Merge d55cec6b8741a6007eea53150f2ef4cdf82e19cb into dca21c2a17b6e62f687e49398517cb57f62203b0

egs/wenetspeech/ASR/lstm_transducer_stateless/asr_datamodule.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless2/asr_datamodule.py

egs/wenetspeech/ASR/lstm_transducer_stateless/beam_search.py

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

egs/wenetspeech/ASR/lstm_transducer_stateless/decode.py

@@ -0,0 +1,880 @@
+#!/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:
+(1) greedy search
+./lstm_transducer_stateless/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./lstm_transducer_stateless/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./lstm_transducer_stateless/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./lstm_transducer_stateless/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./lstm_transducer_stateless/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./lstm_transducer_stateless/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./lstm_transducer_stateless/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./lstm_transducer_stateless/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./lstm_transducer_stateless/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./lstm_transducer_stateless/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./lstm_transducer_stateless/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./lstm_transducer_stateless/decode.py \
+    --epoch 35 \
+    --avg 15 \
+    --exp-dir ./lstm_transducer_stateless/exp \
+    --lang-dir data/lang_char \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import WenetSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+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,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="lstm_transducer_stateless/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(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    #  graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    decoding_graph: 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 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)
+
+    # tail padding here to alleviate the tail deletion problem
+    num_tail_padded_frames = 35
+    feature = torch.nn.functional.pad(
+        feature,
+        (0, 0, 0, num_tail_padded_frames),
+        mode="constant",
+        value=LOG_EPS,
+    )
+    feature_lens += num_tail_padded_frames
+
+    encoder_out, encoder_out_lens, _ = model.encoder(
+        x=feature, x_lens=feature_lens
+    )
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            sentence = "".join([lexicon.word_table[i] for i in hyp])
+            hyps.append(list(sentence))
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=graph_compiler.texts_to_ids(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif (
+        params.decoding_method == "greedy_search"
+        and params.max_sym_per_frame == 1
+    ):
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i + 1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append([lexicon.token_table[idx] for idx in hyp])
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> 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 = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        texts = [list(str(text)) for text in texts]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                this_batch.append((cut_id, ref_text, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(
+                f"batch {batch_str}, cuts processed until now is {num_cuts}"
+            )
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[int], List[int]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir
+        / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    WenetSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += (
+            f"-{params.decoding_method}-beam-size-{params.beam_size}"
+        )
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(
+                params.exp_dir, iteration=-params.iter
+            )[: params.avg]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(
+                params.exp_dir, iteration=-params.iter
+            )[: params.avg + 1]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            #  word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            #  word_table = None
+            decoding_graph = k2.trivial_graph(
+                params.vocab_size - 1, device=device
+            )
+    else:
+        decoding_graph = None
+        #  word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # Note: Please use "pip install webdataset==0.1.103"
+    # for installing the webdataset.
+    import glob
+    import os
+
+    from lhotse import CutSet
+    from lhotse.dataset.webdataset import export_to_webdataset
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    wenetspeech = WenetSpeechAsrDataModule(args)
+
+    dev = "dev"
+    test_net = "test_net"
+    test_meeting = "test_meeting"
+
+    if not os.path.exists(f"{dev}/shared-0.tar"):
+        os.makedirs(dev, exist_ok=True)
+        dev_cuts = wenetspeech.valid_cuts()
+        export_to_webdataset(
+            dev_cuts,
+            output_path=f"{dev}/shared-%d.tar",
+            shard_size=300,
+        )
+
+    if not os.path.exists(f"{test_net}/shared-0.tar"):
+        os.makedirs(test_net, exist_ok=True)
+        test_net_cuts = wenetspeech.test_net_cuts()
+        export_to_webdataset(
+            test_net_cuts,
+            output_path=f"{test_net}/shared-%d.tar",
+            shard_size=300,
+        )
+
+    if not os.path.exists(f"{test_meeting}/shared-0.tar"):
+        os.makedirs(test_meeting, exist_ok=True)
+        test_meeting_cuts = wenetspeech.test_meeting_cuts()
+        export_to_webdataset(
+            test_meeting_cuts,
+            output_path=f"{test_meeting}/shared-%d.tar",
+            shard_size=300,
+        )
+
+    print("done")
+
+    dev_shards = [
+        str(path)
+        for path in sorted(glob.glob(os.path.join(dev, "shared-*.tar")))
+    ]
+    cuts_dev_webdataset = CutSet.from_webdataset(
+        dev_shards,
+        split_by_worker=True,
+        split_by_node=True,
+        shuffle_shards=True,
+    )
+
+    test_net_shards = [
+        str(path)
+        for path in sorted(glob.glob(os.path.join(test_net, "shared-*.tar")))
+    ]
+    cuts_test_net_webdataset = CutSet.from_webdataset(
+        test_net_shards,
+        split_by_worker=True,
+        split_by_node=True,
+        shuffle_shards=True,
+    )
+
+    test_meeting_shards = [
+        str(path)
+        for path in sorted(
+            glob.glob(os.path.join(test_meeting, "shared-*.tar"))
+        )
+    ]
+    cuts_test_meeting_webdataset = CutSet.from_webdataset(
+        test_meeting_shards,
+        split_by_worker=True,
+        split_by_node=True,
+        shuffle_shards=True,
+    )
+
+    dev_dl = wenetspeech.valid_dataloaders(cuts_dev_webdataset)
+    test_net_dl = wenetspeech.test_dataloaders(cuts_test_net_webdataset)
+    test_meeting_dl = wenetspeech.test_dataloaders(cuts_test_meeting_webdataset)
+
+    test_sets = ["DEV", "TEST_NET", "TEST_MEETING"]
+    test_dl = [dev_dl, test_net_dl, test_meeting_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            #  word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()

egs/wenetspeech/ASR/lstm_transducer_stateless/decoder.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/decoder.py

egs/wenetspeech/ASR/lstm_transducer_stateless/encoder_interface.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/encoder_interface.py

egs/wenetspeech/ASR/lstm_transducer_stateless/export.py

@@ -0,0 +1,403 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+(1) Export to torchscript model using torch.jit.trace()
+
+./lstm_transducer_stateless/export.py \
+  --exp-dir ./lstm_transducer_stateless/exp \
+  --lang-dir data/lang_char \
+  --epoch 35 \
+  --avg 10 \
+  --jit-trace 1
+
+It will generate 3 files: `encoder_jit_trace.pt`,
+`decoder_jit_trace.pt`, and `joiner_jit_trace.pt`.
+
+(2) Export `model.state_dict()`
+
+./lstm_transducer_stateless/export.py \
+  --exp-dir ./lstm_transducer_stateless/exp \
+  --lang-dir data/lang_char \
+  --epoch 35 \
+  --avg 10
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+To use the generated file with `lstm_transducer_stateless/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./lstm_transducer_stateless/decode.py \
+        --exp-dir ./lstm_transducer_stateless/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --lang-dir data/lang_char \
+
+Check ./pretrained.py for its usage.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+import torch.nn as nn
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+from icefall.lexicon import Lexicon
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless3/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_char",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--jit-trace",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.trace.
+        It will generate 3 files:
+         - encoder_jit_trace.pt
+         - decoder_jit_trace.pt
+         - joiner_jit_trace.pt
+
+        Check ./jit_pretrained.py for how to use them.
+        """,
+    )
+
+    parser.add_argument(
+        "--pnnx",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.trace for later
+        converting to PNNX. It will generate 3 files:
+         - encoder_jit_trace-pnnx.pt
+         - decoder_jit_trace-pnnx.pt
+         - joiner_jit_trace-pnnx.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def export_encoder_model_jit_trace(
+    encoder_model: nn.Module,
+    encoder_filename: str,
+) -> None:
+    """Export the given encoder model with torch.jit.trace()
+
+    Note: The warmup argument is fixed to 1.
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported model.
+    """
+    x = torch.zeros(1, 100, 80, dtype=torch.float32)
+    x_lens = torch.tensor([100], dtype=torch.int64)
+    states = encoder_model.get_init_states()
+
+    traced_model = torch.jit.trace(encoder_model, (x, x_lens, states))
+    traced_model.save(encoder_filename)
+    logging.info(f"Saved to {encoder_filename}")
+
+
+def export_decoder_model_jit_trace(
+    decoder_model: nn.Module,
+    decoder_filename: str,
+) -> None:
+    """Export the given decoder model with torch.jit.trace()
+
+    Note: The argument need_pad is fixed to False.
+
+    Args:
+      decoder_model:
+        The input decoder model
+      decoder_filename:
+        The filename to save the exported model.
+    """
+    y = torch.zeros(10, decoder_model.context_size, dtype=torch.int64)
+    need_pad = torch.tensor([False])
+
+    traced_model = torch.jit.trace(decoder_model, (y, need_pad))
+    traced_model.save(decoder_filename)
+    logging.info(f"Saved to {decoder_filename}")
+
+
+def export_joiner_model_jit_trace(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+) -> None:
+    """Export the given joiner model with torch.jit.trace()
+
+    Note: The argument project_input is fixed to True. A user should not
+    project the encoder_out/decoder_out by himself/herself. The exported joiner
+    will do that for the user.
+
+    Args:
+      joiner_model:
+        The input joiner model
+      joiner_filename:
+        The filename to save the exported model.
+
+    """
+    encoder_out_dim = joiner_model.encoder_proj.weight.shape[1]
+    decoder_out_dim = joiner_model.decoder_proj.weight.shape[1]
+    encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32)
+    decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32)
+
+    traced_model = torch.jit.trace(joiner_model, (encoder_out, decoder_out))
+    traced_model.save(joiner_filename)
+    logging.info(f"Saved to {joiner_filename}")
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+    lexicon = Lexicon(params.lang_dir)
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    if params.pnnx:
+        params.is_pnnx = params.pnnx
+        logging.info("For PNNX")
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    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.pnnx:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        logging.info("Using torch.jit.trace()")
+        encoder_filename = params.exp_dir / "encoder_jit_trace-pnnx.pt"
+        export_encoder_model_jit_trace(model.encoder, encoder_filename)
+
+        decoder_filename = params.exp_dir / "decoder_jit_trace-pnnx.pt"
+        export_decoder_model_jit_trace(model.decoder, decoder_filename)
+
+        joiner_filename = params.exp_dir / "joiner_jit_trace-pnnx.pt"
+        export_joiner_model_jit_trace(model.joiner, joiner_filename)
+    elif params.jit_trace is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        logging.info("Using torch.jit.trace()")
+        encoder_filename = params.exp_dir / "encoder_jit_trace.pt"
+        export_encoder_model_jit_trace(model.encoder, encoder_filename)
+
+        decoder_filename = params.exp_dir / "decoder_jit_trace.pt"
+        export_decoder_model_jit_trace(model.decoder, decoder_filename)
+
+        joiner_filename = params.exp_dir / "joiner_jit_trace.pt"
+        export_joiner_model_jit_trace(model.joiner, joiner_filename)
+    else:
+        logging.info("Not using torchscript")
+        # 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/wenetspeech/ASR/lstm_transducer_stateless/joiner.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/joiner.py

egs/wenetspeech/ASR/lstm_transducer_stateless/lstm.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/lstm_transducer_stateless/lstm.py

egs/wenetspeech/ASR/lstm_transducer_stateless/model.py

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

egs/wenetspeech/ASR/lstm_transducer_stateless/ncnn-decode.py

@@ -0,0 +1,296 @@
+#!/usr/bin/env python3
+# flake8: noqa
+#
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+  ./lstm_transducer_stateless/ncnn-decode.py \
+   --token-filename ./data/lang_char/tokens.txt \
+   --encoder-param-filename ./lstm_transducer_stateless/exp-new/encoder_jit_trace-epoch-11-avg-2-pnnx.ncnn.param \
+   --encoder-bin-filename ./lstm_transducer_stateless/exp-new/encoder_jit_trace-epoch-11-avg-2-pnnx.ncnn.bin \
+   --decoder-param-filename ./lstm_transducer_stateless/exp-new/decoder_jit_trace-epoch-11-avg-2-pnnx.ncnn.param \
+   --decoder-bin-filename ./lstm_transducer_stateless/exp-new/decoder_jit_trace-epoch-11-avg-2-pnnx.ncnn.bin \
+   --joiner-param-filename ./lstm_transducer_stateless/exp-new/joiner_jit_trace-epoch-11-avg-2-pnnx.ncnn.param \
+   --joiner-bin-filename ./lstm_transducer_stateless/exp-new/joiner_jit_trace-epoch-11-avg-2-pnnx.ncnn.bin \
+   ./test_wavs/DEV_T0000000001.wav
+"""
+
+import argparse
+import logging
+from typing import List
+
+import k2
+import kaldifeat
+import ncnn
+import torch
+import torchaudio
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--token-filename",
+        type=str,
+        help="Path to tokens.txt",
+    )
+
+    parser.add_argument(
+        "--encoder-param-filename",
+        type=str,
+        help="Path to encoder.ncnn.param",
+    )
+
+    parser.add_argument(
+        "--encoder-bin-filename",
+        type=str,
+        help="Path to encoder.ncnn.bin",
+    )
+
+    parser.add_argument(
+        "--decoder-param-filename",
+        type=str,
+        help="Path to decoder.ncnn.param",
+    )
+
+    parser.add_argument(
+        "--decoder-bin-filename",
+        type=str,
+        help="Path to decoder.ncnn.bin",
+    )
+
+    parser.add_argument(
+        "--joiner-param-filename",
+        type=str,
+        help="Path to joiner.ncnn.param",
+    )
+
+    parser.add_argument(
+        "--joiner-bin-filename",
+        type=str,
+        help="Path to joiner.ncnn.bin",
+    )
+
+    parser.add_argument(
+        "sound_filename",
+        type=str,
+        help="Path to foo.wav",
+    )
+
+    return parser.parse_args()
+
+
+class Model:
+    def __init__(self, args):
+        self.init_encoder(args)
+        self.init_decoder(args)
+        self.init_joiner(args)
+
+    def init_encoder(self, args):
+        encoder_net = ncnn.Net()
+        encoder_net.opt.use_packing_layout = False
+        encoder_net.opt.use_fp16_storage = False
+        encoder_param = args.encoder_param_filename
+        encoder_model = args.encoder_bin_filename
+
+        encoder_net.load_param(encoder_param)
+        encoder_net.load_model(encoder_model)
+
+        self.encoder_net = encoder_net
+
+    def init_decoder(self, args):
+        decoder_param = args.decoder_param_filename
+        decoder_model = args.decoder_bin_filename
+
+        decoder_net = ncnn.Net()
+        decoder_net.opt.use_packing_layout = False
+
+        decoder_net.load_param(decoder_param)
+        decoder_net.load_model(decoder_model)
+
+        self.decoder_net = decoder_net
+
+    def init_joiner(self, args):
+        joiner_param = args.joiner_param_filename
+        joiner_model = args.joiner_bin_filename
+        joiner_net = ncnn.Net()
+        joiner_net.opt.use_packing_layout = False
+        joiner_net.load_param(joiner_param)
+        joiner_net.load_model(joiner_model)
+
+        self.joiner_net = joiner_net
+
+    def run_encoder(self, x, states):
+        with self.encoder_net.create_extractor() as ex:
+            ex.set_num_threads(10)
+            ex.input("in0", ncnn.Mat(x.numpy()).clone())
+            x_lens = torch.tensor([x.size(0)], dtype=torch.float32)
+            ex.input("in1", ncnn.Mat(x_lens.numpy()).clone())
+            ex.input("in2", ncnn.Mat(states[0].numpy()).clone())
+            ex.input("in3", ncnn.Mat(states[1].numpy()).clone())
+
+            ret, ncnn_out0 = ex.extract("out0")
+            assert ret == 0, ret
+
+            ret, ncnn_out1 = ex.extract("out1")
+            assert ret == 0, ret
+
+            ret, ncnn_out2 = ex.extract("out2")
+            assert ret == 0, ret
+
+            ret, ncnn_out3 = ex.extract("out3")
+            assert ret == 0, ret
+
+            encoder_out = torch.from_numpy(ncnn_out0.numpy()).clone()
+            encoder_out_lens = torch.from_numpy(ncnn_out1.numpy()).to(
+                torch.int32
+            )
+            hx = torch.from_numpy(ncnn_out2.numpy()).clone()
+            cx = torch.from_numpy(ncnn_out3.numpy()).clone()
+            return encoder_out, encoder_out_lens, hx, cx
+
+    def run_decoder(self, decoder_input):
+        assert decoder_input.dtype == torch.int32
+
+        with self.decoder_net.create_extractor() as ex:
+            ex.set_num_threads(10)
+            ex.input("in0", ncnn.Mat(decoder_input.numpy()).clone())
+            ret, ncnn_out0 = ex.extract("out0")
+            assert ret == 0, ret
+            decoder_out = torch.from_numpy(ncnn_out0.numpy()).clone()
+            return decoder_out
+
+    def run_joiner(self, encoder_out, decoder_out):
+        with self.joiner_net.create_extractor() as ex:
+            ex.set_num_threads(10)
+            ex.input("in0", ncnn.Mat(encoder_out.numpy()).clone())
+            ex.input("in1", ncnn.Mat(decoder_out.numpy()).clone())
+            ret, ncnn_out0 = ex.extract("out0")
+            assert ret == 0, ret
+            joiner_out = torch.from_numpy(ncnn_out0.numpy()).clone()
+            return joiner_out
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert sample_rate == expected_sample_rate, (
+            f"expected sample rate: {expected_sample_rate}. "
+            f"Given: {sample_rate}"
+        )
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def greedy_search(model: Model, encoder_out: torch.Tensor):
+    assert encoder_out.ndim == 2
+    T = encoder_out.size(0)
+
+    context_size = 2
+    blank_id = 0  # hard-code to 0
+    hyp = [blank_id] * context_size
+
+    decoder_input = torch.tensor(hyp, dtype=torch.int32)  # (1, context_size)
+
+    decoder_out = model.run_decoder(decoder_input).squeeze(0)
+    #  print(decoder_out.shape)  # (512,)
+
+    for t in range(T):
+        encoder_out_t = encoder_out[t]
+        joiner_out = model.run_joiner(encoder_out_t, decoder_out)
+        #  print(joiner_out.shape) # [500]
+        y = joiner_out.argmax(dim=0).tolist()
+        if y != blank_id:
+            hyp.append(y)
+            decoder_input = hyp[-context_size:]
+            decoder_input = torch.tensor(decoder_input, dtype=torch.int32)
+            decoder_out = model.run_decoder(decoder_input).squeeze(0)
+    return hyp[context_size:]
+
+
+def main():
+    args = get_args()
+    logging.info(vars(args))
+
+    model = Model(args)
+
+    sound_file = args.sound_filename
+
+    sample_rate = 16000
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = sample_rate
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {sound_file}")
+    wave_samples = read_sound_files(
+        filenames=[sound_file],
+        expected_sample_rate=sample_rate,
+    )[0]
+
+    logging.info("Decoding started")
+    features = fbank(wave_samples)
+
+    num_encoder_layers = 12
+    d_model = 512
+    rnn_hidden_size = 1024
+
+    states = (
+        torch.zeros(num_encoder_layers, d_model),
+        torch.zeros(
+            num_encoder_layers,
+            rnn_hidden_size,
+        ),
+    )
+
+    encoder_out, encoder_out_lens, hx, cx = model.run_encoder(features, states)
+    hyp = greedy_search(model, encoder_out)
+
+    logging.info(sound_file)
+
+    token_table = k2.SymbolTable.from_file(args.token_filename)
+    words = [token_table[i] for i in hyp]
+    logging.info("".join(words))
+
+
+if __name__ == "__main__":
+    formatter = (
+        "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    )
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()

egs/wenetspeech/ASR/lstm_transducer_stateless/optim.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/optim.py

egs/wenetspeech/ASR/lstm_transducer_stateless/scaling.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/scaling.py

egs/wenetspeech/ASR/lstm_transducer_stateless/scaling_converter.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless3/scaling_converter.py

egs/wenetspeech/ASR/lstm_transducer_stateless/test_model.py

@@ -0,0 +1,95 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./lstm_transducer_stateless/test_model.py
+"""
+
+import os
+from pathlib import Path
+
+import torch
+
+#  from export import (
+#      export_decoder_model_jit_trace,
+#      export_encoder_model_jit_trace,
+#      export_joiner_model_jit_trace,
+#  )
+from lstm import stack_states, unstack_states
+from scaling_converter import convert_scaled_to_non_scaled
+from train import get_params, get_transducer_model
+
+
+def test_model():
+    params = get_params()
+    params.vocab_size = 5536
+    params.blank_id = 0
+    params.context_size = 2
+    params.unk_id = 2
+    params.encoder_dim = 256
+    params.rnn_hidden_size = 384
+    params.dim_feedforward = 512
+    params.num_encoder_layers = 6
+    params.aux_layer_period = 0
+    params.exp_dir = Path("exp_test_model")
+
+    model = get_transducer_model(params)
+    model.eval()
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+    return
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+
+    if not os.path.exists(params.exp_dir):
+        os.path.mkdir(params.exp_dir)
+
+    encoder_filename = params.exp_dir / "encoder_jit_trace.pt"
+    export_encoder_model_jit_trace(model.encoder, encoder_filename)
+
+    decoder_filename = params.exp_dir / "decoder_jit_trace.pt"
+    export_decoder_model_jit_trace(model.decoder, decoder_filename)
+
+    joiner_filename = params.exp_dir / "joiner_jit_trace.pt"
+    export_joiner_model_jit_trace(model.joiner, joiner_filename)
+
+    print("The model has been successfully exported using jit.trace.")
+
+
+def test_states_stack_and_unstack():
+    layer, batch, hidden, cell = 12, 100, 512, 1024
+    states = (
+        torch.randn(layer, batch, hidden),
+        torch.randn(layer, batch, cell),
+    )
+    states2 = stack_states(unstack_states(states))
+    assert torch.allclose(states[0], states2[0])
+    assert torch.allclose(states[1], states2[1])
+
+
+def main():
+    test_model()
+    #  test_states_stack_and_unstack()
+
+
+if __name__ == "__main__":
+    main()

egs/wenetspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py

@@ -292,7 +292,7 @@ class WenetSpeechAsrDataModule:
                 max_duration=self.args.max_duration,
                 shuffle=self.args.shuffle,
                 num_buckets=self.args.num_buckets,
-                buffer_size=30000,
+                buffer_size=300000,
                 drop_last=True,
             )
         else:

egs/wenetspeech/ASR/pruned_transducer_stateless2/decode.py

@@ -203,7 +203,7 @@ def get_parser():
         "--beam-size",
         type=int,
         default=4,
-        help="""An interger indicating how many candidates we will keep for each
+        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.""",
     )

icefall/utils.py

@@ -1395,7 +1395,7 @@ def tokenize_by_CJK_char(line: str) -> str:
 def display_and_save_batch(
     batch: dict,
     params: AttributeDict,
-    sp: spm.SentencePieceProcessor,
+    sp: Optional[spm.SentencePieceProcessor] = None,
 ) -> None:
     """Display the batch statistics and save the batch into disk.
 
@@ -1406,7 +1406,7 @@ def display_and_save_batch(
       params:
         Parameters for training. See :func:`get_params`.
       sp:
-        The BPE model.
+        Optional. The BPE model.
     """
     from lhotse.utils import uuid4
 
@@ -1418,9 +1418,14 @@ def display_and_save_batch(
     features = batch["inputs"]
 
     logging.info(f"features shape: {features.shape}")
+    text = supervisions["text"]
+
+    if sp is not None:
+        y = sp.encode(text, out_type=int)
+        num_tokens = sum(len(i) for i in y)
+    else:
+        num_tokens = sum(len(i) for i in text)
 
-    y = sp.encode(supervisions["text"], out_type=int)
-    num_tokens = sum(len(i) for i in y)
     logging.info(f"num tokens: {num_tokens}")