Merge 6d809bad0bfe6ea1beb36eabf7c71e2c5f7716e9 into f6ce135608968bfe736040345743dfbc2847cd8b

.flake8

@@ -21,4 +21,6 @@ exclude =
   .git,
   **/data/**,
   icefall/shared/make_kn_lm.py,
+  egs/librispeech/ASR/pruned_transducer_stateless5/conformer.py,
+  egs/librispeech/ASR/pruned_transducer_stateless5/sampling.py,
   icefall/__init__.py

egs/librispeech/ASR/README.md

@@ -19,6 +19,8 @@ The following table lists the differences among them.
 | `pruned_transducer_stateless`         | Conformer           | Embedding + Conv1d | Using k2 pruned RNN-T loss                        |
 | `pruned_transducer_stateless2`        | Conformer(modified) | Embedding + Conv1d | Using k2 pruned RNN-T loss                        |
 | `pruned_transducer_stateless3`        | Conformer(modified) | Embedding + Conv1d | Using k2 pruned RNN-T loss + using GigaSpeech as extra training data |
+| `pruned_transducer_stateless4`        | Conformer(modified) | Embedding + Conv1d | Same as pruned_transducer_stateless2 but supports saving averaged model periodically.|
+| `pruned_transducer_stateless5`        | Conformer(modified) | Embedding + Conv1d | Same as pruned_transducer_stateless3 but with knowledge bank|
 
 
 The decoder in `transducer_stateless` is modified from the paper

egs/librispeech/ASR/pruned_transducer_stateless4/train.py

@@ -411,7 +411,7 @@ def get_transducer_model(params: AttributeDict) -> nn.Module:
 def load_checkpoint_if_available(
     params: AttributeDict,
     model: nn.Module,
-    model_avg: nn.Module = None,
+    model_avg: Optional[nn.Module] = None,
     optimizer: Optional[torch.optim.Optimizer] = None,
     scheduler: Optional[LRSchedulerType] = None,
 ) -> Optional[Dict[str, Any]]:

egs/librispeech/ASR/pruned_transducer_stateless5/asr_datamodule.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless3/asr_datamodule.py

egs/librispeech/ASR/pruned_transducer_stateless5/beam_search.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless3/beam_search.py

egs/librispeech/ASR/pruned_transducer_stateless5/decode-giga.py

@@ -0,0 +1,681 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang
+#                                                 Zengwei Yao)
+#
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless5/decode-giga.py \
+        --epoch 28 \
+        --avg 15 \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --max-duration 100 \
+        --decoding-method greedy_search
+
+(2) beam search
+./pruned_transducer_stateless5/decode-giga.py \
+        --epoch 28 \
+        --avg 15 \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --max-duration 100 \
+        --decoding-method beam_search \
+        --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless5/decode-giga.py \
+        --epoch 28 \
+        --avg 15 \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --max-duration 100 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+(4) fast beam search
+./pruned_transducer_stateless5/decode-giga.py \
+        --epoch 28 \
+        --avg 15 \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --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 sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import AsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from gigaspeech import GigaSpeech
+from gigaspeech_scoring import asr_text_post_processing
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=False,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is
+        fast_beam_search or fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search or fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search or fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for computed nbest oracle WER
+        when the decoding method is fast_beam_search_nbest_oracle.
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding_method is fast_beam_search_nbest_oracle.
+        """,
+    )
+    return parser
+
+
+def post_processing(
+    results: List[Tuple[List[List[str]], List[List[str]]]],
+) -> List[Tuple[List[List[str]], List[List[str]]]]:
+    new_results = []
+    for ref, hyp in results:
+        new_ref = asr_text_post_processing(" ".join(ref)).split()
+        new_hyp = asr_text_post_processing(" ".join(hyp)).split()
+        new_results.append((new_ref, new_hyp))
+    return new_results
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    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.
+      sp:
+        The BPE 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 or fast_beam_search_nbest_oracle.
+    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_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    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 hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    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(sp.decode(hyp).split())
+
+    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
+        }
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}_"
+                f"num_paths_{params.num_paths}_"
+                f"nbest_scale_{params.nbest_scale}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    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.
+      sp:
+        The BPE 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"]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            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):
+                ref_words = ref_text.split()
+                this_batch.append((ref_words, 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[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = post_processing(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir
+        / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / "giga" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.decoding_method == "fast_beam_search":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        params.suffix += f"-num-paths-{params.num_paths}"
+        params.suffix += f"-nbest-scale-{params.nbest_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += (
+            f"-{params.decoding_method}-beam-size-{params.beam_size}"
+        )
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.unk_id()
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(
+                params.exp_dir, iteration=-params.iter
+            )[: params.avg]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        assert params.iter == 0 and params.avg > 0
+        start = params.epoch - params.avg
+        assert start >= 1
+        filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+        filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+        logging.info(
+            f"Calculating the averaged model over epoch range from "
+            f"{start} (excluded) to {params.epoch}"
+        )
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints_with_averaged_model(
+                filename_start=filename_start,
+                filename_end=filename_end,
+                device=device,
+            )
+        )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+    model.unk_id = params.unk_id
+
+    # In beam_search.py, we are using model.decoder() and model.joiner(),
+    # so we have to switch to the branch for the GigaSpeech dataset.
+    model.decoder = model.decoder_giga
+    model.joiner = model.joiner_giga
+
+    if params.decoding_method in (
+        "fast_beam_search",
+        "fast_beam_search_nbest_oracle",
+    ):
+        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}")
+
+    asr_datamodule = AsrDataModule(args)
+    gigaspeech = GigaSpeech(manifest_dir=args.manifest_dir)
+
+    test_cuts = gigaspeech.test_cuts()
+    dev_cuts = gigaspeech.dev_cuts()
+
+    test_dl = asr_datamodule.test_dataloaders(test_cuts)
+    dev_dl = asr_datamodule.test_dataloaders(dev_cuts)
+
+    test_sets = ["test", "dev"]
+    test_sets_dl = [test_dl, dev_dl]
+
+    for test_set, dl in zip(test_sets, test_sets_dl):
+        results_dict = decode_dataset(
+            dl=dl,
+            params=params,
+            model=model,
+            sp=sp,
+            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/librispeech/ASR/pruned_transducer_stateless5/decode.py

@@ -0,0 +1,663 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang
+#                                                 Zengwei Yao)
+#
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless5/decode.py \
+        --epoch 28 \
+        --avg 15 \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --max-duration 100 \
+        --decoding-method greedy_search
+
+(2) beam search
+./pruned_transducer_stateless5/decode.py \
+        --epoch 28 \
+        --avg 15 \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --max-duration 100 \
+        --decoding-method beam_search \
+        --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless5/decode.py \
+        --epoch 28 \
+        --avg 15 \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --max-duration 100 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+(4) fast beam search
+./pruned_transducer_stateless5/decode.py \
+        --epoch 28 \
+        --avg 15 \
+        --exp-dir ./pruned_transducer_stateless5/exp \
+        --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 sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import AsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from librispeech import LibriSpeech
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=False,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is
+        fast_beam_search or fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search or fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search or fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for computed nbest oracle WER
+        when the decoding method is fast_beam_search_nbest_oracle.
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding_method is fast_beam_search_nbest_oracle.
+        """,
+    )
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    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.
+      sp:
+        The BPE 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 or fast_beam_search_nbest_oracle.
+    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_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    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 hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    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(sp.decode(hyp).split())
+
+    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
+        }
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}_"
+                f"num_paths_{params.num_paths}_"
+                f"nbest_scale_{params.nbest_scale}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    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.
+      sp:
+        The BPE 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"]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            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):
+                ref_words = ref_text.split()
+                this_batch.append((ref_words, 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()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.decoding_method == "fast_beam_search":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        params.suffix += f"-num-paths-{params.num_paths}"
+        params.suffix += f"-nbest-scale-{params.nbest_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += (
+            f"-{params.decoding_method}-beam-size-{params.beam_size}"
+        )
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.unk_id()
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(
+                params.exp_dir, iteration=-params.iter
+            )[: params.avg]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        assert params.iter == 0 and params.avg > 0
+        start = params.epoch - params.avg
+        assert start >= 1
+        filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+        filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+        logging.info(
+            f"Calculating the averaged model over epoch range from "
+            f"{start} (excluded) to {params.epoch}"
+        )
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints_with_averaged_model(
+                filename_start=filename_start,
+                filename_end=filename_end,
+                device=device,
+            )
+        )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+    model.unk_id = params.unk_id
+
+    if params.decoding_method in (
+        "fast_beam_search",
+        "fast_beam_search_nbest_oracle",
+    ):
+        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}")
+
+    asr_datamodule = AsrDataModule(args)
+    librispeech = LibriSpeech(manifest_dir=args.manifest_dir)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = asr_datamodule.test_dataloaders(test_clean_cuts)
+    test_other_dl = asr_datamodule.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            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/librispeech/ASR/pruned_transducer_stateless5/decoder.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless3/decoder.py

egs/librispeech/ASR/pruned_transducer_stateless5/encoder_interface.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless3/encoder_interface.py

egs/librispeech/ASR/pruned_transducer_stateless5/gigaspeech.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless3/gigaspeech.py

egs/librispeech/ASR/pruned_transducer_stateless5/gigaspeech_scoring.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless3/gigaspeech_scoring.py

egs/librispeech/ASR/pruned_transducer_stateless5/joiner.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless3/joiner.py

egs/librispeech/ASR/pruned_transducer_stateless5/librispeech.py

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

egs/librispeech/ASR/pruned_transducer_stateless5/model.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless3/model.py

egs/librispeech/ASR/pruned_transducer_stateless5/optim.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless3/optim.py

egs/librispeech/ASR/pruned_transducer_stateless5/sampling.py

@@ -0,0 +1,293 @@
+#!/usr/bin/env python3
+
+# This was copied from /ceph-dan/torch-sampling/torch_sampling/sampling_ref.py,
+# its git history is there.
+
+import timeit
+import torch
+from torch import Tensor
+from torch import nn
+from torch.cuda.amp import GradScaler, custom_fwd, custom_bwd
+from typing import Tuple, Optional
+from scaling import ScaledLinear
+import random
+from torch_scheduled_sampling import sample_combined
+
+# The main exports of this file are the module KnowledgeBaseLookup and the
+# function create_knowledge_base.
+
+
+
+
+
+
+def create_knowledge_base(M: int, N: int, D: int) -> nn.Parameter:
+    std = 0.1
+    a = (3 ** 0.5) * std  # this sqrt(3) thing is intended to get variance of
+                          # 0.1 from uniform distribution
+    ans = nn.Parameter(torch.ones(M ** N, D))
+    nn.init.uniform_(ans, -a, a)
+    return ans
+
+def join_indexes(indexes: Tensor, M: int) -> Tensor:
+    """
+    Combines N-tuples of indexes into single indexes that can be used for
+    lookup in the knowledge base.  Args:
+      indexes: tensor of torch.int64 of shape (*, K, N), with elements in
+         {0..M-1}
+         M: the size of the original softmaxes, is upper bound on elements
+           in indexes
+       Returns:
+          joined_indexes: of shape (*, K), joined_indexes[...,k] equals
+            joined_indexes[...,0,k] + joined_indexes[...,1,k]*(M**1) ... + joined_indexes[...,1,k]*(M**(N-1))]
+    """
+    N = indexes.shape[-1]
+    n_powers = M ** torch.arange(N, device=indexes.device)  # [ 1, M, ..., M**(N-1) ]
+    return (indexes * n_powers).sum(dim=-1)
+
+
+# Note, we don't use this, we
+def weighted_matrix_lookup(weights: Tensor,
+                           indexes: Tensor,
+                           knowledge_base: Tensor) -> Tensor:
+    """
+    Weighted combination of specified rows of a matrix.
+         weights: Tensor of shape (*, K), can contain any value but probably in [0..1].
+         indexes: Tensor of shape (*, K), with elements in [0..C-1]
+         knowledge_base: Tensor of shape (C-1, D), whose rows we'll be looking up
+      Returns:
+         tensor of shape (*, D), containing weighted sums of rows of
+         `knowledge_base`
+    """
+    if True:
+        return WeightedMatrixLookupFunction.apply(weights, indexes, knowledge_base)
+    else:
+        # simpler but less memory-efficient implementation
+        lookup = torch.index_select(knowledge_base, dim=0, index=indexes.flatten())
+        D = knowledge_base.shape[-1]
+        weights = weights.unsqueeze(-2)   # (*, 1, K)
+        lookup = lookup.reshape(*indexes.shape, D) # (*, K, D)
+        ans = torch.matmul(weights, lookup) # ans: (*, 1, D)
+        ans = ans.squeeze(-2)
+        assert list(ans.shape) == list(weights.shape[:-2]) + [D]
+        return ans
+
+
+class WeightedMatrixLookupFunction(torch.autograd.Function):
+    @staticmethod
+    @custom_fwd
+    def forward(ctx, weights: Tensor, indexes: Tensor, knowledge_base: Tensor) -> Tensor:
+        """
+        Weighted combination of specified rows of a matrix.
+         weights: Tensor of shape (*, K), can contain any value but probably in [0..1].
+         indexes: Tensor of shape (*, K), with elements in [0..C-1]
+         knowledge_base: Tensor of shape (C, D), whose rows we'll be looking up
+        Returns:
+         tensor of shape (*, D), containing weighted sums of rows of
+         `knowledge_base`
+        """
+        #  if random.random() < 0.001:
+        #      print("dtype[1] = ", weights.dtype)
+        ctx.save_for_backward(weights.detach(), indexes.detach(),
+                              knowledge_base.detach())
+        with torch.no_grad():
+            lookup = torch.index_select(knowledge_base, dim=0, index=indexes.flatten())
+            D = knowledge_base.shape[-1]
+            weights = weights.unsqueeze(-2)   # (*, 1, K)
+            lookup = lookup.reshape(*indexes.shape, D) # (*, K, D)
+            ans = torch.matmul(weights, lookup) # ans: (*, 1, D)
+            ans = ans.squeeze(-2) #(*, D)
+        return ans
+
+    @staticmethod
+    @custom_bwd
+    def backward(ctx, ans_grad: Tensor) -> Tuple[Tensor, None, Tensor]:
+        # ans_grad: (*, D)
+        weights, indexes, knowledge_base = ctx.saved_tensors
+        knowledge_base.requires_grad = True
+        dtype = ans_grad.dtype
+        ans_grad = ans_grad.to(weights.dtype)
+        assert weights.requires_grad == False
+        D = knowledge_base.shape[-1]
+        with torch.enable_grad():
+            # we'll use torch's autograd to differentiate this operation, which
+            # is nontrivial [and anyway we need `lookup` to compute weight grad.
+            # We don't save `lookup` because it's large, that is the reason
+            # we override Torch autograd.
+            lookup = torch.index_select(knowledge_base, dim=0, index=indexes.flatten())
+            lookup = lookup.reshape(*indexes.shape, D) # (*, K, D)
+        weights = weights.unsqueeze(-1)   # (*, K, 1)
+        # forward pass: was:
+        ## ans = torch.matmul(weights, lookup)
+        ## ans: (*, 1, D)
+        ## ans = ans.squeeze(-2) # ans, ans_grad: (*, D)
+        weights_grad = torch.matmul(lookup,  # (*, K, D)
+                                    ans_grad.unsqueeze(-1))  # (*, D, 1)
+        weights_grad = weights_grad.squeeze(-1) # (*, K, 1) -> (*, K)
+        lookup_grad = weights * ans_grad.unsqueeze(-2) # (*, K, 1) * (*, 1, D) = (*, K, D)
+        lookup.backward(gradient=lookup_grad)
+        return weights_grad.to(dtype), None, knowledge_base.grad.to(dtype)
+
+
+class KnowledgeBaseLookup(nn.Module):
+    """
+    Create knowledge-base lookup module.  (The knowledge-base parameter, which is
+    large, is shared between these modules).
+    Args:
+       M: int, softmax size, e.g. in [32..128]
+       N: int, number of softmaxes, in [2..3]
+       D: int, embedding dimension in knowledge base, e.g. 256
+       K: number of samples (affects speed/accuracy tradeoff), e.g. 16.
+      embedding_dim:  the dimension to project from and to, e.g. the
+        d_model of the conformer.
+    """
+    def __init__(self, M: int, N: int, D: int,
+                 K: int, embedding_dim: int,
+                 knowledge_base: nn.Parameter):
+        super(KnowledgeBaseLookup, self).__init__()
+        self.knowledge_base = knowledge_base  # shared!
+        self.in_proj = ScaledLinear(embedding_dim, M * N,
+                                    initial_scale=1.0)
+        # initial_scale = 4.0 because the knowlege_base activations are
+        # quite small -- if we use our optimizer they'll have stddev <= 0.1.
+        self.out_proj = ScaledLinear(D, embedding_dim,
+                                     initial_scale = 4.0)
+        self.M = M
+        self.N = N
+        self.K = K
+
+    def forward(self, x: Tensor) -> Tensor:
+        """
+        Forward function that does knowledge-base lookup.
+        Args:
+             x: input, of shape (*, E) where E is embedding_dim
+                as passed to constructor
+             y: output of knowledge-base lookup, of shape (*, E)
+
+        # TODO: later we can try multiplying by a projection of x or something like that.
+        """
+        assert torch.all(x - x == 0)
+        x = self.in_proj(x) # now (*, M*N)
+        assert torch.all(x - x == 0)
+        x = x.reshape(*x.shape[:-1], self.N, self.M) # now (*, N, M)
+        x = x.log_softmax(dim=-1) # now normalized logprobs, dim= (*, N, M)
+        assert torch.all(x - x == 0)
+        if random.random() < 0.001:
+            entropy = (x * x.exp()).sum(dim=-1).mean()
+            #  print("Entropy = ", entropy)
+        # only need 'combined_indexes', call them 'indexes'.
+        _, indexes, weights = sample_combined(x, self.K, input_is_log=True)
+        x = weighted_matrix_lookup(weights, indexes, self.knowledge_base) # now (*, D)
+        x = self.out_proj(x) # now (*, self.embedding_dim)
+        return x
+
+
+def _test_knowledge_base_lookup():
+    K = 16
+    N = 2
+    M = 128
+    D = 256
+    E = 255
+
+    knowledge_base: nn.Parameter = create_knowledge_base(M, N, D)
+    m = KnowledgeBaseLookup(M, N, D, K, E, knowledge_base)
+
+    B = 30
+    T = 40
+    x = torch.randn(B, T, E)
+    x.requires_grad = True
+    y = m(x)
+    assert y.shape == x.shape
+    y.sum().backward() # make sure backward doesn't crash..
+    print("y = ", y)
+    print("x.grad = ", x.grad)
+    print("knowlege_base.grad norm = ", knowledge_base.grad.norm())
+
+    dtype = torch.float32
+    device = torch.device('cuda')
+    train_pairs = [ (torch.randn(B, T, E, device=device, dtype=dtype),  torch.randn(B, T, E, device=device, dtype=dtype)) for _ in range(10) ]
+    from optim import Eve
+    optimizer = Eve(m.parameters(), lr=0.005, eps=1.0e-04)
+    m = m.to(device).to(dtype)
+
+
+    start = timeit.default_timer()
+
+# Epoch 0, batch 0, loss 1.0109944343566895
+# Epoch 10, batch 0, loss 1.0146660804748535
+# Epoch 20, batch 0, loss 1.0119813680648804
+# Epoch 30, batch 0, loss 1.0105408430099487
+# Epoch 40, batch 0, loss 1.0077732801437378
+# Epoch 50, batch 0, loss 1.0050103664398193
+# Epoch 60, batch 0, loss 1.0033129453659058
+# Epoch 70, batch 0, loss 1.0014232397079468
+# Epoch 80, batch 0, loss 0.9977912306785583
+# Epoch 90, batch 0, loss 0.8274348974227905
+# Epoch 100, batch 0, loss 0.3368612825870514
+# Epoch 110, batch 0, loss 0.11323091387748718
+# Time taken:  17.591704960912466
+    for epoch in range(150):
+        for n, (x,y) in enumerate(train_pairs):
+            y_out = m(x)
+            loss = ((y_out - y)**2).mean() * 100.0
+            if n % 10 == 0 and epoch % 10 == 0:
+                print(f"Epoch {epoch}, batch {n}, loss {loss.item()}")
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+
+    stop = timeit.default_timer()
+    print('Time taken: ', stop - start)
+
+def _test_knowledge_base_lookup_autocast():
+    K = 16
+    N = 2
+    M = 128
+    D = 256
+    E = 255
+
+    knowledge_base: nn.Parameter = create_knowledge_base(M, N, D)
+    m = KnowledgeBaseLookup(M, N, D, K, E, knowledge_base)
+
+    B = 30
+    T = 40
+    x = torch.randn(B, T, E)
+    x.requires_grad = True
+    y = m(x)
+    assert y.shape == x.shape
+    y.sum().backward() # make sure backward doesn't crash..
+    print("y = ", y)
+    print("x.grad = ", x.grad)
+    print("knowlege_base.grad norm = ", knowledge_base.grad.norm())
+
+    device = torch.device('cuda')
+    train_pairs = [ (torch.randn(B, T, E, device=device),  torch.randn(B, T, E, device=device)) for _ in range(10) ]
+    from optim import Eve
+    optimizer = Eve(m.parameters(), lr=0.005, eps=1.0e-04)
+    m = m.to(device)
+
+    scaler = GradScaler(enabled=True)
+
+    start = timeit.default_timer()
+
+
+    for epoch in range(150):
+        for n, (x,y) in enumerate(train_pairs):
+            y_out = m(x)
+            with torch.cuda.amp.autocast(enabled=True):
+                loss = ((y_out - y)**2).mean() * 100.0
+            if n % 10 == 0 and epoch % 10 == 0:
+                print(f"Epoch {epoch}, batch {n}, loss {loss.item()}")
+            scaler.scale(loss).backward()
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+
+    stop = timeit.default_timer()
+    print('Time taken: ', stop - start)
+
+
+
+if __name__ == '__main__':
+    _test_knowledge_base_lookup()
+    _test_knowledge_base_lookup_autocast()

egs/librispeech/ASR/pruned_transducer_stateless5/scaling.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless3/scaling.py

egs/librispeech/ASR/pruned_transducer_stateless5/test_model.py

@@ -0,0 +1,44 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./pruned_transducer_stateless5/test_model.py
+"""
+
+from train import get_params, get_transducer_model
+
+
+def test_model():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    model = get_transducer_model(params)
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+
+def main():
+    test_model()
+
+
+if __name__ == "__main__":
+    main()

icefall/checkpoint.py

@@ -346,7 +346,7 @@ def remove_checkpoints(
     for c in to_remove:
         os.remove(c)
 
-
+@torch.no_grad()
 def update_averaged_model(
     params: Dict[str, Tensor],
     model_cur: Union[nn.Module, DDP],
@@ -442,7 +442,7 @@ def average_checkpoints_with_averaged_model(
 
     return avg
 
-
+@torch.no_grad()
 def average_state_dict(
     state_dict_1: Dict[str, Tensor],
     state_dict_2: Dict[str, Tensor],