Add streaming Emformer stateless RNN-T. (#390)

* Add streaming Emformer stateless RNN-T.

* Update results for streaming Emformer.

* Minor fixes.

egs/librispeech/ASR/README.md

@@ -22,6 +22,7 @@ The following table lists the differences among them.
 | `pruned_transducer_stateless4`        | Conformer(modified) | Embedding + Conv1d | same as pruned_transducer_stateless2 + save averaged models periodically during training                        |
 | `pruned_transducer_stateless5`        | Conformer(modified) | Embedding + Conv1d | same as pruned_transducer_stateless4 + more layers + random combiner|
 | `pruned_transducer_stateless6`        | Conformer(modified) | Embedding + Conv1d | same as pruned_transducer_stateless4 + distillation with hubert|
+| `pruned_stateless_emformer_rnnt2`     | Emformer(from torchaudio) | Embedding + Conv1d | Using Emformer from torchaudio for streaming ASR|
 
 
 The decoder in `transducer_stateless` is modified from the paper

egs/librispeech/ASR/RESULTS.md

@@ -1,5 +1,71 @@
 ## Results
 
+### LibriSpeech BPE training results (Pruned Stateless Emformer RNN-T)
+
+[pruned_stateless_emformer_rnnt2](./pruned_stateless_emformer_rnnt2)
+
+Use <https://github.com/k2-fsa/icefall/pull/390>.
+
+Use [Emformer](https://arxiv.org/abs/2010.10759) from [torchaudio](https://github.com/pytorch/audio)
+for streaming ASR. The Emformer model is imported from torchaudio without modifications.
+
+|                                     | test-clean | test-other | comment                                |
+|-------------------------------------|------------|------------|----------------------------------------|
+| greedy search (max sym per frame 1) | 4.28       | 11.42       | --epoch 39 --avg 6  --max-duration 600 |
+| modified beam search                | 4.22       | 11.16       | --epoch 39 --avg 6  --max-duration 600 |
+| fast beam search                    | 4.29       | 11.26       | --epoch 39 --avg 6 --max-duration 600  |
+
+
+The training commands are:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./pruned_stateless_emformer_rnnt2/train.py \
+  --world-size 8 \
+  --num-epochs 40 \
+  --start-epoch 1 \
+  --exp-dir pruned_stateless_emformer_rnnt2/exp-full \
+  --full-libri 1 \
+  --use-fp16 0 \
+  --max-duration 200 \
+  --prune-range 5 \
+  --lm-scale 0.25 \
+  --master-port 12358 \
+  --num-encoder-layers 18 \
+  --left-context-length 128 \
+  --segment-length 8 \
+  --right-context-length 4
+```
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/ZyiqhAhmRjmr49xml4ofLw/>
+
+The decoding commands are:
+```bash
+for m in greedy_search fast_beam_search modified_beam_search; do
+  for epoch in 39; do
+    for avg in 6; do
+      ./pruned_stateless_emformer_rnnt2/decode.py \
+        --epoch $epoch \
+        --avg $avg \
+        --use-averaged-model 1 \
+        --exp-dir pruned_stateless_emformer_rnnt2/exp-full \
+        --max-duration 50 \
+        --decoding-method $m \
+        --num-encoder-layers 18 \
+        --left-context-length 128 \
+        --segment-length 8 \
+        --right-context-length 4
+    done
+  done
+done
+```
+
+You can find a pretrained model, training logs, decoding logs, and decoding
+results at:
+<https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-stateless-emformer-rnnt2-2022-06-01>
+
+
 ### LibriSpeech BPE training results (Pruned Stateless Transducer 5)
 
 [pruned_transducer_stateless5](./pruned_transducer_stateless5)

egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/asr_datamodule.py

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

egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/beam_search.py

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

egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/decode.py

@@ -0,0 +1,645 @@
+#!/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_stateless_emformer_rnnt/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_stateless_emformer_rnnt/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_stateless_emformer_rnnt/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_stateless_emformer_rnnt/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_stateless_emformer_rnnt/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_stateless_emformer_rnnt/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search
+./pruned_stateless_emformer_rnnt/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_stateless_emformer_rnnt/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 4 \
+    --max-contexts 4 \
+    --max-states 8
+"""
+
+
+import argparse
+import logging
+import math
+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 LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.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,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=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_stateless_emformer_rnnt/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
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    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.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    feature_lens += params.left_context_length
+    feature = torch.nn.functional.pad(
+        feature,
+        pad=(0, 0, 0, params.left_context_length),
+        value=LOG_EPS,
+    )
+
+    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 == "greedy_search"
+        and params.max_sym_per_frame == 1
+    ):
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for 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,
+            encoder_out_lens=encoder_out_lens,
+            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
+        }
+    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 = 50
+    else:
+        log_interval = 10
+
+    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()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += (
+            f"-{params.decoding_method}-beam-size-{params.beam_size}"
+        )
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    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:
+        if params.iter > 0:
+            filenames = find_checkpoints(
+                params.exp_dir, iteration=-params.iter
+            )[: params.avg + 1]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.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_stateless_emformer_rnnt2/decoder.py

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

egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/emformer.py

@@ -0,0 +1,313 @@
+# Copyright      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.
+
+import math
+from typing import List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from subsampling import Conv2dSubsampling, VggSubsampling
+
+try:
+    from torchaudio.models import Emformer as _Emformer
+except ImportError:
+    import torchaudio
+
+    print(
+        "Please install torchaudio >= 0.11.0. "
+        f"Current version: {torchaudio.__version__}"
+    )
+    raise
+
+
+def unstack_states(
+    states: List[List[torch.Tensor]],
+) -> List[List[List[torch.Tensor]]]:
+    """Unstack the emformer state corresponding to a batch of utterances
+    into a list of states, were the i-th entry is the state from the i-th
+    utterance in the batch.
+
+    Args:
+      states:
+        A list-of-list of tensors. ``len(states)`` equals to number of
+        layers in the emformer. ``states[i]]`` contains the states for
+        the i-th layer. ``states[i][k]`` is either a 3-D tensor of shape
+        ``(T, N, C)`` or a 2-D tensor of shape ``(C, N)``
+    """
+    batch_size = states[0][0].size(1)
+    num_layers = len(states)
+
+    ans = [None] * batch_size
+    for i in range(batch_size):
+        ans[i] = [[] for _ in range(num_layers)]
+
+    for li, layer in enumerate(states):
+        for s in layer:
+            s_list = s.unbind(dim=1)
+            # We will use stack(dim=1) later in stack_states()
+            for bi, b in enumerate(ans):
+                b[li].append(s_list[bi])
+    return ans
+
+
+def stack_states(
+    state_list: List[List[List[torch.Tensor]]],
+) -> List[List[torch.Tensor]]:
+    """Stack list of emformer states that correspond to separate utterances
+    into a single emformer state so that it can be used as an input for
+    emformer when those utterances are formed into a batch.
+
+    Note:
+      It is the inverse of :func:`unstack_states`.
+
+    Args:
+      state_list:
+        Each element in state_list corresponding to the internal state
+        of the emformer model for a single utterance.
+    Returns:
+      Return a new state corresponding to a batch of utterances.
+      See the input argument of :func:`unstack_states` for the meaning
+      of the returned tensor.
+    """
+    batch_size = len(state_list)
+    ans = []
+    for layer in state_list[0]:
+        # layer is a list of tensors
+        if batch_size > 1:
+            ans.append([[s] for s in layer])
+            # Note: We will stack ans[layer][s][] later to get ans[layer][s]
+        else:
+            ans.append([s.unsqueeze(1) for s in layer])
+
+    for b, states in enumerate(state_list[1:], 1):
+        for li, layer in enumerate(states):
+            for si, s in enumerate(layer):
+                ans[li][si].append(s)
+                if b == batch_size - 1:
+                    ans[li][si] = torch.stack(ans[li][si], dim=1)
+                    # We will use unbind(dim=1) later in unstack_states()
+    return ans
+
+
+class Emformer(EncoderInterface):
+    """This is just a simple wrapper around torchaudio.models.Emformer.
+    We may replace it with our own implementation some time later.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        output_dim: int,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int,
+        num_encoder_layers: int,
+        segment_length: int,
+        left_context_length: int,
+        right_context_length: int,
+        max_memory_size: int = 0,
+        dropout: float = 0.1,
+        subsampling_factor: int = 4,
+        vgg_frontend: bool = False,
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          output_dim:
+            The output dimension of the model.
+          d_model:
+            Attention dimension.
+          nhead:
+            Number of heads in multi-head attention.
+          dim_feedforward:
+            The output dimension of the feedforward layers in encoder.
+          num_encoder_layers:
+            Number of encoder layers.
+          segment_length:
+            Number of frames per segment before subsampling.
+          left_context_length:
+            Number of frames in the left context before subsampling.
+          right_context_length:
+            Number of frames in the right context before subsampling.
+          max_memory_size:
+            TODO.
+          dropout:
+            Dropout in encoder.
+          subsampling_factor:
+            Number of output frames is num_in_frames // subsampling_factor.
+            Currently, subsampling_factor MUST be 4.
+          vgg_frontend:
+            True to use vgg style frontend for subsampling.
+        """
+        super().__init__()
+
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_features -> d_model
+        if vgg_frontend:
+            self.encoder_embed = VggSubsampling(num_features, d_model)
+        else:
+            self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.segment_length = segment_length  # before subsampling
+        self.right_context_length = right_context_length
+
+        assert right_context_length % subsampling_factor == 0
+        assert segment_length % subsampling_factor == 0
+        assert left_context_length % subsampling_factor == 0
+
+        left_context_length = left_context_length // subsampling_factor
+        right_context_length = right_context_length // subsampling_factor
+        segment_length = segment_length // subsampling_factor
+
+        self.model = _Emformer(
+            input_dim=d_model,
+            num_heads=nhead,
+            ffn_dim=dim_feedforward,
+            num_layers=num_encoder_layers,
+            segment_length=segment_length,
+            dropout=dropout,
+            activation="relu",
+            left_context_length=left_context_length,
+            right_context_length=right_context_length,
+            max_memory_size=max_memory_size,
+            weight_init_scale_strategy="depthwise",
+            tanh_on_mem=False,
+            negative_inf=-1e8,
+        )
+
+        self.encoder_output_layer = nn.Sequential(
+            nn.Dropout(p=dropout), nn.Linear(d_model, output_dim)
+        )
+        self.log_eps = math.log(1e-10)
+
+        self._init_state = torch.jit.Attribute([], List[List[torch.Tensor]])
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            Input features of shape (N, T, C).
+          x_lens:
+            A int32 tensor of shape (N,) containing valid frames in `x` before
+            padding. We have `x.size(1) == x_lens.max()`
+        Returns:
+          Return a tuple containing two tensors:
+
+            - encoder_out, a tensor of shape (N, T', C)
+            - encoder_out_lens, a int32 tensor of shape (N,) containing the
+              valid frames in `encoder_out` before padding
+        """
+        x = nn.functional.pad(
+            x,
+            # (left, right, top, bottom)
+            # left/right are for the channel dimension, i.e., axis 2
+            # top/bottom are for the time dimension, i.e., axis 1
+            (0, 0, 0, self.right_context_length),
+            value=self.log_eps,
+        )  # (N, T, C) -> (N, T+right_context_length, C)
+
+        x = self.encoder_embed(x)
+
+        # Caution: We assume the subsampling factor is 4!
+        x_lens = (((x_lens - 1) >> 1) - 1) >> 1
+
+        emformer_out, emformer_out_lens = self.model(x, x_lens)
+        logits = self.encoder_output_layer(emformer_out)
+
+        return logits, emformer_out_lens
+
+    @torch.jit.export
+    def streaming_forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        states: Optional[List[List[torch.Tensor]]] = None,
+    ):
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C). Note: x also contains right
+            context frames.
+          x_lens:
+            A 2-D tensor of shap containing the number of valid frames for each
+            element in `x` before padding. Note: It also counts right context
+            frames.
+          states:
+            Internal states of the model.
+        Returns:
+          Return a tuple containing 3 tensors:
+            - encoder_out, a 3-D tensor of shape (N, T, C)
+            - encoder_out_lens: a 1-D tensor of shape (N,)
+            - next_state, internal model states for the next invocation
+        """
+        x = self.encoder_embed(x)
+
+        # Caution: We assume the subsampling factor is 4!
+        x_lens = (((x_lens - 1) >> 1) - 1) >> 1
+
+        emformer_out, emformer_out_lens, states = self.model.infer(
+            x, x_lens, states
+        )
+
+        if x.size(1) != (
+            self.model.segment_length + self.model.right_context_length
+        ):
+            raise ValueError(
+                "Incorrect input shape."
+                f"{x.size(1)} vs {self.model.segment_length} + "
+                f"{self.model.right_context_length}"
+            )
+
+        logits = self.encoder_output_layer(emformer_out)
+
+        return logits, emformer_out_lens, states
+
+    @torch.jit.export
+    def get_init_state(self, device: torch.device) -> List[List[torch.Tensor]]:
+        """Return the initial state of each layer.
+
+        Returns:
+          Return the initial state of each layer. NOTE: the returned
+          tensors are on the given device. `len(ans) == num_emformer_layers`.
+        """
+        if self._init_state:
+            # Note(fangjun): It is OK to share the init state as it is
+            # not going to be modified by the model
+            return self._init_state
+
+        batch_size = 1
+
+        ans: List[List[torch.Tensor]] = []
+        for layer in self.model.emformer_layers:
+            s = layer._init_state(batch_size=batch_size, device=device)
+            ans.append(s)
+
+        self._init_state = ans
+
+        return ans

egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/encoder_interface.py

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

egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/export.py

@@ -0,0 +1,281 @@
+#!/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.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./prunted_stateless_emformer_rnnt/export.py \
+  --exp-dir ./prunted_stateless_emformer_rnnt/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `prunted_stateless_emformer_rnnt/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./prunted_stateless_emformer_rnnt/decode.py \
+        --exp-dir ./prunted_stateless_emformer_rnnt/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 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="prunted_stateless_emformer_rnnt/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    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",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    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:
+        if params.iter > 0:
+            filenames = find_checkpoints(
+                params.exp_dir, iteration=-params.iter
+            )[: params.avg + 1]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+    for p in model.parameters():
+        p.requires_grad_(False)
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        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/librispeech/ASR/pruned_stateless_emformer_rnnt2/joiner.py

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

egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/model.py

@@ -0,0 +1,169 @@
+# 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 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,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, C) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, C) 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, C). It should contain
+            one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, C) and (N, U, C). Its
+            output shape is (N, T, U, C). 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
+
+    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,
+    ) -> 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
+        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)
+        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, C]
+        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
+
+        simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+            lm=decoder_out,
+            am=encoder_out,
+            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, C]
+        # lm_pruned : [B, T, prune_range, C]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=encoder_out, lm=decoder_out, ranges=ranges
+        )
+
+        # logits : [B, T, prune_range, C]
+        logits = self.joiner(am_pruned, lm_pruned)
+
+        pruned_loss = k2.rnnt_loss_pruned(
+            logits=logits,
+            symbols=y_padded,
+            ranges=ranges,
+            termination_symbol=blank_id,
+            boundary=boundary,
+            reduction="sum",
+        )
+
+        return (simple_loss, pruned_loss)

egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/noam.py

@@ -0,0 +1,104 @@
+# Copyright    2021 University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# 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
+
+
+class Noam(object):
+    """
+    Implements Noam optimizer.
+
+    Proposed in
+    "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf
+
+    Modified from
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/optimizer.py  # noqa
+
+    Args:
+      params:
+        iterable of parameters to optimize or dicts defining parameter groups
+      model_size:
+        attention dimension of the transformer model
+      factor:
+        learning rate factor
+      warm_step:
+        warmup steps
+    """
+
+    def __init__(
+        self,
+        params,
+        model_size: int = 256,
+        factor: float = 10.0,
+        warm_step: int = 25000,
+        weight_decay=0,
+    ) -> None:
+        """Construct an Noam object."""
+        self.optimizer = torch.optim.Adam(
+            params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=weight_decay
+        )
+        self._step = 0
+        self.warmup = warm_step
+        self.factor = factor
+        self.model_size = model_size
+        self._rate = 0
+
+    @property
+    def param_groups(self):
+        """Return param_groups."""
+        return self.optimizer.param_groups
+
+    def step(self):
+        """Update parameters and rate."""
+        self._step += 1
+        rate = self.rate()
+        for p in self.optimizer.param_groups:
+            p["lr"] = rate
+        self._rate = rate
+        self.optimizer.step()
+
+    def rate(self, step=None):
+        """Implement `lrate` above."""
+        if step is None:
+            step = self._step
+        return (
+            self.factor
+            * self.model_size ** (-0.5)
+            * min(step ** (-0.5), step * self.warmup ** (-1.5))
+        )
+
+    def zero_grad(self):
+        """Reset gradient."""
+        self.optimizer.zero_grad()
+
+    def state_dict(self):
+        """Return state_dict."""
+        return {
+            "_step": self._step,
+            "warmup": self.warmup,
+            "factor": self.factor,
+            "model_size": self.model_size,
+            "_rate": self._rate,
+            "optimizer": self.optimizer.state_dict(),
+        }
+
+    def load_state_dict(self, state_dict):
+        """Load state_dict."""
+        for key, value in state_dict.items():
+            if key == "optimizer":
+                self.optimizer.load_state_dict(state_dict["optimizer"])
+            else:
+                setattr(self, key, value)

egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/subsampling.py

@@ -0,0 +1 @@
+../conformer_ctc/subsampling.py

egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/test_emformer.py

@@ -0,0 +1,142 @@
+#!/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_stateless_emformer_rnnt/test_emformer.py
+"""
+
+import torch
+from emformer import Emformer, stack_states, unstack_states
+
+
+def test_emformer():
+    N = 3
+    T = 300
+    C = 80
+
+    output_dim = 500
+
+    encoder = Emformer(
+        num_features=C,
+        output_dim=output_dim,
+        d_model=512,
+        nhead=8,
+        dim_feedforward=2048,
+        num_encoder_layers=20,
+        segment_length=16,
+        left_context_length=120,
+        right_context_length=4,
+        vgg_frontend=False,
+    )
+
+    x = torch.rand(N, T, C)
+    x_lens = torch.randint(100, T, (N,))
+    x_lens[0] = T
+
+    y, y_lens = encoder(x, x_lens)
+
+    y_lens = (((x_lens - 1) >> 1) - 1) >> 1
+    assert x.size(0) == x.size(0)
+    assert y.size(1) == max(y_lens)
+    assert y.size(2) == output_dim
+
+    num_param = sum([p.numel() for p in encoder.parameters()])
+    print(f"Number of encoder parameters: {num_param}")
+
+
+def test_emformer_streaming_forward():
+    N = 3
+    C = 80
+
+    output_dim = 500
+
+    encoder = Emformer(
+        num_features=C,
+        output_dim=output_dim,
+        d_model=512,
+        nhead=8,
+        dim_feedforward=2048,
+        num_encoder_layers=20,
+        segment_length=16,
+        left_context_length=120,
+        right_context_length=4,
+        vgg_frontend=False,
+    )
+
+    x = torch.rand(N, 23, C)
+    x_lens = torch.full((N,), 23)
+    y, y_lens, states = encoder.streaming_forward(x=x, x_lens=x_lens)
+
+    state_list = unstack_states(states)
+    states2 = stack_states(state_list)
+
+    for ss, ss2 in zip(states, states2):
+        for s, s2 in zip(ss, ss2):
+            assert torch.allclose(s, s2), f"{s.sum()}, {s2.sum()}"
+
+
+def test_emformer_init_state():
+    num_encoder_layers = 20
+    d_model = 512
+    encoder = Emformer(
+        num_features=80,
+        output_dim=500,
+        d_model=512,
+        nhead=8,
+        dim_feedforward=2048,
+        num_encoder_layers=num_encoder_layers,
+        segment_length=16,
+        left_context_length=120,
+        right_context_length=4,
+        vgg_frontend=False,
+    )
+    init_state = encoder.get_init_state()
+    assert len(init_state) == num_encoder_layers
+    layer0_state = init_state[0]
+    assert len(layer0_state) == 4
+
+    assert layer0_state[0].shape == (
+        0,  # max_memory_size
+        1,  # batch_size
+        d_model,  # input_dim
+    )
+
+    assert layer0_state[1].shape == (
+        encoder.model.left_context_length,
+        1,  # batch_size
+        d_model,  # input_dim
+    )
+    assert layer0_state[2].shape == layer0_state[1].shape
+    assert layer0_state[3].shape == (
+        1,  # always 1
+        1,  # batch_size
+    )
+
+
+@torch.no_grad()
+def main():
+    test_emformer()
+    test_emformer_streaming_forward()
+    test_emformer_init_state()
+
+
+if __name__ == "__main__":
+    torch.manual_seed(20220329)
+    main()

egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/test_model.py

@@ -0,0 +1,59 @@
+#!/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_stateless_emformer_rnnt/test_model.py
+"""
+
+import torch
+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
+    params.unk_id = 2
+
+    params.attention_dim = 512
+    params.nhead = 8
+    params.dim_feedforward = 2048
+    params.num_encoder_layers = 18
+    params.left_context_length = 128
+    params.segment_length = 8
+    params.right_context_length = 4
+    params.memory_size = 0
+
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+    model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+    torch.jit.script(model)
+
+
+def main():
+    test_model()
+
+
+if __name__ == "__main__":
+    main()

egs/librispeech/ASR/pruned_transducer_stateless/joiner.py

@@ -32,13 +32,15 @@ class Joiner(nn.Module):
         """
         Args:
           encoder_out:
-            Output from the encoder. Its shape is (N, T, s_range, C).
+            Output from the encoder. Its shape is (N, T, s_range, C) during
+            training or (N, C) in case of streaming decoding.
           decoder_out:
-            Output from the decoder. Its shape is (N, T, s_range, C).
+            Output from the decoder. Its shape is (N, T, s_range, C) during
-        Returns:
+            training or (N, C) in case of streaming decoding.
           Return a tensor of shape (N, T, s_range, C).
         """
-        assert encoder_out.ndim == decoder_out.ndim == 4
+        assert encoder_out.ndim == decoder_out.ndim
+        assert encoder_out.ndim in (2, 4)
         assert encoder_out.shape == decoder_out.shape
 
         logit = encoder_out + decoder_out

egs/librispeech/ASR/pruned_transducer_stateless3/asr_datamodule.py

@@ -277,10 +277,12 @@ class AsrDataModule:
                 cut_transforms=transforms,
                 return_cuts=self.args.return_cuts,
             )
-        valid_sampler = BucketingSampler(
+        valid_sampler = DynamicBucketingSampler(
             cuts_valid,
             max_duration=self.args.max_duration,
             shuffle=False,
+            num_buckets=self.args.num_buckets,
+            drop_last=False,
         )
         logging.info("About to create dev dataloader")
         valid_dl = DataLoader(
@@ -301,8 +303,12 @@ class AsrDataModule:
             else PrecomputedFeatures(),
             return_cuts=self.args.return_cuts,
         )
-        sampler = BucketingSampler(
+        sampler = DynamicBucketingSampler(
-            cuts, max_duration=self.args.max_duration, shuffle=False
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+            num_buckets=self.args.num_buckets,
+            drop_last=True,
         )
         logging.debug("About to create test dataloader")
         test_dl = DataLoader(

egs/librispeech/ASR/pruned_transducer_stateless3/train.py

@@ -799,7 +799,7 @@ def train_one_epoch(
                 f"tot_loss[{tot_loss}], "
                 f"libri_tot_loss[{libri_tot_loss}], "
                 f"giga_tot_loss[{giga_tot_loss}], "
-                f"batch size: {batch_size}"
+                f"batch size: {batch_size}, "
                 f"lr: {cur_lr:.2e}"
             )
 

egs/librispeech/ASR/tdnn_lstm_ctc/asr_datamodule.py

@@ -26,9 +26,9 @@ from typing import Any, Dict, Optional
 import torch
 from lhotse import CutSet, Fbank, FbankConfig, load_manifest
 from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
-    BucketingSampler,
     CutConcatenate,
     CutMix,
+    DynamicBucketingSampler,
     K2SpeechRecognitionDataset,
     PrecomputedFeatures,
     SingleCutSampler,
@@ -113,7 +113,7 @@ class LibriSpeechAsrDataModule:
             "--num-buckets",
             type=int,
             default=30,
-            help="The number of buckets for the BucketingSampler"
+            help="The number of buckets for the DynamicBucketingSampler"
             "(you might want to increase it for larger datasets).",
         )
         group.add_argument(
@@ -306,13 +306,12 @@ class LibriSpeechAsrDataModule:
             )
 
         if self.args.bucketing_sampler:
-            logging.info("Using BucketingSampler.")
+            logging.info("Using DynamicBucketingSampler.")
-            train_sampler = BucketingSampler(
+            train_sampler = DynamicBucketingSampler(
                 cuts_train,
                 max_duration=self.args.max_duration,
                 shuffle=self.args.shuffle,
                 num_buckets=self.args.num_buckets,
-                bucket_method="equal_duration",
                 drop_last=self.args.drop_last,
             )
         else:
@@ -367,7 +366,7 @@ class LibriSpeechAsrDataModule:
                 cut_transforms=transforms,
                 return_cuts=self.args.return_cuts,
             )
-        valid_sampler = BucketingSampler(
+        valid_sampler = DynamicBucketingSampler(
             cuts_valid,
             max_duration=self.args.max_duration,
             shuffle=False,
@@ -391,8 +390,10 @@ class LibriSpeechAsrDataModule:
             else eval(self.args.input_strategy)(),
             return_cuts=self.args.return_cuts,
         )
-        sampler = BucketingSampler(
+        sampler = DynamicBucketingSampler(
-            cuts, max_duration=self.args.max_duration, shuffle=False
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
         )
         logging.debug("About to create test dataloader")
         test_dl = DataLoader(