Update RESULTS.md (#388)

* update RESULT.md about pruned_transducer_stateless4

* Update RESULT.md

This PR is only to update RESULT.md about pruned_transducer_stateless4.

* set default value of --use-averaged-model to True

* update RESULTS.md and add decode command

* minor fix

* update export.py

* add uploaded files links

* update link

* fix typos

egs/librispeech/ASR/RESULTS.md

@@ -259,6 +259,126 @@ You can find a pretrained model, training logs, decoding logs, and decoding
 results at:
 <https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless5-narrower-2022-05-13>
 
+
+### LibriSpeech BPE training results (Pruned Stateless Transducer 4)
+
+[pruned_transducer_stateless4](./pruned_transducer_stateless4)
+
+This version saves averaged model during training, and decodes with averaged model.
+
+See <https://github.com/k2-fsa/icefall/issues/337> for details about the idea of model averaging.
+
+#### Training on full librispeech
+
+See <https://github.com/k2-fsa/icefall/pull/344>
+
+Using commit `ec0b0e92297cc03fdb09f48cd235e84d2c04156b`.
+
+The WERs are:
+
+|                                     | test-clean | test-other | comment                                                                       |
+|-------------------------------------|------------|------------|-------------------------------------------------------------------------------|
+| greedy search (max sym per frame 1) | 2.75       | 6.74       | --epoch 30 --avg 6  --use_averaged_model False                                |
+| greedy search (max sym per frame 1) | 2.69       | 6.64       | --epoch 30 --avg 6  --use_averaged_model True                                 |
+| fast beam search                    | 2.72       | 6.67       | --epoch 30 --avg 6  --use_averaged_model False                                |
+| fast beam search                    | 2.66       | 6.6        | --epoch 30 --avg 6  --use_averaged_model True                                 |
+| modified beam search                | 2.67       | 6.68       | --epoch 30 --avg 6  --use_averaged_model False                                |
+| modified beam search                | 2.62       | 6.57       | --epoch 30 --avg 6  --use_averaged_model True                                 |
+
+The training command is:
+
+```bash
+./pruned_transducer_stateless4/train.py \
+  --world-size 6 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless4/exp \
+  --full-libri 1 \
+  --max-duration 300 \
+  --save-every-n 8000 \
+  --keep-last-k 20 \
+  --average-period 100
+```
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/QOGSPBgsR8KzcRMmie9JGw/>
+
+The decoding command using greedy search is:
+```bash
+./pruned_transducer_stateless4/decode.py \
+  --epoch 30 \
+  --avg 6 \
+  --exp-dir pruned_transducer_stateless4/exp \
+  --max-duration 300 \
+  --decoding-method greedy_search \
+  --use-averaged-model True
+```
+
+The decoding command using fast beam search is:
+```bash
+./pruned_transducer_stateless4/decode.py \
+  --epoch 30 \
+  --avg 6 \
+  --exp-dir pruned_transducer_stateless4/exp \
+  --max-duration 300 \
+  --decoding-method fast_beam_search \
+  --use-averaged-model True \
+  --beam 4 \
+  --max-contexts 4 \
+  --max-states 8
+```
+
+The decoding command using modified beam search is:
+```bash
+./pruned_transducer_stateless4/decode.py \
+  --epoch 30 \
+  --avg 6 \
+  --exp-dir pruned_transducer_stateless4/exp \
+  --max-duration 300 \
+  --decoding-method modified_beam_search \
+  --use-averaged-model True \
+  --beam-size 4
+```
+
+Pretrained models, training logs, decoding logs, and decoding results
+are available at
+<https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless4-2022-06-03>
+
+#### Training on train-clean-100
+
+See <https://github.com/k2-fsa/icefall/pull/344>
+
+Using commit `ec0b0e92297cc03fdb09f48cd235e84d2c04156b`.
+
+The WERs are:
+
+|                                     | test-clean | test-other | comment                                                                       |
+|-------------------------------------|------------|------------|-------------------------------------------------------------------------------|
+| greedy search (max sym per frame 1) | 7.0        | 18.95      | --epoch 30 --avg 10 --use_averaged_model False                                |
+| greedy search (max sym per frame 1) | 6.92       | 18.65      | --epoch 30 --avg 10 --use_averaged_model True                                 |
+| fast beam search                    | 6.82       | 18.47      | --epoch 30 --avg 10 --use_averaged_model False                                |
+| fast beam search                    | 6.74       | 18.2       | --epoch 30 --avg 10 --use_averaged_model True                                 |
+| modified beam search                | 6.74       | 18.39      | --epoch 30 --avg 10 --use_averaged_model False                                |
+| modified beam search                | 6.74       | 18.12      | --epoch 30 --avg 10 --use_averaged_model True                                 |
+
+The training command is:
+
+```bash
+./pruned_transducer_stateless4/train.py \
+  --world-size 3 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless4/exp \
+  --full-libri 0 \
+  --max-duration 300 \
+  --save-every-n 8000 \
+  --keep-last-k 20 \
+  --average-period 100
+```
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/YVYHq1irQS69s9bW1vQ06Q/>
+
 ### LibriSpeech BPE training results (Pruned Stateless Transducer 3, 2022-04-29)
 
 [pruned_transducer_stateless3](./pruned_transducer_stateless3)

egs/librispeech/ASR/pruned_transducer_stateless4/decode.py

@@ -128,7 +128,7 @@ def get_parser():
     parser.add_argument(
         "--use-averaged-model",
         type=str2bool,
-        default=False,
+        default=True,
         help="Whether to load averaged model. Currently it only supports "
         "using --epoch. If True, it would decode with the averaged model "
         "over the epoch range from `epoch-avg` (excluded) to `epoch`."

egs/librispeech/ASR/pruned_transducer_stateless4/export.py

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

egs/librispeech/ASR/pruned_transducer_stateless4/export.py

@@ -0,0 +1,273 @@
+#!/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:
+./pruned_transducer_stateless4/export.py \
+  --exp-dir ./pruned_transducer_stateless4/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 `pruned_transducer_stateless4/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./pruned_transducer_stateless4/decode.py \
+        --exp-dir ./pruned_transducer_stateless4/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100 \
+        --bpe-model data/lang_bpe_500/bpe.model \
+        --use-averaged-model False
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+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 str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--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",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    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")
+
+    logging.info(f"device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    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()
+
+    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()