From c689b018d7de3dabc4f0f25fdf916aa560e18013 Mon Sep 17 00:00:00 2001
From: marcoyang <marcoyang1998@gmail.com>
Date: Mon, 13 Feb 2023 12:45:22 +0800
Subject: [PATCH] add new files

---
 .../ASR/lstm_transducer_stateless3/decode.py  | 746 ++++++++++++++++++
 .../ASR/lstm_transducer_stateless3/lstmp.py   |   1 +
 .../ASR/lstm_transducer_stateless3/model.py   |   1 +
 3 files changed, 748 insertions(+)
 create mode 100755 egs/tal_csasr/ASR/lstm_transducer_stateless3/decode.py
 create mode 120000 egs/tal_csasr/ASR/lstm_transducer_stateless3/lstmp.py
 create mode 120000 egs/tal_csasr/ASR/lstm_transducer_stateless3/model.py

diff --git a/egs/tal_csasr/ASR/lstm_transducer_stateless3/decode.py b/egs/tal_csasr/ASR/lstm_transducer_stateless3/decode.py
new file mode 100755
index 000000000..bdbe69b06
--- /dev/null
+++ b/egs/tal_csasr/ASR/lstm_transducer_stateless3/decode.py
@@ -0,0 +1,746 @@
+#!/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.
+"""
+Usage:
+(1) greedy search
+./lstm_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./lstm_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./lstm_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./lstm_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./lstm_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./lstm_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search
+./lstm_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./lstm_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 4 \
+    --max-contexts 4 \
+    --max-states 8
+"""
+
+
+import argparse
+import logging
+import re
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import TAL_CSASRAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from lhotse.cut import Cut
+from local.text_normalize import text_normalize
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=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=False,
+        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_stateless3/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    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=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+    sp: spm.SentencePieceProcessor = 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 HLG, Used
+        only when --decoding_method is fast_beam_search.
+    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)
+
+    encoder_out, encoder_out_lens, _ = model.encoder(x=feature, x_lens=feature_lens)
+    hyps = []
+    zh_hyps = []
+    en_hyps = []
+    pattern = re.compile(r"([\u4e00-\u9fff])")
+    en_letter = "[\u0041-\u005a|\u0061-\u007a]+"  # English letters
+    zh_char = "[\u4e00-\u9fa5]+"  # Chinese chars
+    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)):
+            hyp = sp.decode([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+            chars = pattern.split(hyp.upper())
+            chars_new = []
+            zh_text = []
+            en_text = []
+            for char in chars:
+                if char != "":
+                    tokens = char.strip().split(" ")
+                    chars_new.extend(tokens)
+                    for token in tokens:
+                        zh_text.extend(re.findall(zh_char, token))
+                        en_text.extend(re.findall(en_letter, token))
+            hyps.append(chars_new)
+            zh_hyps.append(zh_text)
+            en_hyps.append(en_text)
+    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)):
+            hyp = sp.decode([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+            chars = pattern.split(hyp.upper())
+            chars_new = []
+            zh_text = []
+            en_text = []
+            for char in chars:
+                if char != "":
+                    tokens = char.strip().split(" ")
+                    chars_new.extend(tokens)
+                    for token in tokens:
+                        zh_text.extend(re.findall(zh_char, token))
+                        en_text.extend(re.findall(en_letter, token))
+            hyps.append(chars_new)
+            zh_hyps.append(zh_text)
+            en_hyps.append(en_text)
+    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)):
+            hyp = sp.decode([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+            chars = pattern.split(hyp.upper())
+            chars_new = []
+            zh_text = []
+            en_text = []
+            for char in chars:
+                if char != "":
+                    tokens = char.strip().split(" ")
+                    chars_new.extend(tokens)
+                    for token in tokens:
+                        zh_text.extend(re.findall(zh_char, token))
+                        en_text.extend(re.findall(en_letter, token))
+            hyps.append(chars_new)
+            zh_hyps.append(zh_text)
+            en_hyps.append(en_text)
+    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}"
+                )
+            for i in range(encoder_out.size(0)):
+                hyp = sp.decode([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+                chars = pattern.split(hyp.upper())
+                chars_new = []
+                zh_text = []
+                en_text = []
+                for char in chars:
+                    if char != "":
+                        tokens = char.strip().split(" ")
+                        chars_new.extend(tokens)
+                        for token in tokens:
+                            zh_text.extend(re.findall(zh_char, token))
+                            en_text.extend(re.findall(en_letter, token))
+                hyps.append(chars_new)
+                zh_hyps.append(zh_text)
+                en_hyps.append(en_text)
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": (hyps, zh_hyps, en_hyps)}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): (hyps, zh_hyps, en_hyps)
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": (hyps, zh_hyps, en_hyps)}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    decoding_graph: Optional[k2.Fsa] = None,
+    sp: spm.SentencePieceProcessor = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    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)
+    zh_results = defaultdict(list)
+    en_results = defaultdict(list)
+    pattern = re.compile(r"([\u4e00-\u9fff])")
+    en_letter = "[\u0041-\u005a|\u0061-\u007a]+"  # English letters
+    zh_char = "[\u4e00-\u9fa5]+"  # Chinese chars
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+        zh_texts = []
+        en_texts = []
+        for i in range(len(texts)):
+            text = texts[i]
+            chars = pattern.split(text.upper())
+            chars_new = []
+            zh_text = []
+            en_text = []
+            for char in chars:
+                if char != "":
+                    tokens = char.strip().split(" ")
+                    chars_new.extend(tokens)
+                    for token in tokens:
+                        zh_text.extend(re.findall(zh_char, token))
+                        en_text.extend(re.findall(en_letter, token))
+            zh_texts.append(zh_text)
+            en_texts.append(en_text)
+            texts[i] = chars_new
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            decoding_graph=decoding_graph,
+            batch=batch,
+            sp=sp,
+        )
+
+        for name, hyps_texts in hyps_dict.items():
+            this_batch = []
+            this_batch_zh = []
+            this_batch_en = []
+            # print(hyps_texts)
+            hyps, zh_hyps, en_hyps = hyps_texts
+            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))
+
+            for cut_id, hyp_words, ref_text in zip(cut_ids, zh_hyps, zh_texts):
+                this_batch_zh.append((cut_id, ref_text, hyp_words))
+
+            for cut_id, hyp_words, ref_text in zip(cut_ids, en_hyps, en_texts):
+                this_batch_en.append((cut_id, ref_text, hyp_words))
+
+            results[name].extend(this_batch)
+            zh_results[name + "_zh"].extend(this_batch_zh)
+            en_results[name + "_en"].extend(this_batch_en)
+
+        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, zh_results, en_results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    TAL_CSASRAsrDataModule.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",
+        "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}"
+    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}")
+
+    bpe_model = params.lang_dir + "/bpe.model"
+    sp = spm.SentencePieceProcessor()
+    sp.load(bpe_model)
+
+    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 params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    def text_normalize_for_cut(c: Cut):
+        # Text normalize for each sample
+        text = c.supervisions[0].text
+        text = text.strip("\n").strip("\t")
+        c.supervisions[0].text = text_normalize(text)
+        return c
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    tal_csasr = TAL_CSASRAsrDataModule(args)
+
+    dev_cuts = tal_csasr.valid_cuts()
+    dev_cuts = dev_cuts.map(text_normalize_for_cut)
+    dev_dl = tal_csasr.valid_dataloaders(dev_cuts)
+
+    test_cuts = tal_csasr.test_cuts()
+    test_cuts = test_cuts.map(text_normalize_for_cut)
+    test_dl = tal_csasr.test_dataloaders(test_cuts)
+
+    test_sets = ["dev", "test"]
+    test_dl = [dev_dl, test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict, zh_results_dict, en_results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            decoding_graph=decoding_graph,
+            sp=sp,
+        )
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=zh_results_dict,
+        )
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=en_results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/tal_csasr/ASR/lstm_transducer_stateless3/lstmp.py b/egs/tal_csasr/ASR/lstm_transducer_stateless3/lstmp.py
new file mode 120000
index 000000000..b82e115fc
--- /dev/null
+++ b/egs/tal_csasr/ASR/lstm_transducer_stateless3/lstmp.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/lstm_transducer_stateless2/lstmp.py
\ No newline at end of file
diff --git a/egs/tal_csasr/ASR/lstm_transducer_stateless3/model.py b/egs/tal_csasr/ASR/lstm_transducer_stateless3/model.py
new file mode 120000
index 000000000..ce6f089fd
--- /dev/null
+++ b/egs/tal_csasr/ASR/lstm_transducer_stateless3/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/lstm_transducer_stateless3/model.py
\ No newline at end of file