add char-based pruned-rnnt2 for wenetspeech

egs/wenetspeech/ASR/README.md

@@ -0,0 +1,19 @@
+
+# Introduction
+
+This recipe includes some different ASR models trained with WenetSpeech.
+
+[./RESULTS.md](./RESULTS.md) contains the latest results.
+
+# Transducers
+
+There are various folders containing the name `transducer` in this folder.
+The following table lists the differences among them.
+
+|                                       | Encoder             | Decoder            | Comment                     |
+|---------------------------------------|---------------------|--------------------|-----------------------------|
+| `pruned_transducer_stateless2`        | Conformer(modified) | Embedding + Conv1d | Using k2 pruned RNN-T loss  |                      |
+
+The decoder in `transducer_stateless` is modified from the paper
+[Rnn-Transducer with Stateless Prediction Network](https://ieeexplore.ieee.org/document/9054419/).
+We place an additional Conv1d layer right after the input embedding layer.

egs/wenetspeech/ASR/RESULTS.md

@@ -0,0 +1,93 @@
+## Results
+
+### WenetSpeech char-based training results (Pruned Transducer 2)
+
+#### 2022-05-06
+
+Using the codes from this PR https://github.com/k2-fsa/icefall/pull/349 and the Lhotse v1.1.
+
+When training with the L subset, the WERs are
+
+|                                    |  dev  | test-net | test-meeting | comment                                 |
+|------------------------------------|-------|----------|--------------|-----------------------------------------|
+|          greedy search             | 8.06  | 9.16     | 14.07        | --epoch 6, --avg 3, --max-duration 100  |
+| modified beam search (beam size 4) | 7.97  | 9.18     | 13.91        | --epoch 6, --avg 3, --max-duration 100  |
+| fast beam search (set as default)  | 8.13  | 9.12     | 14.33        | --epoch 6, --avg 3, --max-duration 1500 |
+
+The training command for reproducing is given below:
+
+```
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./pruned_transducer_stateless2/train.py \
+  --lang-dir data/lang_char \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --world-size 8 \
+  --num-epochs 10 \
+  --start-epoch 0 \
+  --max-duration 180 \
+  --valid-interval 3000 \
+  --model-warm-step 3000 \
+  --save-every-n 8000 \
+  --training-subset L
+```
+
+The tensorboard training log can be found at
+https://tensorboard.dev/experiment/VpA8b7SZQ7CEjZs9WZ5HNA/#scalars
+
+The decoding command is:
+```
+epoch=6
+avg=3
+
+## greedy search
+./pruned_transducer_stateless2/decode.py \
+        --epoch 6 \
+        --avg 3 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir data/lang_char \
+        --max-duration 100 \
+        --decoding-method greedy_search
+
+## modified beam search
+./pruned_transducer_stateless2/decode.py \
+        --epoch 6 \
+        --avg 3 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir data/lang_char \
+        --max-duration 100 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+## fast beam search
+./pruned_transducer_stateless2/decode.py \
+        --epoch 6 \
+        --avg 3 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir data/lang_char \
+        --max-duration 1500 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+```
+
+When training with the M subset, the WERs are
+
+|                                    |   dev  | test-net  | test-meeting  | comment                                   |
+|------------------------------------|--------|-----------|---------------|-------------------------------------------|
+|          greedy search             | 10.40  | 11.31     | 19.64         | --epoch 29, --avg 11, --max-duration 100  |
+| modified beam search (beam size 4) |  9.85  | 11.04     | 18.20         | --epoch 29, --avg 11, --max-duration 100  |
+| fast beam search (set as default)  | 10.18  | 11.10     | 19.32         | --epoch 29, --avg 11, --max-duration 1500 |
+
+
+When training with the S subset, the WERs are
+
+|                                    |  dev   | test-net  | test-meeting  | comment                                   |
+|------------------------------------|--------|-----------|---------------|-------------------------------------------|
+|          greedy search             | 19.92  | 25.20     | 35.35         | --epoch 29, --avg 24, --max-duration 100  |
+| modified beam search (beam size 4) | 18.62  | 23.88     | 33.80         | --epoch 29, --avg 24, --max-duration 100  |
+| fast beam search (set as default)  | 19.31  | 24.41     | 34.87         | --epoch 29, --avg 24, --max-duration 1500 |
+
+
+A pre-trained model and decoding logs can be found at <https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless2>

egs/wenetspeech/ASR/local/compute_fbank_musan.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compute_fbank_musan.py

egs/wenetspeech/ASR/local/compute_fbank_wenetspeech_dev_test.py

@@ -0,0 +1,93 @@
+#!/usr/bin/env python3
+# Copyright    2021  Johns Hopkins University (Piotr Żelasko)
+# Copyright    2021  Xiaomi Corp.             (Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+from pathlib import Path
+
+import torch
+from lhotse import (
+    CutSet,
+    KaldifeatFbank,
+    KaldifeatFbankConfig,
+    LilcomHdf5Writer,
+)
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_wenetspeech_dev_test():
+    in_out_dir = Path("data/fbank")
+    # number of workers in dataloader
+    num_workers = 42
+
+    # number of seconds in a batch
+    batch_duration = 600
+
+    subsets = ("S", "M", "DEV", "TEST_NET", "TEST_MEETING")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device))
+
+    logging.info(f"device: {device}")
+
+    for partition in subsets:
+        cuts_path = in_out_dir / f"cuts_{partition}.jsonl.gz"
+        if cuts_path.is_file():
+            logging.info(f"{cuts_path} exists - skipping")
+            continue
+
+        raw_cuts_path = in_out_dir / f"cuts_{partition}_raw.jsonl.gz"
+
+        logging.info(f"Loading {raw_cuts_path}")
+        cut_set = CutSet.from_file(raw_cuts_path)
+
+        logging.info("Computing features")
+
+        cut_set = cut_set.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=f"{in_out_dir}/feats_{partition}",
+            num_workers=num_workers,
+            batch_duration=batch_duration,
+            storage_type=LilcomHdf5Writer,
+        )
+        cut_set = cut_set.trim_to_supervisions(
+            keep_overlapping=False, min_duration=None
+        )
+
+        logging.info(f"Saving to {cuts_path}")
+        cut_set.to_file(cuts_path)
+
+
+def main():
+    formatter = (
+        "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    )
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    compute_fbank_wenetspeech_dev_test()
+
+
+if __name__ == "__main__":
+    main()

egs/wenetspeech/ASR/local/compute_fbank_wenetspeech_splits.py

@@ -0,0 +1,181 @@
+#!/usr/bin/env python3
+# Copyright    2021  Johns Hopkins University (Piotr Żelasko)
+# Copyright    2021  Xiaomi Corp.             (Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+from datetime import datetime
+from pathlib import Path
+
+import torch
+from lhotse import (
+    ChunkedLilcomHdf5Writer,
+    CutSet,
+    KaldifeatFbank,
+    KaldifeatFbankConfig,
+    set_audio_duration_mismatch_tolerance,
+    set_caching_enabled,
+)
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--training-subset",
+        type=str,
+        default="L",
+        help="The training subset for computing fbank feature.",
+    )
+
+    parser.add_argument(
+        "--num-workers",
+        type=int,
+        default=20,
+        help="Number of dataloading workers used for reading the audio.",
+    )
+
+    parser.add_argument(
+        "--batch-duration",
+        type=float,
+        default=600.0,
+        help="The maximum number of audio seconds in a batch."
+        "Determines batch size dynamically.",
+    )
+
+    parser.add_argument(
+        "--num-splits",
+        type=int,
+        required=True,
+        help="The number of splits of the L subset",
+    )
+
+    parser.add_argument(
+        "--start",
+        type=int,
+        default=0,
+        help="Process pieces starting from this number (inclusive).",
+    )
+
+    parser.add_argument(
+        "--stop",
+        type=int,
+        default=-1,
+        help="Stop processing pieces until this number (exclusive).",
+    )
+    return parser
+
+
+def compute_fbank_wenetspeech_splits(args):
+    subset = args.training_subset
+    subset = str(subset) 
+    num_splits = args.num_splits
+    output_dir = f"data/fbank/{subset}_split_{num_splits}"
+    output_dir = Path(output_dir)
+    assert output_dir.exists(), f"{output_dir} does not exist!"
+
+    num_digits = len(str(num_splits))
+
+    start = args.start
+    stop = args.stop
+    if stop < start:
+        stop = num_splits
+
+    stop = min(stop, num_splits)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device))
+    logging.info(f"device: {device}")
+
+    set_audio_duration_mismatch_tolerance(0.01)  # 10ms tolerance
+    set_caching_enabled(False)
+    for i in range(start, stop):
+        idx = f"{i + 1}".zfill(num_digits)
+        logging.info(f"Processing {idx}/{num_splits}")
+
+        cuts_path = output_dir / f"cuts_{subset}.{idx}.jsonl.gz"
+        if cuts_path.is_file():
+            logging.info(f"{cuts_path} exists - skipping")
+            continue
+
+        raw_cuts_path = output_dir / f"cuts_{subset}_raw.{idx}.jsonl.gz"
+
+        logging.info(f"Loading {raw_cuts_path}")
+        cut_set = CutSet.from_file(raw_cuts_path)
+
+        logging.info("Computing features")
+
+        cut_set = cut_set.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=f"{output_dir}/feats_{subset}_{idx}",
+            num_workers=args.num_workers,
+            batch_duration=args.batch_duration,
+            storage_type=ChunkedLilcomHdf5Writer,
+        )
+
+        logging.info("About to split cuts into smaller chunks.")
+        cut_set = cut_set.trim_to_supervisions(
+            keep_overlapping=False, min_duration=None
+        )
+
+        logging.info(f"Saving to {cuts_path}")
+        cut_set.to_file(cuts_path)
+        logging.info(f"Saved to {cuts_path}")
+
+
+def main():
+    now = datetime.now()
+    date_time = now.strftime("%Y-%m-%d-%H-%M-%S")
+
+    log_filename = "log-compute_fbank_wenetspeech_splits"
+    formatter = (
+        "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    )
+    log_filename = f"{log_filename}-{date_time}"
+
+    logging.basicConfig(
+        filename=log_filename,
+        format=formatter,
+        level=logging.INFO,
+        filemode="w",
+    )
+
+    console = logging.StreamHandler()
+    console.setLevel(logging.INFO)
+    console.setFormatter(logging.Formatter(formatter))
+    logging.getLogger("").addHandler(console)
+
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    compute_fbank_wenetspeech_splits(args)
+
+
+if __name__ == "__main__":
+    main()

egs/wenetspeech/ASR/local/display_manifest_statistics.py

@@ -0,0 +1,133 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+# 						                           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.
+
+"""
+This file displays duration statistics of utterances in a manifest.
+You can use the displayed value to choose minimum/maximum duration
+to remove short and long utterances during the training.
+See the function `remove_short_and_long_utt()`
+in ../../../librispeech/ASR/transducer/train.py
+for usage.
+"""
+
+
+from lhotse import load_manifest
+
+
+def main():
+    paths = [
+        #"./data/fbank/cuts_S.jsonl.gz",
+        #"./data/fbank/cuts_M.jsonl.gz",
+        "./data/fbank/cuts_L.jsonl.gz",
+        #"./data/fbank/cuts_DEV.jsonl.gz",
+        #"./data/fbank/cuts_TEST_NET.jsonl.gz",
+        #"./data/fbank/cuts_TEST_MEETING.jsonl.gz"
+    ]
+
+    for path in paths:
+        print(f"Starting display the statistics for {path}")
+        cuts = load_manifest(path)
+        cuts.describe()
+
+
+if __name__ == "__main__":
+    main()
+
+"""
+Starting display the statistics for ./data/fbank/cuts_S.jsonl.gz
+Duration statistics (seconds):
+mean    2.4
+std     1.8
+min     0.2
+25%     1.4
+50%     2.0
+75%     2.9
+99%     8.0
+99.5%   8.7
+99.9%   11.9
+max     405.1
+
+Starting display the statistics for ./data/fbank/cuts_M.jsonl.gz
+Cuts count: 4543341
+Total duration (hours): 3021.1
+Speech duration (hours): 3021.1 (100.0%)
+***
+Duration statistics (seconds):
+mean    2.4
+std     1.6
+min     0.2
+25%     1.4
+50%     2.0
+75%     2.9
+99%     8.0
+99.5%   8.8
+99.9%   12.1
+max     405.1
+
+Starting display the statistics for ./data/fbank/cuts_DEV.jsonl.gz
+Cuts count: 13825
+Total duration (hours): 20.0
+Speech duration (hours): 20.0 (100.0%)
+***
+Duration statistics (seconds):
+mean    5.2
+std     2.2
+min     1.0
+25%     3.3
+50%     4.9
+75%     7.0
+99%     9.6
+99.5%   9.8
+99.9%   10.0
+max     10.0
+
+Starting display the statistics for ./data/fbank/cuts_TEST_NET.jsonl.gz
+Cuts count: 24774
+Total duration (hours): 23.1
+Speech duration (hours): 23.1 (100.0%)
+***
+Duration statistics (seconds):
+mean    3.4
+std     2.6
+min     0.1
+25%     1.4
+50%     2.4
+75%     4.8
+99%     13.1
+99.5%   14.5
+99.9%   18.5
+max     33.3
+
+Starting display the statistics for ./data/fbank/cuts_TEST_MEETING.jsonl.gz
+Cuts count: 8370
+Total duration (hours): 15.2
+Speech duration (hours): 15.2 (100.0%)
+***
+Duration statistics (seconds):
+mean    6.5
+std     3.5
+min     0.8
+25%     3.7
+50%     5.8
+75%     8.8
+99%     15.2
+99.5%   16.0
+99.9%   18.8
+max     24.6
+
+"""

egs/wenetspeech/ASR/local/prepare_char.py

@@ -0,0 +1,246 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang,
+#                                                  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.
+
+
+"""
+This script takes as input `lang_dir`, which should contain::
+    - lang_dir/text,
+    - lang_dir/words.txt
+and generates the following files in the directory `lang_dir`:
+    - lexicon.txt
+    - lexicon_disambig.txt
+    - L.pt
+    - L_disambig.pt
+    - tokens.txt
+"""
+
+import argparse
+import re
+from pathlib import Path
+from typing import Dict, List
+
+import k2
+import torch
+from prepare_lang import (
+    Lexicon,
+    add_disambig_symbols,
+    add_self_loops,
+    write_lexicon,
+    write_mapping,
+)
+
+
+def lexicon_to_fst_no_sil(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format).
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    loop_state = 0  # words enter and leave from here
+    next_state = 1  # the next un-allocated state, will be incremented as we go
+
+    arcs = []
+
+    # The blank symbol <blk> is defined in local/train_bpe_model.py
+    assert token2id["<blk>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    for word, pieces in lexicon:
+        assert len(pieces) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        pieces = [
+            token2id[i] if i in token2id else token2id["<unk>"] for i in pieces
+        ]
+
+        for i in range(len(pieces) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, pieces[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last piece of this word
+        i = len(pieces) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, pieces[i], w, 0])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def contain_oov(token_sym_table: Dict[str, int], tokens: List[str]) -> bool:
+    """Check if all the given tokens are in token symbol table.
+    Args:
+      token_sym_table:
+        Token symbol table that contains all the valid tokens.
+      tokens:
+        A list of tokens.
+    Returns:
+      Return True if there is any token not in the token_sym_table,
+      otherwise False.
+    """
+    for tok in tokens:
+        if tok not in token_sym_table:
+            return True
+    return False
+
+
+def generate_lexicon(
+    token_sym_table: Dict[str, int], words: List[str]
+) -> Lexicon:
+    """Generate a lexicon from a word list and token_sym_table.
+    Args:
+      token_sym_table:
+        Token symbol table that mapping token to token ids.
+      words:
+        A list of strings representing words.
+    Returns:
+      Return a dict whose keys are words and values are the corresponding
+          tokens.
+    """
+    lexicon = []
+    for word in words:
+        chars = list(word.strip(" \t"))
+        if contain_oov(token_sym_table, chars):
+            continue
+        lexicon.append((word, chars))
+
+    # The OOV word is <UNK>
+    lexicon.append(("<UNK>", ["<unk>"]))
+    return lexicon
+
+
+def generate_tokens(text_file: str) -> Dict[str, int]:
+    """Generate tokens from the given text file.
+    Args:
+      text_file:
+        A file that contains text lines to generate tokens.
+    Returns:
+      Return a dict whose keys are tokens and values are token ids ranged
+      from 0 to len(keys) - 1.
+    """
+    tokens: Dict[str, int] = dict()
+    tokens["<blk>"] = 0
+    tokens["<sos/eos>"] = 1
+    tokens["<unk>"] = 2
+    whitespace = re.compile(r"([ \t\r\n]+)")
+    with open(text_file, "r", encoding="utf-8") as f:
+        for line in f:
+            line = re.sub(whitespace, "", line)
+            tokens_list = list(line)
+            for token in tokens_list:
+                if token not in tokens:
+                    tokens[token] = len(tokens)
+    return tokens
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--lang-dir", type=str, help="The lang directory.")
+    args = parser.parse_args()
+
+    lang_dir = Path(args.lang_dir)
+    text_file = lang_dir / "text"
+
+    word_sym_table = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+    words = word_sym_table.symbols
+
+    excluded = ["<eps>", "!SIL", "<SPOKEN_NOISE>", "<UNK>", "#0", "<s>", "</s>"]
+    for w in excluded:
+        if w in words:
+            words.remove(w)
+
+    token_sym_table = generate_tokens(text_file)
+
+    lexicon = generate_lexicon(token_sym_table, words)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    next_token_id = max(token_sym_table.values()) + 1
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in token_sym_table
+        token_sym_table[disambig] = next_token_id
+        next_token_id += 1
+
+    word_sym_table.add("#0")
+    word_sym_table.add("<s>")
+    word_sym_table.add("</s>")
+
+    write_mapping(lang_dir / "tokens.txt", token_sym_table)
+
+    write_lexicon(lang_dir / "lexicon.txt", lexicon)
+    write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst_no_sil(
+        lexicon,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+    )
+
+    L_disambig = lexicon_to_fst_no_sil(
+        lexicon_disambig,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), lang_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), lang_dir / "L_disambig.pt")
+
+
+if __name__ == "__main__":
+    main()

egs/wenetspeech/ASR/local/prepare_lang.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang.py

egs/wenetspeech/ASR/local/preprocess_wenetspeech.py

@@ -0,0 +1,120 @@
+#!/usr/bin/env python3
+# Copyright    2021  Johns Hopkins University (Piotr Żelasko)
+# Copyright    2021  Xiaomi Corp.             (Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+import re
+from pathlib import Path
+
+from lhotse import CutSet, SupervisionSegment
+from lhotse.recipes.utils import read_manifests_if_cached
+
+# Similar text filtering and normalization procedure as in:
+# https://github.com/SpeechColab/WenetSpeech/blob/main/toolkits/kaldi/wenetspeech_data_prep.sh
+
+
+def normalize_text(
+    utt: str,
+    # punct_pattern=re.compile(r"<(COMMA|PERIOD|QUESTIONMARK|EXCLAMATIONPOINT)>"),
+    punct_pattern=re.compile(r"<(PERIOD|QUESTIONMARK|EXCLAMATIONPOINT)>"),
+    whitespace_pattern=re.compile(r"\s\s+"),
+) -> str:
+    return whitespace_pattern.sub(" ", punct_pattern.sub("", utt))
+
+
+def has_no_oov(
+    sup: SupervisionSegment,
+    oov_pattern=re.compile(r"<(SIL|MUSIC|NOISE|OTHER)>"),
+) -> bool:
+    return oov_pattern.search(sup.text) is None
+
+
+def preprocess_wenet_speech():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+    output_dir.mkdir(exist_ok=True)
+
+    dataset_parts = (
+        "L",
+        "M",
+        "S",
+        "DEV",
+        "TEST_NET",
+        "TEST_MEETING",
+    )
+
+    logging.info("Loading manifest (may take 10 minutes)")
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        suffix="jsonl.gz",
+    )
+    assert manifests is not None
+
+    for partition, m in manifests.items():
+        logging.info(f"Processing {partition}")
+        raw_cuts_path = output_dir / f"cuts_{partition}_raw.jsonl.gz"
+        if raw_cuts_path.is_file():
+            logging.info(f"{partition} already exists - skipping")
+            continue
+
+        # Note this step makes the recipe different than LibriSpeech:
+        # We must filter out some utterances and remove punctuation
+        # to be consistent with Kaldi.
+        logging.info("Filtering OOV utterances from supervisions")
+        m["supervisions"] = m["supervisions"].filter(has_no_oov)
+        logging.info(f"Normalizing text in {partition}")
+        for sup in m["supervisions"]:
+            text = str(sup.text)
+            logging.info(f"Original text: {text}")
+            sup.text = normalize_text(sup.text)
+            text = str(sup.text)
+            logging.info(f"Normalize text: {text}")
+
+        # Create long-recording cut manifests.
+        logging.info(f"Processing {partition}")
+        cut_set = CutSet.from_manifests(
+            recordings=m["recordings"],
+            supervisions=m["supervisions"],
+        )
+        # Run data augmentation that needs to be done in the
+        # time domain.
+        if partition not in ["DEV", "TEST_NET", "TEST_MEETING"]:
+            logging.info(
+                f"Speed perturb for {partition} with factors 0.9 and 1.1 "
+                "(Perturbing may take 8 minutes and saving may take 20 minutes)"
+            )
+            cut_set = (
+                cut_set
+                + cut_set.perturb_speed(0.9)
+                + cut_set.perturb_speed(1.1)
+            )
+        logging.info(f"Saving to {raw_cuts_path}")
+        cut_set.to_file(raw_cuts_path)
+
+
+def main():
+    formatter = (
+        "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    )
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    preprocess_wenet_speech()
+
+
+if __name__ == "__main__":
+    main()

egs/wenetspeech/ASR/local/text2token.py

@@ -0,0 +1,196 @@
+#!/usr/bin/env python3
+# Copyright    2017  Johns Hopkins University   (authors: Shinji Watanabe)
+#              2022  Xiaomi Corp.               (authors: Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import codecs
+import re
+import sys
+from typing import List
+
+from pypinyin import lazy_pinyin, pinyin
+
+is_python2 = sys.version_info[0] == 2
+
+
+def exist_or_not(i, match_pos):
+    start_pos = None
+    end_pos = None
+    for pos in match_pos:
+        if pos[0] <= i < pos[1]:
+            start_pos = pos[0]
+            end_pos = pos[1]
+            break
+
+    return start_pos, end_pos
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        description="convert raw text to tokenized text",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument(
+        "--nchar",
+        "-n",
+        default=1,
+        type=int,
+        help="number of characters to split, i.e., \
+                        aabb -> a a b b with -n 1 and aa bb with -n 2",
+    )
+    parser.add_argument(
+        "--skip-ncols", "-s", default=0, type=int, help="skip first n columns"
+    )
+    parser.add_argument(
+        "--space", default="<space>", type=str, help="space symbol"
+    )
+    parser.add_argument(
+        "--non-lang-syms",
+        "-l",
+        default=None,
+        type=str,
+        help="list of non-linguistic symobles, e.g., <NOISE> etc.",
+    )
+    parser.add_argument(
+        "text", type=str, default=False, nargs="?", help="input text"
+    )
+    parser.add_argument(
+        "--trans_type",
+        "-t",
+        type=str,
+        default="char",
+        choices=["char", "pinyin", "lazy_pinyin"],
+        help="""Transcript type. char/pinyin/lazy_pinyin""",
+    )
+    return parser
+
+
+def token2id(
+    texts, token_table, token_type: str = "lazy_pinyin", oov: str = "<unk>"
+) -> List[List[int]]:
+    """Convert token to id.
+    Args:
+      texts:
+        The input texts, it refers to the chinese text here.
+      token_table:
+        The token table is built based on "data/lang_xxx/token.txt"
+      token_type:
+        The type of token, such as "pinyin" and "lazy_pinyin".
+      oov:
+        Out of vocabulary token. When a word(token) in the transcript
+        does not exist in the token list, it is replaced with `oov`.
+
+    Returns:
+      The list of ids for the input texts.
+    """
+    if texts is None:
+        raise ValueError("texts can't be None!")
+    else:
+        oov_id = token_table[oov]
+        ids: List[List[int]] = []
+        for text in texts:
+            chars_list = list(str(text))
+            if token_type == "lazy_pinyin":
+                text = lazy_pinyin(chars_list)
+                sub_ids = [
+                    token_table[txt] if txt in token_table else oov_id
+                    for txt in text
+                ]
+                ids.append(sub_ids)
+            else:  # token_type = "pinyin"
+                text = pinyin(chars_list)
+                sub_ids = [
+                    token_table[txt[0]] if txt[0] in token_table else oov_id
+                    for txt in text
+                ]
+                ids.append(sub_ids)
+        return ids
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    rs = []
+    if args.non_lang_syms is not None:
+        with codecs.open(args.non_lang_syms, "r", encoding="utf-8") as f:
+            nls = [x.rstrip() for x in f.readlines()]
+            rs = [re.compile(re.escape(x)) for x in nls]
+
+    if args.text:
+        f = codecs.open(args.text, encoding="utf-8")
+    else:
+        f = codecs.getreader("utf-8")(
+            sys.stdin if is_python2 else sys.stdin.buffer
+        )
+
+    sys.stdout = codecs.getwriter("utf-8")(
+        sys.stdout if is_python2 else sys.stdout.buffer
+    )
+    line = f.readline()
+    n = args.nchar
+    while line:
+        x = line.split()
+        print(" ".join(x[: args.skip_ncols]), end=" ")
+        a = " ".join(x[args.skip_ncols :])  # noqa E203
+
+        # get all matched positions
+        match_pos = []
+        for r in rs:
+            i = 0
+            while i >= 0:
+                m = r.search(a, i)
+                if m:
+                    match_pos.append([m.start(), m.end()])
+                    i = m.end()
+                else:
+                    break
+        if len(match_pos) > 0:
+            chars = []
+            i = 0
+            while i < len(a):
+                start_pos, end_pos = exist_or_not(i, match_pos)
+                if start_pos is not None:
+                    chars.append(a[start_pos:end_pos])
+                    i = end_pos
+                else:
+                    chars.append(a[i])
+                    i += 1
+            a = chars
+
+        if args.trans_type == "pinyin":
+            a = pinyin(list(str(a)))
+            a = [one[0] for one in a]
+
+        if args.trans_type == "lazy_pinyin":
+            a = lazy_pinyin(list(str(a)))
+
+        a = [a[j : j + n] for j in range(0, len(a), n)]  # noqa E203
+
+        a_flat = []
+        for z in a:
+            a_flat.append("".join(z))
+
+        a_chars = [z.replace(" ", args.space) for z in a_flat]
+
+        print("".join(a_chars))
+        line = f.readline()
+
+
+if __name__ == "__main__":
+    main()

egs/wenetspeech/ASR/prepare.sh

@@ -0,0 +1,210 @@
+#!/usr/bin/env bash
+
+set -eou pipefail
+
+nj=15
+stage=0
+stop_stage=100
+
+# Split L subset to this number of pieces
+# This is to avoid OOM during feature extraction.
+num_splits=1000
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/WenetSpeech
+#      You can find audio, WenetSpeech.json inside it.
+#      You can apply for the download credentials by following
+#      https://github.com/wenet-e2e/WenetSpeech#download
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  [ ! -e $dl_dir/WenetSpeech ] && mkdir -p $dl_dir/WenetSpeech
+
+  # If you have pre-downloaded it to /path/to/WenetSpeech,
+  # you can create a symlink
+  #
+  # ln -sfv /path/to/WenetSpeech $dl_dir/WenetSpeech
+  #
+  if [ ! -d $dl_dir/WenetSpeech/wenet_speech ] && [ ! -f $dl_dir/WenetSpeech/metadata/v1.list ]; then
+    log "Stage 0: should download WenetSpeech first"
+    exit 1;
+  fi
+
+  # If you have pre-downloaded it to /path/to/musan,
+  # you can create a symlink
+  #
+  #ln -sfv /path/to/musan $dl_dir/musan
+
+  if [ ! -d $dl_dir/musan ]; then
+    lhotse download musan $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare WenetSpeech manifest"
+  # We assume that you have downloaded the WenetSpeech corpus
+  # to $dl_dir/WenetSpeech
+  mkdir -p data/manifests
+  lhotse prepare wenet-speech $dl_dir/WenetSpeech data/manifests -j $nj
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to data/musan
+  mkdir -p data/manifests
+  lhotse prepare musan $dl_dir/musan data/manifests
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Preprocess WenetSpeech manifest"
+  if [ ! -f data/fbank/.preprocess_complete ]; then
+    python3 ./local/preprocess_wenetspeech.py
+    touch data/fbank/.preprocess_complete
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute features for DEV and TEST subsets of WenetSpeech (may take 2 minutes)"
+  python3 ./local/compute_fbank_wenetspeech_dev_test.py
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Split S subset into ${num_splits} pieces"
+  split_dir=data/fbank/S_split_${num_splits}_test
+  if [ ! -f $split_dir/.split_completed ]; then
+    lhotse split $num_splits ./data/fbank/cuts_S_raw.jsonl.gz $split_dir
+    touch $split_dir/.split_completed
+  fi
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Split M subset into ${num_splits} piece"
+  split_dir=data/fbank/M_split_${num_splits}
+  if [ ! -f $split_dir/.split_completed ]; then
+    lhotse split $num_splits ./data/fbank/cuts_M_raw.jsonl.gz $split_dir
+    touch $split_dir/.split_completed
+  fi
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+  log "Stage 7: Split L subset into ${num_splits} pieces"
+  split_dir=data/fbank/L_split_${num_splits}
+  if [ ! -f $split_dir/.split_completed ]; then
+    lhotse split $num_splits ./data/fbank/cuts_L_raw.jsonl.gz $split_dir
+    touch $split_dir/.split_completed
+  fi
+fi
+
+if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
+  log "Stage 8: Compute features for S"
+  python3 ./local/compute_fbank_wenetspeech_splits.py \
+    --training-subset S \
+    --num-workers 20 \
+    --batch-duration 600 \
+    --start 0 \
+    --num-splits $num_splits
+fi
+
+if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then
+  log "Stage 9: Compute features for M"
+  python3 ./local/compute_fbank_wenetspeech_splits.py \
+    --training-subset M \
+    --num-workers 20 \
+    --batch-duration 600 \
+    --start 0 \
+    --num-splits $num_splits
+fi
+
+if [ $stage -le 10 ] && [ $stop_stage -ge 10 ]; then
+  log "Stage 10: Compute features for L"
+  python3 ./local/compute_fbank_wenetspeech_splits.py \
+    --training-subset L \
+    --num-workers 20 \
+    --batch-duration 600 \
+    --start 0 \
+    --num-splits $num_splits
+fi
+
+if [ $stage -le 11 ] && [ $stop_stage -ge 11 ]; then
+  log "Stage 11: Combine features for S"
+  if [ ! -f data/fbank/cuts_S.jsonl.gz ]; then
+    pieces=$(find data/fbank/S_split_1000 -name "cuts_S.*.jsonl.gz")
+    lhotse combine $pieces data/fbank/cuts_S.jsonl.gz
+  fi
+fi
+
+if [ $stage -le 12 ] && [ $stop_stage -ge 12 ]; then
+  log "Stage 12: Combine features for M"
+  if [ ! -f data/fbank/cuts_M.jsonl.gz ]; then
+    pieces=$(find data/fbank/M_split_1000 -name "cuts_M.*.jsonl.gz")
+    lhotse combine $pieces data/fbank/cuts_M.jsonl.gz
+  fi
+fi
+
+if [ $stage -le 13 ] && [ $stop_stage -ge 13 ]; then
+  log "Stage 13: Combine features for L"
+  if [ ! -f data/fbank/cuts_L.jsonl.gz ]; then
+    pieces=$(find data/fbank/L_split_1000 -name "cuts_L.*.jsonl.gz")
+    lhotse combine $pieces data/fbank/cuts_L.jsonl.gz
+  fi
+fi
+
+if [ $stage -le 14 ] && [ $stop_stage -ge 14 ]; then
+  log "Stage 14: Compute fbank for musan"
+  mkdir -p data/fbank
+  ./local/compute_fbank_musan.py
+fi
+
+if [ $stage -le 15 ] && [ $stop_stage -ge 15 ]; then
+  log "Stage 15: Prepare char based lang"
+  lang_char_dir=data/lang_char
+  mkdir -p $lang_char_dir
+
+  # Prepare text.
+  if [ ! -f $lang_char_dir/text ]; then
+    gunzip -c data/manifests/supervisions_L.jsonl.gz \
+      | jq 'text' | sed 's/"//g' \
+      | ./local/text2token.py -t "char" > $lang_char_dir/text
+    # if use the whole text to generate the text, you can use
+    # the following command:
+    # grep "\"text\":" $dl_dir/WenetSpeech/WenetSpeech.json |
+    #   sed -e 's/["text:\t ]*//g' > $lang_char_dir/text
+  fi
+fi
+
+if [ $stage -le 16 ] && [ $stop_stage -ge 16 ]; then
+  log "Stage 16: Prepare char based L_disambig.pt"
+  if [ ! -f data/lang_char/L_disambig.pt ]; then
+    python ./local/prepare_char.py \
+      --lang-dir data/lang_char
+  fi
+fi

egs/wenetspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py

@@ -0,0 +1,450 @@
+# 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,
+    set_caching_enabled,
+)
+from lhotse.dataset import (
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SingleCutSampler,
+    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
+
+set_caching_enabled(False)
+torch.set_num_threads(1)
+
+
+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=300,
+            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(
+            "--lazy-load",
+            type=str2bool,
+            default=True,
+            help="lazily open CutSets to avoid OOM (for L|XL subset)",
+        )
+
+        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 / "cuts_musan.json.gz"
+        )
+
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            transforms.append(
+                CutMix(
+                    cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True
+                )
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(
+                f"Time warp factor: {self.args.spec_aug_time_warp_factor}"
+            )
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            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=30000,
+                drop_last=True,
+            )
+        else:
+            logging.info("Using SingleCutSampler.")
+            train_sampler = SingleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        # '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,
+            rank=0,
+            world_size=1,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+
+        from lhotse.dataset.iterable_dataset import IterableDatasetWrapper
+
+        dev_iter_dataset = IterableDatasetWrapper(
+            dataset=validate,
+            sampler=valid_sampler,
+        )
+        valid_dl = DataLoader(
+            dev_iter_dataset,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            rank=0,
+            world_size=1,
+            shuffle=False,
+        )
+        from lhotse.dataset.iterable_dataset import IterableDatasetWrapper
+
+        test_iter_dataset = IterableDatasetWrapper(
+            dataset=test,
+            sampler=sampler,
+        )
+        test_dl = DataLoader(
+            test_iter_dataset,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get train cuts")
+        if self.args.lazy_load:
+            logging.info("use lazy cuts")
+            cuts_train = CutSet.from_jsonl_lazy(
+                self.args.manifest_dir
+                / f"cuts_{self.args.training_subset}.jsonl.gz"
+            )
+        else:
+            cuts_train = CutSet.from_file(
+                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(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(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(
+            self.args.manifest_dir / "cuts_TEST_MEETING.jsonl.gz"
+        )

egs/wenetspeech/ASR/pruned_transducer_stateless2/beam_search.py

@@ -0,0 +1,767 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Dict, List, Optional
+
+import k2
+import torch
+from model import Transducer
+
+from icefall.decode import one_best_decoding
+from icefall.utils import get_texts
+
+
+def fast_beam_search(
+    model: Transducer,
+    decoding_graph: k2.Fsa,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+) -> List[List[int]]:
+    """It limits the maximum number of symbols per frame to 1.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      decoding_graph:
+        Decoding graph used for decoding, may be a TrivialGraph or a HLG.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      encoder_out_lens:
+        A tensor of shape (N,) containing the number of frames in `encoder_out`
+        before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+    Returns:
+      Return the decoded result.
+    """
+    assert encoder_out.ndim == 3
+
+    context_size = model.decoder.context_size
+    vocab_size = model.decoder.vocab_size
+
+    B, T, C = encoder_out.shape
+
+    config = k2.RnntDecodingConfig(
+        vocab_size=vocab_size,
+        decoder_history_len=context_size,
+        beam=beam,
+        max_contexts=max_contexts,
+        max_states=max_states,
+    )
+    individual_streams = []
+    for i in range(B):
+        individual_streams.append(k2.RnntDecodingStream(decoding_graph))
+    decoding_streams = k2.RnntDecodingStreams(individual_streams, config)
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    for t in range(T):
+        # shape is a RaggedShape of shape (B, context)
+        # contexts is a Tensor of shape (shape.NumElements(), context_size)
+        shape, contexts = decoding_streams.get_contexts()
+        # `nn.Embedding()` in torch below v1.7.1 supports only torch.int64
+        contexts = contexts.to(torch.int64)
+        # decoder_out is of shape (shape.NumElements(), 1, decoder_out_dim)
+        decoder_out = model.decoder(contexts, need_pad=False)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # current_encoder_out is of shape
+        # (shape.NumElements(), 1, joiner_dim)
+        # fmt: off
+        current_encoder_out = torch.index_select(
+            encoder_out[:, t:t + 1, :], 0, shape.row_ids(1).to(torch.int64)
+        )
+        # fmt: on
+        logits = model.joiner(
+            current_encoder_out.unsqueeze(2),
+            decoder_out.unsqueeze(1),
+            project_input=False,
+        )
+        logits = logits.squeeze(1).squeeze(1)
+        log_probs = logits.log_softmax(dim=-1)
+        decoding_streams.advance(log_probs)
+    decoding_streams.terminate_and_flush_to_streams()
+    lattice = decoding_streams.format_output(encoder_out_lens.tolist())
+
+    best_path = one_best_decoding(lattice)
+    hyps = get_texts(best_path)
+    return hyps
+
+
+def greedy_search(
+    model: Transducer, encoder_out: torch.Tensor, max_sym_per_frame: int
+) -> List[int]:
+    """Greedy search for a single utterance.
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      max_sym_per_frame:
+        Maximum number of symbols per frame. If it is set to 0, the WER
+        would be 100%.
+    Returns:
+      Return the decoded result.
+    """
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+
+    device = model.device
+
+    decoder_input = torch.tensor(
+        [blank_id] * context_size, device=device, dtype=torch.int64
+    ).reshape(1, context_size)
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    decoder_out = model.joiner.decoder_proj(decoder_out)
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    T = encoder_out.size(1)
+    t = 0
+    hyp = [blank_id] * context_size
+
+    # Maximum symbols per utterance.
+    max_sym_per_utt = 1000
+
+    # symbols per frame
+    sym_per_frame = 0
+
+    # symbols per utterance decoded so far
+    sym_per_utt = 0
+
+    while t < T and sym_per_utt < max_sym_per_utt:
+        if sym_per_frame >= max_sym_per_frame:
+            sym_per_frame = 0
+            t += 1
+            continue
+
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :].unsqueeze(2)
+        # fmt: on
+        logits = model.joiner(
+            current_encoder_out, decoder_out.unsqueeze(1), project_input=False
+        )
+        # logits is (1, 1, 1, vocab_size)
+
+        y = logits.argmax().item()
+        if y != blank_id:
+            hyp.append(y)
+            decoder_input = torch.tensor(
+                [hyp[-context_size:]], device=device
+            ).reshape(1, context_size)
+
+            decoder_out = model.decoder(decoder_input, need_pad=False)
+            decoder_out = model.joiner.decoder_proj(decoder_out)
+
+            sym_per_utt += 1
+            sym_per_frame += 1
+        else:
+            sym_per_frame = 0
+            t += 1
+    hyp = hyp[context_size:]  # remove blanks
+
+    return hyp
+
+
+def greedy_search_batch(
+    model: Transducer, encoder_out: torch.Tensor
+) -> List[List[int]]:
+    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C), where N >= 1.
+    Returns:
+      Return a list-of-list of token IDs containing the decoded results.
+      len(ans) equals to encoder_out.size(0).
+    """
+    assert encoder_out.ndim == 3
+    assert encoder_out.size(0) >= 1, encoder_out.size(0)
+
+    device = model.device
+
+    batch_size = encoder_out.size(0)
+    T = encoder_out.size(1)
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+
+    hyps = [[blank_id] * context_size for _ in range(batch_size)]
+
+    decoder_input = torch.tensor(
+        hyps,
+        device=device,
+        dtype=torch.int64,
+    )  # (batch_size, context_size)
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    decoder_out = model.joiner.decoder_proj(decoder_out)
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    # decoder_out: (batch_size, 1, decoder_out_dim)
+    for t in range(T):
+        current_encoder_out = encoder_out[:, t : t + 1, :].unsqueeze(2)  # noqa
+        # current_encoder_out's shape: (batch_size, 1, 1, encoder_out_dim)
+        logits = model.joiner(
+            current_encoder_out, decoder_out.unsqueeze(1), project_input=False
+        )
+        # logits'shape (batch_size, 1, 1, vocab_size)
+
+        logits = logits.squeeze(1).squeeze(1)  # (batch_size, vocab_size)
+        assert logits.ndim == 2, logits.shape
+        y = logits.argmax(dim=1).tolist()
+        emitted = False
+        for i, v in enumerate(y):
+            if v != blank_id:
+                hyps[i].append(v)
+                emitted = True
+        if emitted:
+            # update decoder output
+            decoder_input = [h[-context_size:] for h in hyps]
+            decoder_input = torch.tensor(
+                decoder_input,
+                device=device,
+                dtype=torch.int64,
+            )
+            decoder_out = model.decoder(decoder_input, need_pad=False)
+            decoder_out = model.joiner.decoder_proj(decoder_out)
+
+    ans = [h[context_size:] for h in hyps]
+    return ans
+
+
+@dataclass
+class Hypothesis:
+    # The predicted tokens so far.
+    # Newly predicted tokens are appended to `ys`.
+    ys: List[int]
+
+    # The log prob of ys.
+    # It contains only one entry.
+    log_prob: torch.Tensor
+
+    @property
+    def key(self) -> str:
+        """Return a string representation of self.ys"""
+        return "_".join(map(str, self.ys))
+
+
+class HypothesisList(object):
+    def __init__(self, data: Optional[Dict[str, Hypothesis]] = None) -> None:
+        """
+        Args:
+          data:
+            A dict of Hypotheses. Its key is its `value.key`.
+        """
+        if data is None:
+            self._data = {}
+        else:
+            self._data = data
+
+    @property
+    def data(self) -> Dict[str, Hypothesis]:
+        return self._data
+
+    def add(self, hyp: Hypothesis) -> None:
+        """Add a Hypothesis to `self`.
+
+        If `hyp` already exists in `self`, its probability is updated using
+        `log-sum-exp` with the existed one.
+
+        Args:
+          hyp:
+            The hypothesis to be added.
+        """
+        key = hyp.key
+        if key in self:
+            old_hyp = self._data[key]  # shallow copy
+            torch.logaddexp(
+                old_hyp.log_prob, hyp.log_prob, out=old_hyp.log_prob
+            )
+        else:
+            self._data[key] = hyp
+
+    def get_most_probable(self, length_norm: bool = False) -> Hypothesis:
+        """Get the most probable hypothesis, i.e., the one with
+        the largest `log_prob`.
+
+        Args:
+          length_norm:
+            If True, the `log_prob` of a hypothesis is normalized by the
+            number of tokens in it.
+        Returns:
+          Return the hypothesis that has the largest `log_prob`.
+        """
+        if length_norm:
+            return max(
+                self._data.values(), key=lambda hyp: hyp.log_prob / len(hyp.ys)
+            )
+        else:
+            return max(self._data.values(), key=lambda hyp: hyp.log_prob)
+
+    def remove(self, hyp: Hypothesis) -> None:
+        """Remove a given hypothesis.
+
+        Caution:
+          `self` is modified **in-place**.
+
+        Args:
+          hyp:
+            The hypothesis to be removed from `self`.
+            Note: It must be contained in `self`. Otherwise,
+            an exception is raised.
+        """
+        key = hyp.key
+        assert key in self, f"{key} does not exist"
+        del self._data[key]
+
+    def filter(self, threshold: torch.Tensor) -> "HypothesisList":
+        """Remove all Hypotheses whose log_prob is less than threshold.
+
+        Caution:
+          `self` is not modified. Instead, a new HypothesisList is returned.
+
+        Returns:
+          Return a new HypothesisList containing all hypotheses from `self`
+          with `log_prob` being greater than the given `threshold`.
+        """
+        ans = HypothesisList()
+        for _, hyp in self._data.items():
+            if hyp.log_prob > threshold:
+                ans.add(hyp)  # shallow copy
+        return ans
+
+    def topk(self, k: int) -> "HypothesisList":
+        """Return the top-k hypothesis."""
+        hyps = list(self._data.items())
+
+        hyps = sorted(hyps, key=lambda h: h[1].log_prob, reverse=True)[:k]
+
+        ans = HypothesisList(dict(hyps))
+        return ans
+
+    def __contains__(self, key: str):
+        return key in self._data
+
+    def __iter__(self):
+        return iter(self._data.values())
+
+    def __len__(self) -> int:
+        return len(self._data)
+
+    def __str__(self) -> str:
+        s = []
+        for key in self:
+            s.append(key)
+        return ", ".join(s)
+
+
+def _get_hyps_shape(hyps: List[HypothesisList]) -> k2.RaggedShape:
+    """Return a ragged shape with axes [utt][num_hyps].
+
+    Args:
+      hyps:
+        len(hyps) == batch_size. It contains the current hypothesis for
+        each utterance in the batch.
+    Returns:
+      Return a ragged shape with 2 axes [utt][num_hyps]. Note that
+      the shape is on CPU.
+    """
+    num_hyps = [len(h) for h in hyps]
+
+    # torch.cumsum() is inclusive sum, so we put a 0 at the beginning
+    # to get exclusive sum later.
+    num_hyps.insert(0, 0)
+
+    num_hyps = torch.tensor(num_hyps)
+    row_splits = torch.cumsum(num_hyps, dim=0, dtype=torch.int32)
+    ans = k2.ragged.create_ragged_shape2(
+        row_splits=row_splits, cached_tot_size=row_splits[-1].item()
+    )
+    return ans
+
+
+def modified_beam_search(
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    beam: int = 4,
+) -> List[List[int]]:
+    """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).
+      beam:
+        Number of active paths during the beam search.
+    Returns:
+      Return a list-of-list of token IDs. ans[i] is the decoding results
+      for the i-th utterance.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+
+    batch_size = encoder_out.size(0)
+    T = encoder_out.size(1)
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    device = model.device
+    B = [HypothesisList() for _ in range(batch_size)]
+    for i in range(batch_size):
+        B[i].add(
+            Hypothesis(
+                ys=[blank_id] * context_size,
+                log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+            )
+        )
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    for t in range(T):
+        current_encoder_out = encoder_out[:, t : t + 1, :].unsqueeze(2)  # noqa
+        # current_encoder_out's shape is (batch_size, 1, 1, encoder_out_dim)
+
+        hyps_shape = _get_hyps_shape(B).to(device)
+
+        A = [list(b) for b in B]
+        B = [HypothesisList() for _ in range(batch_size)]
+
+        ys_log_probs = torch.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)
+
+        log_probs = logits.log_softmax(dim=-1)  # (num_hyps, vocab_size)
+
+        log_probs.add_(ys_log_probs)
+
+        vocab_size = log_probs.size(-1)
+
+        log_probs = log_probs.reshape(-1)
+
+        row_splits = hyps_shape.row_splits(1) * vocab_size
+        log_probs_shape = k2.ragged.create_ragged_shape2(
+            row_splits=row_splits, cached_tot_size=log_probs.numel()
+        )
+        ragged_log_probs = k2.RaggedTensor(
+            shape=log_probs_shape, value=log_probs
+        )
+
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam)
+
+            #topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+            topk_hyp_indexes = torch.div(topk_indexes, vocab_size, rounding_mode="trunc")
+            topk_token_indexes = (topk_indexes % vocab_size).tolist()
+
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+
+                new_ys = hyp.ys[:]
+                new_token = topk_token_indexes[k]
+                if new_token != blank_id:
+                    new_ys.append(new_token)
+
+                new_log_prob = topk_log_probs[k]
+                new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob)
+                B[i].add(new_hyp)
+
+    best_hyps = [b.get_most_probable(length_norm=True) for b in B]
+    ans = [h.ys[context_size:] for h in best_hyps]
+
+    return ans
+
+
+def _deprecated_modified_beam_search(
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    beam: int = 4,
+) -> List[int]:
+    """It limits the maximum number of symbols per frame to 1.
+
+    It decodes only one utterance at a time. We keep it only for reference.
+    The function :func:`modified_beam_search` should be preferred as it
+    supports batch decoding.
+
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      beam:
+        Beam size.
+    Returns:
+      Return the decoded result.
+    """
+
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+
+    device = model.device
+
+    T = encoder_out.size(1)
+
+    B = HypothesisList()
+    B.add(
+        Hypothesis(
+            ys=[blank_id] * context_size,
+            log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+        )
+    )
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    for t in range(T):
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :].unsqueeze(2)
+        # current_encoder_out is of shape (1, 1, 1, encoder_out_dim)
+        # fmt: on
+        A = list(B)
+        B = HypothesisList()
+
+        ys_log_probs = torch.cat([hyp.log_prob.reshape(1, 1) for hyp in A])
+        # ys_log_probs is of shape (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyp in A],
+            device=device,
+            dtype=torch.int64,
+        )
+        # decoder_input is of shape (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+        # decoder_output is of shape (num_hyps, 1, 1, joiner_dim)
+
+        current_encoder_out = current_encoder_out.expand(
+            decoder_out.size(0), 1, 1, -1
+        )  # (num_hyps, 1, 1, encoder_out_dim)
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+            project_input=False,
+        )
+        # logits is of shape (num_hyps, 1, 1, vocab_size)
+        logits = logits.squeeze(1).squeeze(1)
+
+        # now logits is of shape (num_hyps, vocab_size)
+        log_probs = logits.log_softmax(dim=-1)
+
+        log_probs.add_(ys_log_probs)
+
+        log_probs = log_probs.reshape(-1)
+        topk_log_probs, topk_indexes = log_probs.topk(beam)
+
+        # topk_hyp_indexes are indexes into `A`
+        topk_hyp_indexes = topk_indexes // logits.size(-1)
+        topk_token_indexes = topk_indexes % logits.size(-1)
+
+        topk_hyp_indexes = topk_hyp_indexes.tolist()
+        topk_token_indexes = topk_token_indexes.tolist()
+
+        for i in range(len(topk_hyp_indexes)):
+            hyp = A[topk_hyp_indexes[i]]
+            new_ys = hyp.ys[:]
+            new_token = topk_token_indexes[i]
+            if new_token != blank_id:
+                new_ys.append(new_token)
+            new_log_prob = topk_log_probs[i]
+            new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob)
+            B.add(new_hyp)
+
+    best_hyp = B.get_most_probable(length_norm=True)
+    ys = best_hyp.ys[context_size:]  # [context_size:] to remove blanks
+
+    return ys
+
+
+def beam_search(
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    beam: int = 4,
+) -> List[int]:
+    """
+    It implements Algorithm 1 in https://arxiv.org/pdf/1211.3711.pdf
+
+    espnet/nets/beam_search_transducer.py#L247 is used as a reference.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      beam:
+        Beam size.
+    Returns:
+      Return the decoded result.
+    """
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+
+    device = model.device
+
+    decoder_input = torch.tensor(
+        [blank_id] * context_size,
+        device=device,
+        dtype=torch.int64,
+    ).reshape(1, context_size)
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    decoder_out = model.joiner.decoder_proj(decoder_out)
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    T = encoder_out.size(1)
+    t = 0
+
+    B = HypothesisList()
+    B.add(Hypothesis(ys=[blank_id] * context_size, log_prob=0.0))
+
+    max_sym_per_utt = 20000
+
+    sym_per_utt = 0
+
+    decoder_cache: Dict[str, torch.Tensor] = {}
+
+    while t < T and sym_per_utt < max_sym_per_utt:
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :].unsqueeze(2)
+        # fmt: on
+        A = B
+        B = HypothesisList()
+
+        joint_cache: Dict[str, torch.Tensor] = {}
+
+        # TODO(fangjun): Implement prefix search to update the `log_prob`
+        # of hypotheses in A
+
+        while True:
+            y_star = A.get_most_probable()
+            A.remove(y_star)
+
+            cached_key = y_star.key
+
+            if cached_key not in decoder_cache:
+                decoder_input = torch.tensor(
+                    [y_star.ys[-context_size:]],
+                    device=device,
+                    dtype=torch.int64,
+                ).reshape(1, context_size)
+
+                decoder_out = model.decoder(decoder_input, need_pad=False)
+                decoder_out = model.joiner.decoder_proj(decoder_out)
+                decoder_cache[cached_key] = decoder_out
+            else:
+                decoder_out = decoder_cache[cached_key]
+
+            cached_key += f"-t-{t}"
+            if cached_key not in joint_cache:
+                logits = model.joiner(
+                    current_encoder_out,
+                    decoder_out.unsqueeze(1),
+                    project_input=False,
+                )
+
+                # TODO(fangjun): Scale the blank posterior
+                log_prob = logits.log_softmax(dim=-1)
+                # log_prob is (1, 1, 1, vocab_size)
+                log_prob = log_prob.squeeze()
+                # Now log_prob is (vocab_size,)
+                joint_cache[cached_key] = log_prob
+            else:
+                log_prob = joint_cache[cached_key]
+
+            # First, process the blank symbol
+            skip_log_prob = log_prob[blank_id]
+            new_y_star_log_prob = y_star.log_prob + skip_log_prob
+
+            # ys[:] returns a copy of ys
+            B.add(Hypothesis(ys=y_star.ys[:], log_prob=new_y_star_log_prob))
+
+            # Second, process other non-blank labels
+            values, indices = log_prob.topk(beam + 1)
+            for i, v in zip(indices.tolist(), values.tolist()):
+                if i == blank_id:
+                    continue
+                new_ys = y_star.ys + [i]
+                new_log_prob = y_star.log_prob + v
+                A.add(Hypothesis(ys=new_ys, log_prob=new_log_prob))
+
+            # Check whether B contains more than "beam" elements more probable
+            # than the most probable in A
+            A_most_probable = A.get_most_probable()
+
+            kept_B = B.filter(A_most_probable.log_prob)
+
+            if len(kept_B) >= beam:
+                B = kept_B.topk(beam)
+                break
+
+        t += 1
+
+    best_hyp = B.get_most_probable(length_norm=True)
+    ys = best_hyp.ys[context_size:]  # [context_size:] to remove blanks
+    return ys

egs/wenetspeech/ASR/pruned_transducer_stateless2/decode.py

@@ -0,0 +1,617 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+When training with the L subset, usage:
+(1) greedy search
+./pruned_transducer_stateless2/decode.py \
+        --epoch 6 \
+        --avg 3 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir data/lang_char \
+        --max-duration 100 \
+        --decoding-method greedy_search
+
+(2) modified beam search
+./pruned_transducer_stateless2/decode.py \
+        --epoch 6 \
+        --avg 3 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir data/lang_char \
+        --max-duration 100 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+(3) fast beam search
+./pruned_transducer_stateless2/decode.py \
+        --epoch 6 \
+        --avg 3 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir data/lang_char \
+        --max-duration 1500 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+"""
+
+
+import argparse
+import logging
+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,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--batch",
+        type=int,
+        default=None,
+        help="It specifies the batch checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--avg-last-n",
+        type=int,
+        default=0,
+        help="""If positive, --epoch and --avg are ignored and it
+        will use the last n checkpoints exp_dir/checkpoint-xxx.pt
+        where xxx is the number of processed batches while
+        saving that checkpoint.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/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 interger 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""",
+    )
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    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 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 = model.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 = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search(
+            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 == "greedy_search"
+        and params.max_sym_per_frame == 1
+    ):
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+        )
+        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,
+            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 params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): 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 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 = 100
+    else:
+        log_interval = 2
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        texts = [list(str(text)) for text in texts]
+
+        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 hyp_words, ref_text in zip(hyps, texts):
+                this_batch.append((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"
+        )
+        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",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    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"-beam-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    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 params.avg_last_n > 0:
+        filenames = find_checkpoints(params.exp_dir)[: params.avg_last_n]
+        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)
+    elif params.batch is not None:
+        filenames = f"{params.exp_dir}/checkpoint-{params.batch}.pt"
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints([filenames], device=device))
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    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}")
+
+    # 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
+
+    wenetspeech = WenetSpeechAsrDataModule(args)
+
+    dev = "dev"
+    test_net = "test_net"
+    test_meet = "test_meet"
+
+    if not os.path.exists(f"{dev}/shared-0.tar"):
+        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"):
+        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_meet}/shared-0.tar"):
+        test_meeting_cuts = wenetspeech.test_meeting_cuts()
+        export_to_webdataset(
+            test_meeting_cuts,
+            output_path=f"{test_meet}/shared-%d.tar",
+            shard_size=300,
+        )
+
+    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_meet_shards = [
+        str(path)
+        for path in sorted(glob.glob(os.path.join(test_meet, "shared-*.tar")))
+    ]
+    cuts_test_meet_webdataset = CutSet.from_webdataset(
+        test_meet_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_meet_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,
+            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/pruned_transducer_stateless2/decoder.py

@@ -0,0 +1,103 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from scaling import ScaledConv1d, ScaledEmbedding
+
+
+class Decoder(nn.Module):
+    """This class modifies the stateless decoder from the following paper:
+
+        RNN-transducer with stateless prediction network
+        https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+
+    It removes the recurrent connection from the decoder, i.e., the prediction
+    network. Different from the above paper, it adds an extra Conv1d
+    right after the embedding layer.
+
+    TODO: Implement https://arxiv.org/pdf/2109.07513.pdf
+    """
+
+    def __init__(
+        self,
+        vocab_size: int,
+        decoder_dim: int,
+        blank_id: int,
+        context_size: int,
+    ):
+        """
+        Args:
+          vocab_size:
+            Number of tokens of the modeling unit including blank.
+          decoder_dim:
+            Dimension of the input embedding, and of the decoder output.
+          blank_id:
+            The ID of the blank symbol.
+          context_size:
+            Number of previous words to use to predict the next word.
+            1 means bigram; 2 means trigram. n means (n+1)-gram.
+        """
+        super().__init__()
+
+        self.embedding = ScaledEmbedding(
+            num_embeddings=vocab_size,
+            embedding_dim=decoder_dim,
+            padding_idx=blank_id,
+        )
+        self.blank_id = blank_id
+
+        assert context_size >= 1, context_size
+        self.context_size = context_size
+        self.vocab_size = vocab_size
+        if context_size > 1:
+            self.conv = ScaledConv1d(
+                in_channels=decoder_dim,
+                out_channels=decoder_dim,
+                kernel_size=context_size,
+                padding=0,
+                groups=decoder_dim,
+                bias=False,
+            )
+
+    def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, U).
+          need_pad:
+            True to left pad the input. Should be True during training.
+            False to not pad the input. Should be False during inference.
+        Returns:
+          Return a tensor of shape (N, U, decoder_dim).
+        """
+        y = y.to(torch.int64)
+        embedding_out = self.embedding(y)
+        if self.context_size > 1:
+            embedding_out = embedding_out.permute(0, 2, 1)
+            if need_pad is True:
+                embedding_out = F.pad(
+                    embedding_out, pad=(self.context_size - 1, 0)
+                )
+            else:
+                # During inference time, there is no need to do extra padding
+                # as we only need one output
+                assert embedding_out.size(-1) == self.context_size
+            embedding_out = self.conv(embedding_out)
+            embedding_out = embedding_out.permute(0, 2, 1)
+        embedding_out = F.relu(embedding_out)
+        return embedding_out

egs/wenetspeech/ASR/pruned_transducer_stateless2/encoder_interface.py

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

egs/wenetspeech/ASR/pruned_transducer_stateless2/export.py

@@ -0,0 +1,179 @@
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#           2022 Xiaomi Corporation (Author: Mingshuang Luo)    
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./pruned_transducer_stateless2/export.py \
+  --exp-dir ./pruned_transducer_stateless2/exp \
+  --lang-dir data/lang_char \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `pruned_transducer_stateless2/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/wenetspeech/ASR
+    ./pruned_transducer_stateless2/decode.py \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100 \
+        --lang-dir data/lang_char
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.lexicon import Lexicon
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/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",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    assert args.jit is False, "Support torchscript will be added later"
+
+    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)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = (
+        "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    )
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()

egs/wenetspeech/ASR/pruned_transducer_stateless2/joiner.py

@@ -0,0 +1,67 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+from scaling import ScaledLinear
+
+
+class Joiner(nn.Module):
+    def __init__(
+        self,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+    ):
+        super().__init__()
+
+        self.encoder_proj = ScaledLinear(encoder_dim, joiner_dim)
+        self.decoder_proj = ScaledLinear(decoder_dim, joiner_dim)
+        self.output_linear = ScaledLinear(joiner_dim, vocab_size)
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+        project_input: bool = True,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            Output from the encoder. Its shape is (N, T, s_range, C).
+          decoder_out:
+            Output from the decoder. Its shape is (N, T, s_range, C).
+           project_input:
+            If true, apply input projections encoder_proj and decoder_proj.
+            If this is false, it is the user's responsibility to do this
+            manually.
+        Returns:
+          Return a tensor of shape (N, T, s_range, C).
+        """
+        assert encoder_out.ndim == decoder_out.ndim == 4
+        assert encoder_out.shape[:-1] == decoder_out.shape[:-1]
+
+        if project_input:
+            logit = self.encoder_proj(encoder_out) + self.decoder_proj(
+                decoder_out
+            )
+        else:
+            logit = encoder_out + decoder_out
+
+        logit = self.output_linear(torch.tanh(logit))
+
+        return logit

egs/wenetspeech/ASR/pruned_transducer_stateless2/model.py

@@ -0,0 +1,193 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from scaling import ScaledLinear
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dm) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, decoder_dim).
+            It should contain one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, encoder_dim) and (N, U, decoder_dim).
+            Its output shape is (N, T, U, vocab_size). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+        self.simple_am_proj = ScaledLinear(
+            encoder_dim, vocab_size, initial_speed=0.5
+        )
+        self.simple_lm_proj = ScaledLinear(decoder_dim, vocab_size)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+        warmup: float = 1.0,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+          warmup:
+            A value warmup >= 0 that determines which modules are active, values
+            warmup > 1 "are fully warmed up" and all modules will be active.
+        Returns:
+          Return the transducer loss.
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, x_lens = self.encoder(x, x_lens, warmup=warmup)
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, decoder_dim]
+        decoder_out = self.decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros(
+            (x.size(0), 4), dtype=torch.int64, device=x.device
+        )
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = x_lens
+
+        lm = self.simple_lm_proj(decoder_out)
+        am = self.simple_am_proj(encoder_out)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+                lm=lm.float(),
+                am=am.float(),
+                symbols=y_padded,
+                termination_symbol=blank_id,
+                lm_only_scale=lm_scale,
+                am_only_scale=am_scale,
+                boundary=boundary,
+                reduction="sum",
+                return_grad=True,
+            )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, encoder_dim]
+        # lm_pruned : [B, T, prune_range, decoder_dim]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=self.joiner.encoder_proj(encoder_out),
+            lm=self.joiner.decoder_proj(decoder_out),
+            ranges=ranges,
+        )
+
+        # logits : [B, T, prune_range, vocab_size]
+
+        # project_input=False since we applied the decoder's input projections
+        # prior to do_rnnt_pruning (this is an optimization for speed).
+        logits = self.joiner(am_pruned, lm_pruned, project_input=False)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            pruned_loss = k2.rnnt_loss_pruned(
+                logits=logits.float(),
+                symbols=y_padded,
+                ranges=ranges,
+                termination_symbol=blank_id,
+                boundary=boundary,
+                reduction="sum",
+            )
+
+        return (simple_loss, pruned_loss)

egs/wenetspeech/ASR/pruned_transducer_stateless2/optim.py

@@ -0,0 +1,331 @@
+# Copyright      2022  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import List, Optional, Union
+
+import torch
+from torch.optim import Optimizer
+
+
+class Eve(Optimizer):
+    r"""
+    Implements Eve algorithm.  This is a modified version of AdamW with a special
+    way of setting the weight-decay / shrinkage-factor, which is designed to make the
+    rms of the parameters approach a particular target_rms (default: 0.1).  This is
+    for use with networks with 'scaled' versions of modules (see scaling.py), which
+    will be close to invariant to the absolute scale on the parameter matrix.
+
+    The original Adam algorithm was proposed in `Adam: A Method for Stochastic Optimization`_.
+    The AdamW variant was proposed in `Decoupled Weight Decay Regularization`_.
+    Eve is unpublished so far.
+
+    Arguments:
+        params (iterable): iterable of parameters to optimize or dicts defining
+            parameter groups
+        lr (float, optional): learning rate (default: 1e-3)
+        betas (Tuple[float, float], optional): coefficients used for computing
+            running averages of gradient and its square (default: (0.9, 0.999))
+        eps (float, optional): term added to the denominator to improve
+            numerical stability (default: 1e-8)
+        weight_decay (float, optional): weight decay coefficient (default: 3e-4;
+            this value means that the weight would decay significantly after
+            about 3k minibatches.  Is not multiplied by learning rate, but
+            is conditional on RMS-value of parameter being > target_rms.
+        target_rms (float, optional): target root-mean-square value of
+           parameters, if they fall below this we will stop applying weight decay.
+
+
+    .. _Adam\: A Method for Stochastic Optimization:
+        https://arxiv.org/abs/1412.6980
+    .. _Decoupled Weight Decay Regularization:
+        https://arxiv.org/abs/1711.05101
+    .. _On the Convergence of Adam and Beyond:
+        https://openreview.net/forum?id=ryQu7f-RZ
+    """
+
+    def __init__(
+        self,
+        params,
+        lr=1e-3,
+        betas=(0.9, 0.98),
+        eps=1e-8,
+        weight_decay=1e-3,
+        target_rms=0.1,
+    ):
+
+        if not 0.0 <= lr:
+            raise ValueError("Invalid learning rate: {}".format(lr))
+        if not 0.0 <= eps:
+            raise ValueError("Invalid epsilon value: {}".format(eps))
+        if not 0.0 <= betas[0] < 1.0:
+            raise ValueError(
+                "Invalid beta parameter at index 0: {}".format(betas[0])
+            )
+        if not 0.0 <= betas[1] < 1.0:
+            raise ValueError(
+                "Invalid beta parameter at index 1: {}".format(betas[1])
+            )
+        if not 0 <= weight_decay <= 0.1:
+            raise ValueError(
+                "Invalid weight_decay value: {}".format(weight_decay)
+            )
+        if not 0 < target_rms <= 10.0:
+            raise ValueError("Invalid target_rms value: {}".format(target_rms))
+        defaults = dict(
+            lr=lr,
+            betas=betas,
+            eps=eps,
+            weight_decay=weight_decay,
+            target_rms=target_rms,
+        )
+        super(Eve, self).__init__(params, defaults)
+
+    def __setstate__(self, state):
+        super(Eve, self).__setstate__(state)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """Performs a single optimization step.
+
+        Arguments:
+            closure (callable, optional): A closure that reevaluates the model
+                and returns the loss.
+        """
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        for group in self.param_groups:
+            for p in group["params"]:
+                if p.grad is None:
+                    continue
+
+                # Perform optimization step
+                grad = p.grad
+                if grad.is_sparse:
+                    raise RuntimeError(
+                        "AdamW does not support sparse gradients"
+                    )
+
+                state = self.state[p]
+
+                # State initialization
+                if len(state) == 0:
+                    state["step"] = 0
+                    # Exponential moving average of gradient values
+                    state["exp_avg"] = torch.zeros_like(
+                        p, memory_format=torch.preserve_format
+                    )
+                    # Exponential moving average of squared gradient values
+                    state["exp_avg_sq"] = torch.zeros_like(
+                        p, memory_format=torch.preserve_format
+                    )
+
+                exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"]
+
+                beta1, beta2 = group["betas"]
+
+                state["step"] += 1
+                bias_correction1 = 1 - beta1 ** state["step"]
+                bias_correction2 = 1 - beta2 ** state["step"]
+
+                # Decay the first and second moment running average coefficient
+                exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
+                exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
+                denom = (exp_avg_sq.sqrt() * (bias_correction2 ** -0.5)).add_(
+                    group["eps"]
+                )
+
+                step_size = group["lr"] / bias_correction1
+                target_rms = group["target_rms"]
+                weight_decay = group["weight_decay"]
+
+                if p.numel() > 1:
+                    # avoid applying this weight-decay on "scaling factors"
+                    # (which are scalar).
+                    is_above_target_rms = p.norm() > (
+                        target_rms * (p.numel() ** 0.5)
+                    )
+                    p.mul_(1 - (weight_decay * is_above_target_rms))
+                p.addcdiv_(exp_avg, denom, value=-step_size)
+
+        return loss
+
+
+class LRScheduler(object):
+    """
+    Base-class for learning rate schedulers where the learning-rate depends on both the
+    batch and the epoch.
+    """
+
+    def __init__(self, optimizer: Optimizer, verbose: bool = False):
+        # Attach optimizer
+        if not isinstance(optimizer, Optimizer):
+            raise TypeError(
+                "{} is not an Optimizer".format(type(optimizer).__name__)
+            )
+        self.optimizer = optimizer
+        self.verbose = verbose
+
+        for group in optimizer.param_groups:
+            group.setdefault("initial_lr", group["lr"])
+
+        self.base_lrs = [
+            group["initial_lr"] for group in optimizer.param_groups
+        ]
+
+        self.epoch = 0
+        self.batch = 0
+
+    def state_dict(self):
+        """Returns the state of the scheduler as a :class:`dict`.
+
+        It contains an entry for every variable in self.__dict__ which
+        is not the optimizer.
+        """
+        return {
+            "base_lrs": self.base_lrs,
+            "epoch": self.epoch,
+            "batch": self.batch,
+        }
+
+    def load_state_dict(self, state_dict):
+        """Loads the schedulers state.
+
+        Args:
+            state_dict (dict): scheduler state. Should be an object returned
+                from a call to :meth:`state_dict`.
+        """
+        self.__dict__.update(state_dict)
+
+    def get_last_lr(self) -> List[float]:
+        """Return last computed learning rate by current scheduler.  Will be a list of float."""
+        return self._last_lr
+
+    def get_lr(self):
+        # Compute list of learning rates from self.epoch and self.batch and
+        # self.base_lrs; this must be overloaded by the user.
+        # e.g. return [some_formula(self.batch, self.epoch, base_lr) for base_lr in self.base_lrs ]
+        raise NotImplementedError
+
+    def step_batch(self, batch: Optional[int] = None) -> None:
+        # Step the batch index, or just set it.  If `batch` is specified, it
+        # must be the batch index from the start of training, i.e. summed over
+        # all epochs.
+        # You can call this in any order; if you don't provide 'batch', it should
+        # of course be called once per batch.
+        if batch is not None:
+            self.batch = batch
+        else:
+            self.batch = self.batch + 1
+        self._set_lrs()
+
+    def step_epoch(self, epoch: Optional[int] = None):
+        # Step the epoch index, or just set it.  If you provide the 'epoch' arg,
+        # you should call this at the start of the epoch; if you don't provide the 'epoch'
+        # arg, you should call it at the end of the epoch.
+        if epoch is not None:
+            self.epoch = epoch
+        else:
+            self.epoch = self.epoch + 1
+        self._set_lrs()
+
+    def _set_lrs(self):
+        values = self.get_lr()
+        assert len(values) == len(self.optimizer.param_groups)
+
+        for i, data in enumerate(zip(self.optimizer.param_groups, values)):
+            param_group, lr = data
+            param_group["lr"] = lr
+            self.print_lr(self.verbose, i, lr)
+        self._last_lr = [group["lr"] for group in self.optimizer.param_groups]
+
+    def print_lr(self, is_verbose, group, lr):
+        """Display the current learning rate."""
+        if is_verbose:
+            print(
+                f"Epoch={self.epoch}, batch={self.batch}: adjusting learning rate"
+                f" of group {group} to {lr:.4e}."
+            )
+
+
+class Eden(LRScheduler):
+    """
+    Eden scheduler.
+     lr = initial_lr * (((batch**2 + lr_batches**2) / lr_batches**2) ** -0.25 *
+                       (((epoch**2 + lr_epochs**2) / lr_epochs**2) ** -0.25))
+
+     E.g. suggest initial-lr = 0.003 (passed to optimizer).
+
+    Args:
+        optimizer: the optimizer to change the learning rates on
+        lr_batches: the number of batches after which we start significantly
+              decreasing the learning rate, suggest 5000.
+        lr_epochs: the number of epochs after which we start significantly
+              decreasing the learning rate, suggest 6 if you plan to do e.g.
+              20 to 40 epochs, but may need smaller number if dataset is huge
+              and you will do few epochs.
+    """
+
+    def __init__(
+        self,
+        optimizer: Optimizer,
+        lr_batches: Union[int, float],
+        lr_epochs: Union[int, float],
+        verbose: bool = False,
+    ):
+        super(Eden, self).__init__(optimizer, verbose)
+        self.lr_batches = lr_batches
+        self.lr_epochs = lr_epochs
+
+    def get_lr(self):
+        factor = (
+            (self.batch ** 2 + self.lr_batches ** 2) / self.lr_batches ** 2
+        ) ** -0.25 * (
+            ((self.epoch ** 2 + self.lr_epochs ** 2) / self.lr_epochs ** 2)
+            ** -0.25
+        )
+        return [x * factor for x in self.base_lrs]
+
+
+def _test_eden():
+    m = torch.nn.Linear(100, 100)
+    optim = Eve(m.parameters(), lr=0.003)
+
+    scheduler = Eden(optim, lr_batches=30, lr_epochs=2, verbose=True)
+
+    for epoch in range(10):
+        scheduler.step_epoch(epoch)  # sets epoch to `epoch`
+
+        for step in range(20):
+            x = torch.randn(200, 100).detach()
+            x.requires_grad = True
+            y = m(x)
+            dy = torch.randn(200, 100).detach()
+            f = (y * dy).sum()
+            f.backward()
+
+            optim.step()
+            scheduler.step_batch()
+            optim.zero_grad()
+    print("last lr = ", scheduler.get_last_lr())
+    print("state dict = ", scheduler.state_dict())
+
+
+if __name__ == "__main__":
+    _test_eden()

egs/wenetspeech/ASR/pruned_transducer_stateless2/scaling.py

@@ -0,0 +1,702 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import collections
+from itertools import repeat
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+from torch import Tensor
+
+
+def _ntuple(n):
+    def parse(x):
+        if isinstance(x, collections.Iterable):
+            return x
+        return tuple(repeat(x, n))
+
+    return parse
+
+
+_single = _ntuple(1)
+_pair = _ntuple(2)
+
+
+class ActivationBalancerFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        channel_dim: int,
+        min_positive: float,  # e.g. 0.05
+        max_positive: float,  # e.g. 0.95
+        max_factor: float,  # e.g. 0.01
+        min_abs: float,  # e.g. 0.2
+        max_abs: float,  # e.g. 100.0
+    ) -> Tensor:
+        if x.requires_grad:
+            if channel_dim < 0:
+                channel_dim += x.ndim
+            sum_dims = [d for d in range(x.ndim) if d != channel_dim]
+            xgt0 = x > 0
+            proportion_positive = torch.mean(
+                xgt0.to(x.dtype), dim=sum_dims, keepdim=True
+            )
+            factor1 = (
+                (min_positive - proportion_positive).relu()
+                * (max_factor / min_positive)
+                if min_positive != 0.0
+                else 0.0
+            )
+            factor2 = (
+                (proportion_positive - max_positive).relu()
+                * (max_factor / (max_positive - 1.0))
+                if max_positive != 1.0
+                else 0.0
+            )
+            factor = factor1 + factor2
+            if isinstance(factor, float):
+                factor = torch.zeros_like(proportion_positive)
+
+            mean_abs = torch.mean(x.abs(), dim=sum_dims, keepdim=True)
+            below_threshold = mean_abs < min_abs
+            above_threshold = mean_abs > max_abs
+
+            ctx.save_for_backward(
+                factor, xgt0, below_threshold, above_threshold
+            )
+            ctx.max_factor = max_factor
+            ctx.sum_dims = sum_dims
+        return x
+
+    @staticmethod
+    def backward(
+        ctx, x_grad: Tensor
+    ) -> Tuple[Tensor, None, None, None, None, None, None]:
+        factor, xgt0, below_threshold, above_threshold = ctx.saved_tensors
+        dtype = x_grad.dtype
+        scale_factor = (
+            (below_threshold.to(dtype) - above_threshold.to(dtype))
+            * (xgt0.to(dtype) - 0.5)
+            * (ctx.max_factor * 2.0)
+        )
+
+        neg_delta_grad = x_grad.abs() * (factor + scale_factor)
+        return x_grad - neg_delta_grad, None, None, None, None, None, None
+
+
+class BasicNorm(torch.nn.Module):
+    """
+    This is intended to be a simpler, and hopefully cheaper, replacement for
+    LayerNorm.  The observation this is based on, is that Transformer-type
+    networks, especially with pre-norm, sometimes seem to set one of the
+    feature dimensions to a large constant value (e.g. 50), which "defeats"
+    the LayerNorm because the output magnitude is then not strongly dependent
+    on the other (useful) features.  Presumably the weight and bias of the
+    LayerNorm are required to allow it to do this.
+
+    So the idea is to introduce this large constant value as an explicit
+    parameter, that takes the role of the "eps" in LayerNorm, so the network
+    doesn't have to do this trick.  We make the "eps" learnable.
+
+    Args:
+       num_channels: the number of channels, e.g. 512.
+      channel_dim: the axis/dimension corresponding to the channel,
+        interprted as an offset from the input's ndim if negative.
+        shis is NOT the num_channels; it should typically be one of
+        {-2, -1, 0, 1, 2, 3}.
+       eps: the initial "epsilon" that we add as ballast in:
+             scale = ((input_vec**2).mean() + epsilon)**-0.5
+          Note: our epsilon is actually large, but we keep the name
+          to indicate the connection with conventional LayerNorm.
+       learn_eps: if true, we learn epsilon; if false, we keep it
+         at the initial value.
+    """
+
+    def __init__(
+        self,
+        num_channels: int,
+        channel_dim: int = -1,  # CAUTION: see documentation.
+        eps: float = 0.25,
+        learn_eps: bool = True,
+    ) -> None:
+        super(BasicNorm, self).__init__()
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        if learn_eps:
+            self.eps = nn.Parameter(torch.tensor(eps).log().detach())
+        else:
+            self.register_buffer("eps", torch.tensor(eps).log().detach())
+
+    def forward(self, x: Tensor) -> Tensor:
+        assert x.shape[self.channel_dim] == self.num_channels
+        scales = (
+            torch.mean(x ** 2, dim=self.channel_dim, keepdim=True)
+            + self.eps.exp()
+        ) ** -0.5
+        return x * scales
+
+
+class ScaledLinear(nn.Linear):
+    """
+    A modified version of nn.Linear where the parameters are scaled before
+    use, via:
+         weight = self.weight * self.weight_scale.exp()
+         bias = self.bias * self.bias_scale.exp()
+
+    Args:
+        Accepts the standard args and kwargs that nn.Linear accepts
+        e.g. in_features, out_features, bias=False.
+
+        initial_scale: you can override this if you want to increase
+           or decrease the initial magnitude of the module's output
+           (affects the initialization of weight_scale and bias_scale).
+           Another option, if you want to do something like this, is
+           to re-initialize the parameters.
+        initial_speed: this affects how fast the parameter will
+           learn near the start of training; you can set it to a
+           value less than one if you suspect that a module
+           is contributing to instability near the start of training.
+           Nnote: regardless of the use of this option, it's best to
+           use schedulers like Noam that have a warm-up period.
+           Alternatively you can set it to more than 1 if you want it to
+           initially train faster.   Must be greater than 0.
+    """
+
+    def __init__(
+        self,
+        *args,
+        initial_scale: float = 1.0,
+        initial_speed: float = 1.0,
+        **kwargs
+    ):
+        super(ScaledLinear, self).__init__(*args, **kwargs)
+        initial_scale = torch.tensor(initial_scale).log()
+        self.weight_scale = nn.Parameter(initial_scale.clone().detach())
+        if self.bias is not None:
+            self.bias_scale = nn.Parameter(initial_scale.clone().detach())
+        else:
+            self.register_parameter("bias_scale", None)
+
+        self._reset_parameters(
+            initial_speed
+        )  # Overrides the reset_parameters in nn.Linear
+
+    def _reset_parameters(self, initial_speed: float):
+        std = 0.1 / initial_speed
+        a = (3 ** 0.5) * std
+        nn.init.uniform_(self.weight, -a, a)
+        if self.bias is not None:
+            nn.init.constant_(self.bias, 0.0)
+        fan_in = self.weight.shape[1] * self.weight[0][0].numel()
+        scale = fan_in ** -0.5  # 1/sqrt(fan_in)
+        with torch.no_grad():
+            self.weight_scale += torch.tensor(scale / std).log()
+
+    def get_weight(self):
+        return self.weight * self.weight_scale.exp()
+
+    def get_bias(self):
+        return None if self.bias is None else self.bias * self.bias_scale.exp()
+
+    def forward(self, input: Tensor) -> Tensor:
+        return torch.nn.functional.linear(
+            input, self.get_weight(), self.get_bias()
+        )
+
+
+class ScaledConv1d(nn.Conv1d):
+    # See docs for ScaledLinear
+    def __init__(
+        self,
+        *args,
+        initial_scale: float = 1.0,
+        initial_speed: float = 1.0,
+        **kwargs
+    ):
+        super(ScaledConv1d, self).__init__(*args, **kwargs)
+        initial_scale = torch.tensor(initial_scale).log()
+        self.weight_scale = nn.Parameter(initial_scale.clone().detach())
+        if self.bias is not None:
+            self.bias_scale = nn.Parameter(initial_scale.clone().detach())
+        else:
+            self.register_parameter("bias_scale", None)
+        self._reset_parameters(
+            initial_speed
+        )  # Overrides the reset_parameters in base class
+
+    def _reset_parameters(self, initial_speed: float):
+        std = 0.1 / initial_speed
+        a = (3 ** 0.5) * std
+        nn.init.uniform_(self.weight, -a, a)
+        if self.bias is not None:
+            nn.init.constant_(self.bias, 0.0)
+        fan_in = self.weight.shape[1] * self.weight[0][0].numel()
+        scale = fan_in ** -0.5  # 1/sqrt(fan_in)
+        with torch.no_grad():
+            self.weight_scale += torch.tensor(scale / std).log()
+
+    def get_weight(self):
+        return self.weight * self.weight_scale.exp()
+
+    def get_bias(self):
+        return None if self.bias is None else self.bias * self.bias_scale.exp()
+
+    def forward(self, input: Tensor) -> Tensor:
+        F = torch.nn.functional
+        if self.padding_mode != "zeros":
+            return F.conv1d(
+                F.pad(
+                    input,
+                    self._reversed_padding_repeated_twice,
+                    mode=self.padding_mode,
+                ),
+                self.get_weight(),
+                self.get_bias(),
+                self.stride,
+                _single(0),
+                self.dilation,
+                self.groups,
+            )
+        return F.conv1d(
+            input,
+            self.get_weight(),
+            self.get_bias(),
+            self.stride,
+            self.padding,
+            self.dilation,
+            self.groups,
+        )
+
+
+class ScaledConv2d(nn.Conv2d):
+    # See docs for ScaledLinear
+    def __init__(
+        self,
+        *args,
+        initial_scale: float = 1.0,
+        initial_speed: float = 1.0,
+        **kwargs
+    ):
+        super(ScaledConv2d, self).__init__(*args, **kwargs)
+        initial_scale = torch.tensor(initial_scale).log()
+        self.weight_scale = nn.Parameter(initial_scale.clone().detach())
+        if self.bias is not None:
+            self.bias_scale = nn.Parameter(initial_scale.clone().detach())
+        else:
+            self.register_parameter("bias_scale", None)
+        self._reset_parameters(
+            initial_speed
+        )  # Overrides the reset_parameters in base class
+
+    def _reset_parameters(self, initial_speed: float):
+        std = 0.1 / initial_speed
+        a = (3 ** 0.5) * std
+        nn.init.uniform_(self.weight, -a, a)
+        if self.bias is not None:
+            nn.init.constant_(self.bias, 0.0)
+        fan_in = self.weight.shape[1] * self.weight[0][0].numel()
+        scale = fan_in ** -0.5  # 1/sqrt(fan_in)
+        with torch.no_grad():
+            self.weight_scale += torch.tensor(scale / std).log()
+
+    def get_weight(self):
+        return self.weight * self.weight_scale.exp()
+
+    def get_bias(self):
+        return None if self.bias is None else self.bias * self.bias_scale.exp()
+
+    def _conv_forward(self, input, weight):
+        F = torch.nn.functional
+        if self.padding_mode != "zeros":
+            return F.conv2d(
+                F.pad(
+                    input,
+                    self._reversed_padding_repeated_twice,
+                    mode=self.padding_mode,
+                ),
+                weight,
+                self.get_bias(),
+                self.stride,
+                _pair(0),
+                self.dilation,
+                self.groups,
+            )
+        return F.conv2d(
+            input,
+            weight,
+            self.get_bias(),
+            self.stride,
+            self.padding,
+            self.dilation,
+            self.groups,
+        )
+
+    def forward(self, input: Tensor) -> Tensor:
+        return self._conv_forward(input, self.get_weight())
+
+
+class ActivationBalancer(torch.nn.Module):
+    """
+    Modifies the backpropped derivatives of a function to try to encourage, for
+    each channel, that it is positive at least a proportion `threshold` of the
+    time.  It does this by multiplying negative derivative values by up to
+    (1+max_factor), and positive derivative values by up to (1-max_factor),
+    interpolated from 1 at the threshold to those extremal values when none
+    of the inputs are positive.
+
+
+    Args:
+           channel_dim: the dimension/axis corresponding to the channel, e.g.
+               -1, 0, 1, 2; will be interpreted as an offset from x.ndim if negative.
+           min_positive: the minimum, per channel, of the proportion of the time
+               that (x > 0), below which we start to modify the derivatives.
+           max_positive: the maximum, per channel, of the proportion of the time
+               that (x > 0), above which we start to modify the derivatives.
+           max_factor: the maximum factor by which we modify the derivatives for
+              either the sign constraint or the magnitude constraint;
+              e.g. with max_factor=0.02, the the derivatives would be multiplied by
+              values in the range [0.98..1.02].
+           min_abs:  the minimum average-absolute-value per channel, which
+              we allow, before we start to modify the derivatives to prevent
+              this.
+           max_abs:  the maximum average-absolute-value per channel, which
+               we allow, before we start to modify the derivatives to prevent
+               this.
+    """
+
+    def __init__(
+        self,
+        channel_dim: int,
+        min_positive: float = 0.05,
+        max_positive: float = 0.95,
+        max_factor: float = 0.01,
+        min_abs: float = 0.2,
+        max_abs: float = 100.0,
+    ):
+        super(ActivationBalancer, self).__init__()
+        self.channel_dim = channel_dim
+        self.min_positive = min_positive
+        self.max_positive = max_positive
+        self.max_factor = max_factor
+        self.min_abs = min_abs
+        self.max_abs = max_abs
+
+    def forward(self, x: Tensor) -> Tensor:
+        return ActivationBalancerFunction.apply(
+            x,
+            self.channel_dim,
+            self.min_positive,
+            self.max_positive,
+            self.max_factor,
+            self.min_abs,
+            self.max_abs,
+        )
+
+
+class DoubleSwishFunction(torch.autograd.Function):
+    """
+      double_swish(x) = x * torch.sigmoid(x-1)
+    This is a definition, originally motivated by its close numerical
+    similarity to swish(swish(x)), where swish(x) =  x * sigmoid(x).
+
+    Memory-efficient derivative computation:
+     double_swish(x) = x * s, where s(x) = torch.sigmoid(x-1)
+     double_swish'(x) = d/dx double_swish(x) =  x * s'(x) + x' * s(x) = x * s'(x) + s(x).
+     Now, s'(x) = s(x) * (1-s(x)).
+     double_swish'(x) =  x * s'(x) + s(x).
+                      =  x * s(x) * (1-s(x)) + s(x).
+                     = double_swish(x) * (1-s(x)) + s(x)
+     ... so we just need to remember s(x) but not x itself.
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor) -> Tensor:
+        x = x.detach()
+        s = torch.sigmoid(x - 1.0)
+        y = x * s
+        ctx.save_for_backward(s, y)
+        return y
+
+    @staticmethod
+    def backward(ctx, y_grad: Tensor) -> Tensor:
+        s, y = ctx.saved_tensors
+        return (y * (1 - s) + s) * y_grad
+
+
+class DoubleSwish(torch.nn.Module):
+    def forward(self, x: Tensor) -> Tensor:
+        """Return double-swish activation function which is an approximation to Swish(Swish(x)),
+        that we approximate closely with x * sigmoid(x-1).
+        """
+        return DoubleSwishFunction.apply(x)
+
+
+class ScaledEmbedding(nn.Module):
+    r"""This is a modified version of nn.Embedding that introduces a learnable scale
+    on the parameters.  Note: due to how we initialize it, it's best used with
+    schedulers like Noam that have a warmup period.
+
+    It is a simple lookup table that stores embeddings of a fixed dictionary and size.
+
+    This module is often used to store word embeddings and retrieve them using indices.
+    The input to the module is a list of indices, and the output is the corresponding
+    word embeddings.
+
+    Args:
+        num_embeddings (int): size of the dictionary of embeddings
+        embedding_dim (int): the size of each embedding vector
+        padding_idx (int, optional): If given, pads the output with the embedding vector at :attr:`padding_idx`
+                                         (initialized to zeros) whenever it encounters the index.
+        max_norm (float, optional): If given, each embedding vector with norm larger than :attr:`max_norm`
+                                    is renormalized to have norm :attr:`max_norm`.
+        norm_type (float, optional): The p of the p-norm to compute for the :attr:`max_norm` option. Default ``2``.
+        scale_grad_by_freq (boolean, optional): If given, this will scale gradients by the inverse of frequency of
+                                                the words in the mini-batch. Default ``False``.
+        sparse (bool, optional): If ``True``, gradient w.r.t. :attr:`weight` matrix will be a sparse tensor.
+                                 See Notes for more details regarding sparse gradients.
+
+        initial_speed (float, optional):  This affects how fast the parameter will
+           learn near the start of training; you can set it to a value less than
+           one if you suspect that a module is contributing to instability near
+           the start of training.  Nnote: regardless of the use of this option,
+           it's best to use schedulers like Noam that have a warm-up period.
+           Alternatively you can set it to more than 1 if you want it to
+           initially train faster.  Must be greater than 0.
+
+
+    Attributes:
+        weight (Tensor): the learnable weights of the module of shape (num_embeddings, embedding_dim)
+                         initialized from :math:`\mathcal{N}(0, 1)`
+
+    Shape:
+        - Input: :math:`(*)`, LongTensor of arbitrary shape containing the indices to extract
+        - Output: :math:`(*, H)`, where `*` is the input shape and :math:`H=\text{embedding\_dim}`
+
+    .. note::
+        Keep in mind that only a limited number of optimizers support
+        sparse gradients: currently it's :class:`optim.SGD` (`CUDA` and `CPU`),
+        :class:`optim.SparseAdam` (`CUDA` and `CPU`) and :class:`optim.Adagrad` (`CPU`)
+
+    .. note::
+        With :attr:`padding_idx` set, the embedding vector at
+        :attr:`padding_idx` is initialized to all zeros. However, note that this
+        vector can be modified afterwards, e.g., using a customized
+        initialization method, and thus changing the vector used to pad the
+        output. The gradient for this vector from :class:`~torch.nn.Embedding`
+        is always zero.
+
+    Examples::
+
+        >>> # an Embedding module containing 10 tensors of size 3
+        >>> embedding = nn.Embedding(10, 3)
+        >>> # a batch of 2 samples of 4 indices each
+        >>> input = torch.LongTensor([[1,2,4,5],[4,3,2,9]])
+        >>> embedding(input)
+        tensor([[[-0.0251, -1.6902,  0.7172],
+                 [-0.6431,  0.0748,  0.6969],
+                 [ 1.4970,  1.3448, -0.9685],
+                 [-0.3677, -2.7265, -0.1685]],
+
+                [[ 1.4970,  1.3448, -0.9685],
+                 [ 0.4362, -0.4004,  0.9400],
+                 [-0.6431,  0.0748,  0.6969],
+                 [ 0.9124, -2.3616,  1.1151]]])
+
+
+        >>> # example with padding_idx
+        >>> embedding = nn.Embedding(10, 3, padding_idx=0)
+        >>> input = torch.LongTensor([[0,2,0,5]])
+        >>> embedding(input)
+        tensor([[[ 0.0000,  0.0000,  0.0000],
+                 [ 0.1535, -2.0309,  0.9315],
+                 [ 0.0000,  0.0000,  0.0000],
+                 [-0.1655,  0.9897,  0.0635]]])
+
+    """
+    __constants__ = [
+        "num_embeddings",
+        "embedding_dim",
+        "padding_idx",
+        "scale_grad_by_freq",
+        "sparse",
+    ]
+
+    num_embeddings: int
+    embedding_dim: int
+    padding_idx: int
+    scale_grad_by_freq: bool
+    weight: Tensor
+    sparse: bool
+
+    def __init__(
+        self,
+        num_embeddings: int,
+        embedding_dim: int,
+        padding_idx: Optional[int] = None,
+        scale_grad_by_freq: bool = False,
+        sparse: bool = False,
+        initial_speed: float = 1.0,
+    ) -> None:
+        super(ScaledEmbedding, self).__init__()
+        self.num_embeddings = num_embeddings
+        self.embedding_dim = embedding_dim
+        if padding_idx is not None:
+            if padding_idx > 0:
+                assert (
+                    padding_idx < self.num_embeddings
+                ), "Padding_idx must be within num_embeddings"
+            elif padding_idx < 0:
+                assert (
+                    padding_idx >= -self.num_embeddings
+                ), "Padding_idx must be within num_embeddings"
+                padding_idx = self.num_embeddings + padding_idx
+        self.padding_idx = padding_idx
+        self.scale_grad_by_freq = scale_grad_by_freq
+
+        self.scale = nn.Parameter(torch.zeros(()))  # see reset_parameters()
+        self.sparse = sparse
+
+        self.weight = nn.Parameter(torch.Tensor(num_embeddings, embedding_dim))
+        self.reset_parameters(initial_speed)
+
+    def reset_parameters(self, initial_speed: float = 1.0) -> None:
+        std = 0.1 / initial_speed
+        nn.init.normal_(self.weight, std=std)
+        nn.init.constant_(self.scale, torch.tensor(1.0 / std).log())
+
+        if self.padding_idx is not None:
+            with torch.no_grad():
+                self.weight[self.padding_idx].fill_(0)
+
+    def forward(self, input: Tensor) -> Tensor:
+        F = torch.nn.functional
+        scale = self.scale.exp()
+        if input.numel() < self.num_embeddings:
+            return (
+                F.embedding(
+                    input,
+                    self.weight,
+                    self.padding_idx,
+                    None,
+                    2.0,  # None, 2.0 relate to normalization
+                    self.scale_grad_by_freq,
+                    self.sparse,
+                )
+                * scale
+            )
+        else:
+            return F.embedding(
+                input,
+                self.weight * scale,
+                self.padding_idx,
+                None,
+                2.0,  # None, 2.0 relates to normalization
+                self.scale_grad_by_freq,
+                self.sparse,
+            )
+
+    def extra_repr(self) -> str:
+        s = "{num_embeddings}, {embedding_dim}, scale={scale}"
+        if self.padding_idx is not None:
+            s += ", padding_idx={padding_idx}"
+        if self.scale_grad_by_freq is not False:
+            s += ", scale_grad_by_freq={scale_grad_by_freq}"
+        if self.sparse is not False:
+            s += ", sparse=True"
+        return s.format(**self.__dict__)
+
+
+def _test_activation_balancer_sign():
+    probs = torch.arange(0, 1, 0.01)
+    N = 1000
+    x = 1.0 * (torch.rand(probs.numel(), N) < probs.unsqueeze(-1))
+    x = x.detach()
+    x.requires_grad = True
+    m = ActivationBalancer(
+        channel_dim=0,
+        min_positive=0.05,
+        max_positive=0.95,
+        max_factor=0.2,
+        min_abs=0.0,
+    )
+
+    y_grad = torch.sign(torch.randn(probs.numel(), N))
+
+    y = m(x)
+    y.backward(gradient=y_grad)
+    print("_test_activation_balancer_sign: x = ", x)
+    print("_test_activation_balancer_sign: y grad = ", y_grad)
+    print("_test_activation_balancer_sign: x grad = ", x.grad)
+
+
+def _test_activation_balancer_magnitude():
+    magnitudes = torch.arange(0, 1, 0.01)
+    N = 1000
+    x = torch.sign(torch.randn(magnitudes.numel(), N)) * magnitudes.unsqueeze(
+        -1
+    )
+    x = x.detach()
+    x.requires_grad = True
+    m = ActivationBalancer(
+        channel_dim=0,
+        min_positive=0.0,
+        max_positive=1.0,
+        max_factor=0.2,
+        min_abs=0.2,
+        max_abs=0.8,
+    )
+
+    y_grad = torch.sign(torch.randn(magnitudes.numel(), N))
+
+    y = m(x)
+    y.backward(gradient=y_grad)
+    print("_test_activation_balancer_magnitude: x = ", x)
+    print("_test_activation_balancer_magnitude: y grad = ", y_grad)
+    print("_test_activation_balancer_magnitude: x grad = ", x.grad)
+
+
+def _test_basic_norm():
+    num_channels = 128
+    m = BasicNorm(num_channels=num_channels, channel_dim=1)
+
+    x = torch.randn(500, num_channels)
+
+    y = m(x)
+
+    assert y.shape == x.shape
+    x_rms = (x ** 2).mean().sqrt()
+    y_rms = (y ** 2).mean().sqrt()
+    print("x rms = ", x_rms)
+    print("y rms = ", y_rms)
+    assert y_rms < x_rms
+    assert y_rms > 0.5 * x_rms
+
+
+def _test_double_swish_deriv():
+    x = torch.randn(10, 12, dtype=torch.double) * 0.5
+    x.requires_grad = True
+    m = DoubleSwish()
+    torch.autograd.gradcheck(m, x)
+
+
+if __name__ == "__main__":
+    _test_activation_balancer_sign()
+    _test_activation_balancer_magnitude()
+    _test_basic_norm()
+    _test_double_swish_deriv()

egs/wenetspeech/ASR/shared

@@ -0,0 +1 @@
+../../librispeech/ASR/shared