    keywords_search,
+)
+from lhotse.cut import Cut
+from train import add_model_arguments, get_model, get_params
+
+from icefall import ContextGraph
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    make_pad_mask,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    text_to_pinyin,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+@dataclass
+class KwMetric:
+    TP: int = 0  # True positive
+    FN: int = 0  # False negative
+    FP: int = 0  # False positive
+    TN: int = 0  # True negative
+    FN_list: List[str] = field(default_factory=list)
+    FP_list: List[str] = field(default_factory=list)
+    TP_list: List[str] = field(default_factory=list)
+
+    def __str__(self) -> str:
+        return f"(TP:{self.TP}, FN:{self.FN}, FP:{self.FP}, TN:{self.TN})"
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_char",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--blank-penalty",
+        type=float,
+        default=0.0,
+        help="""
+        The penalty applied on blank symbol during decoding.
+        Note: It is a positive value that would be applied to logits like
+        this `logits[:, 0] -= blank_penalty` (suppose logits.shape is
+        [batch_size, vocab] and blank id is 0).
+        """,
+    )
+
+    parser.add_argument(
+        "--pinyin-type",
+        type=str,
+        help="The type of pinyin used as the modeling units.",
+    )
+
+    parser.add_argument(
+        "--keyword-file",
+        type=str,
+        help="File contains keywords.",
+    )
+
+    parser.add_argument(
+        "--keyword-score",
+        type=float,
+        default=0.75,
+        help="The threshold (probability) to boost the keyword.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    batch: dict,
+    kws_graph: ContextGraph,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.causal:
+        # this seems to cause insertions at the end of the utterance if used with zipformer.
+        pad_len = 30
+        feature_lens += pad_len
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, pad_len),
+            value=LOG_EPS,
+        )
+
+    x, x_lens = model.encoder_embed(feature, feature_lens)
+
+    src_key_padding_mask = make_pad_mask(x_lens)
+    x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+    encoder_out, encoder_out_lens = model.encoder(x, x_lens, src_key_padding_mask)
+    encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+    ans_dict = keywords_search(
+        model=model,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+        context_graph=kws_graph,
+        beam=params.beam_size,
+        num_tailing_blanks=8,
+    )
+
+    hyps = []
+    for ans in ans_dict:
+        hyp = []
+        for hit in ans:
+            hyp.append(
+                (
+                    hit.phrase,
+                    (hit.timestamps[0], hit.timestamps[-1]),
+                )
+            )
+        hyps.append(hyp)
+
+    return hyps
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    kws_graph: ContextGraph,
+    keywords: Set[str],
+) -> Dict[str, List[Tuple[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.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or LG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    log_interval = 20
+
+    results = []
+    metric = {"all": KwMetric()}
+    for k in keywords:
+        metric[k] = KwMetric()
+
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps = decode_one_batch(
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            kws_graph=kws_graph,
+            batch=batch,
+        )
+
+        this_batch = []
+        assert len(hyps) == len(texts)
+        for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+            ref_words = list(ref_text)
+            hyp_words = [x[0] for x in hyp_words]
+            this_batch.append((cut_id, ref_words, list("".join(hyp_words))))
+            hyp_set = set(hyp_words)
+            hyp_str = " | ".join(hyp_words)
+            for x in hyp_set:
+                assert x in keywords, x
+                if x in ref_text and x in keywords:
+                    metric["all"].TP += 1
+                    metric[x].TP += 1
+                    metric[x].TP_list.append(f"({ref_text} -> {x})")
+                if x not in ref_text and x in keywords:
+                    metric["all"].FP += 1
+                    metric[x].FP += 1
+                    metric[x].FP_list.append(f"({ref_text} -> {x}/{cut_id})")
+            for x in keywords:
+                if x not in ref_text and x not in hyp_set:
+                    metric["all"].TN += 1
+                    metric[x].TN += 1
+
+                if x in ref_text:
+                    fn = True
+                    for y in hyp_set:
+                        if y in ref_text:
+                            fn = False
+                            break
+                    if fn and ref_text.endswith(x):
+                        metric["all"].FN += 1
+                        metric[x].FN += 1
+                        metric[x].FN_list.append(f"({ref_text} -> {hyp_str}/{cut_id})")
+
+        results.extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results, metric
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results: List[Tuple[str, List[str], List[str]]],
+    metric: KwMetric,
+):
+    recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+    results = sorted(results)
+    store_transcripts(filename=recog_path, texts=results)
+    logging.info(f"The transcripts are stored in {recog_path}")
+
+    # The following prints out WERs, per-word error statistics and aligned
+    # ref/hyp pairs.
+    errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+    with open(errs_filename, "w") as f:
+        wer = write_error_stats(f, f"{test_set_name}", results, enable_log=True)
+    logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    metric_filename = params.res_dir / f"metric-{test_set_name}-{params.suffix}.txt"
+
+    print_s = ""
+    with open(metric_filename, "w") as of:
+        width = 10
+        for key, item in sorted(
+            metric.items(), key=lambda x: (x[1].FP, x[1].FN), reverse=True
+        ):
+            acc = (item.TP + item.TN) / (item.TP + item.TN + item.FP + item.FN)
+            precision = (item.TP + 1) / (item.TP + item.FP + 1)
+            recall = (item.TP + 1) / (item.TP + item.FN + 1)
+            fpr = (item.FP + 1) / (item.FP + item.TN + 1)
+            s = f"{key}:\n"
+            s += f"\t{'TP':{width}}{'FP':{width}}{'FN':{width}}{'TN':{width}}\n"
+            s += f"\t{str(item.TP):{width}}{str(item.FP):{width}}{str(item.FN):{width}}{str(item.TN):{width}}\n"
+            s += f"\tAccuracy: {acc:.3f}\n"
+            s += f"\tPrecision: {precision:.3f}\n"
+            s += f"\tRecall(PPR): {recall:.3f}\n"
+            s += f"\tFPR: {fpr:.3f}\n"
+            s += f"\tF1: {2 * precision * recall / (precision + recall):.3f}\n"
+            s += f"\tTP list: {' # '.join(item.TP_list)}\n"
+            s += f"\tFP list: {' # '.join(item.FP_list)}\n"
+            s += f"\tFN list: {' # '.join(item.FN_list)}\n"
+            of.write(s + "\n")
+            if key == "all":
+                logging.info(s)
+
+    logging.info("Wrote metric stats to {}".format(metric_filename))
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    WenetSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    params.res_dir = params.exp_dir / "kws"
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.causal:
+        assert (
+            "," not in params.chunk_size
+        ), "chunk_size should be one value in decoding."
+        assert (
+            "," not in params.left_context_frames
+        ), "left_context_frames should be one value in decoding."
+        params.suffix += f"-chunk-{params.chunk_size}"
+        params.suffix += f"-left-context-{params.left_context_frames}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+    params.suffix += f"-keyword-score-{params.keyword_score}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    keywords = []
+    keywords_id = []
+    with open(params.keyword_file, "r") as f:
+        for line in f.readlines():
+            score = 0
+            kws = line.strip().upper().split()
+            if kws[-1][0] == ":":
+                score = float(kws[-1][1:])
+                kws = kws[0:-1]
+            tmp_ids = []
+            kws = "".join(kws)
+            kws_py = text_to_pinyin(kws, mode=params.pinyin_type)
+            for k in kws_py:
+                if k in lexicon.token_table:
+                    tmp_ids.append(lexicon.token_table[k])
+                else:
+                    logging.warning(f"Containing OOV tokens, skipping line : {line}")
+                    tmp_ids = []
+                    break
+            if tmp_ids:
+                logging.info(f"Adding keyword : {kws}")
+                keywords.append(kws)
+                keywords_id.append((tmp_ids, score, kws))
+    kws_graph = ContextGraph(context_score=params.keyword_score)
+    kws_graph.build(keywords_id)
+
+    logging.info("About to create model")
+    model = get_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}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    wenetspeech = WenetSpeechAsrDataModule(args)
+
+    def remove_short_utt(c: Cut):
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        if T <= 0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from decoding, num_frames : {c.num_frames}."
+            )
+        return T > 0
+
+    def select_keywords(c: Cut):
+        text = c.supervisions[0].text.strip()
+        return text in keywords
+
+    commands_cuts = wenetspeech.test_open_commands_cuts()
+    commands_cuts = commands_cuts.filter(select_keywords)
+    commands_cuts = commands_cuts.filter(remove_short_utt)
+    commands_dl = wenetspeech.test_dataloaders(commands_cuts)
+
+    test_net_cuts = wenetspeech.test_net_cuts()
+    test_net_cuts = test_net_cuts.filter(remove_short_utt)
+    test_net_dl = wenetspeech.test_dataloaders(test_net_cuts)
+
+    test_sets = ["COMMANDS"]  # , "TEST_NET"]
+    test_dls = [commands_dl]  # , test_net_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results, metric = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            kws_graph=kws_graph,
+            keywords=set(keywords),
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results=results,
+            metric=metric,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()

egs/wenetspeech/KWS/zipformer/decoder.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/decoder.py

egs/wenetspeech/KWS/zipformer/encoder_interface.py

@@ -0,0 +1 @@
+../../ASR/pruned_transducer_stateless2/encoder_interface.py

egs/wenetspeech/KWS/zipformer/joiner.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/joiner.py

egs/wenetspeech/KWS/zipformer/model.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/model.py

egs/wenetspeech/KWS/zipformer/optim.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/optim.py

egs/wenetspeech/KWS/zipformer/scaling.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/scaling.py

egs/wenetspeech/KWS/zipformer/subsampling.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/subsampling.py

egs/wenetspeech/KWS/zipformer/train_pinyin.py

@@ -0,0 +1,704 @@
+#!/usr/bin/env python3
+# Copyright    2021-2023  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,
+#                                                       Zengwei Yao,
+#                                                       Daniel Povey)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+# For non-streaming model training:
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp \
+  --max-duration 1000
+
+# For streaming model training:
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp \
+  --causal 1 \
+  --max-duration 1000
+
+It supports training with:
+  - transducer loss (default), with `--use-transducer True --use-ctc False`
+  - ctc loss (not recommended), with `--use-transducer False --use-ctc True`
+  - transducer loss & ctc loss, with `--use-transducer True --use-ctc True`
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import WenetSpeechAsrDataModule
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import AsrModel
+from optim import Eden, ScaledAdam
+from scaling import ScheduledFloat
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    get_parameter_groups_with_lrs,
+    setup_logger,
+    str2bool,
+    text_to_pinyin,
+)
+
+from train import (
+    add_model_arguments,
+    add_training_arguments,
+    compute_validation_loss,
+    display_and_save_batch,
+    get_adjusted_batch_count,
+    get_model,
+    get_params,
+    load_checkpoint_if_available,
+    save_checkpoint,
+    scan_pessimistic_batches_for_oom,
+    set_batch_count,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_partial_tone",
+        help="Path to the pinyin lang directory",
+    )
+
+    parser.add_argument(
+        "--pinyin-type",
+        type=str,
+        default="partial_with_tone",
+        help="""
+            The style of the output pinyin, should be:
+              full_with_tone : zhong1 guo2
+              full_no_tone : zhong guo
+              partial_with_tone : zh ong1 g uo2
+              partial_no_tone : zh ong g uo
+        """,
+    )
+
+    parser.add_argument(
+        "--pinyin-errors",
+        default="split",
+        type=str,
+        help="""How to handle characters that has no pinyin,
+        see `text_to_pinyin` in icefall/utils.py for details
+        """,
+    )
+
+    add_training_arguments(parser)
+    add_model_arguments(parser)
+
+    return parser
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = graph_compiler.texts_to_ids(texts, sep="/")
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss, _ = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    cur_batch_idx = params.get("cur_batch_idx", 0)
+
+    saved_bad_model = False
+
+    def save_bad_model(suffix: str = ""):
+        save_checkpoint_impl(
+            filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt",
+            model=model,
+            model_avg=model_avg,
+            params=params,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=0,
+        )
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch_idx % 10 == 0:
+            set_batch_count(model, get_adjusted_batch_count(params))
+        if batch_idx < cur_batch_idx:
+            continue
+        cur_batch_idx = batch_idx
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            save_bad_model()
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            params.cur_batch_idx = batch_idx
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            del params.cur_batch_idx
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+
+            if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                if not saved_bad_model:
+                    save_bad_model(suffix="-first-warning")
+                    saved_bad_model = True
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                save_bad_model()
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = max(scheduler.get_last_lr())
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale", cur_grad_scale, params.batch_idx_train
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    if not params.use_transducer:
+        params.ctc_loss_scale = 1.0
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    optimizer = ScaledAdam(
+        get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True),
+        lr=params.base_lr,  # should have no effect
+        clipping_scale=2.0,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            2**22
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    wenetspeech = WenetSpeechAsrDataModule(args)
+
+    train_cuts = wenetspeech.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 15.0:
+            # logging.warning(
+            #     f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            # )
+            return False
+
+        return True
+
+    def encode_text(c: Cut):
+        # Text normalize for each sample
+        text = c.supervisions[0].text
+        text = "/".join(
+            text_to_pinyin(text, mode=params.pinyin_type, errors=params.pinyin_errors)
+        )
+        c.supervisions[0].text = text
+        return c
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+    train_cuts = train_cuts.map(encode_text)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = wenetspeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = wenetspeech.valid_cuts()
+    valid_cuts = valid_cuts.map(encode_text)
+    valid_dl = wenetspeech.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        # scan_pessimistic_batches_for_oom(
+        #    model=model,
+        #    train_dl=train_dl,
+        #    optimizer=optimizer,
+        #    graph_compiler=graph_compiler,
+        #    params=params,
+        # )
+        pass
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def main():
+    parser = get_parser()
+    WenetSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.lang_dir = Path(args.lang_dir)
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+if __name__ == "__main__":
+    main()

egs/wenetspeech/KWS/zipformer/zipformer.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/zipformer.py

icefall/context_graph.py

@@ -322,6 +322,24 @@ class ContextGraph:
         )
         return (score + node.output_score, node, matched_node)
 
+    def is_matched(self, state: ContextState) -> Tuple[bool, ContextState]:
+        """Whether current state matches any phrase (i.e. current state is the
+        end state or the output of current state is not None.
+
+        Args:
+          state:
+            The given state(trie node).
+
+        Returns:
+          Return a tuple of status and matched state.
+        """
+        if state.is_end:
+            return True, state
+        else:
+            if state.output is not None:
+                return True, state.output
+            return False, None
+
     def finalize(self, state: ContextState) -> Tuple[float, ContextState]:
         """When reaching the end of the decoded sequence, we need to finalize
         the matching, the purpose is to subtract the added bonus score for the