Model average (#344)

* First upload of model average codes.

* minor fix

* update decode file

* update .flake8

* rename pruned_transducer_stateless3 to pruned_transducer_stateless4

* change epoch number counter starting from 1 instead of 0

* minor fix of pruned_transducer_stateless4/train.py

* refactor the checkpoint.py

* minor fix, update docs, and modify the epoch number to count from 1 in the pruned_transducer_stateless4/decode.py

* update author info

* add docs of the scaling in function average_checkpoints_with_averaged_model

.flake8

@@ -9,6 +9,7 @@ per-file-ignores =
     egs/tedlium3/ASR/*/conformer.py: E501,
     egs/gigaspeech/ASR/*/conformer.py: E501,
     egs/librispeech/ASR/pruned_transducer_stateless2/*.py: E501,
+    egs/librispeech/ASR/pruned_transducer_stateless4/*.py: E501,
     egs/librispeech/ASR/*/optim.py: E501,
     egs/librispeech/ASR/*/scaling.py: E501,
 

egs/librispeech/ASR/pruned_transducer_stateless4/__init__.py

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

egs/librispeech/ASR/pruned_transducer_stateless4/asr_datamodule.py

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

egs/librispeech/ASR/pruned_transducer_stateless4/beam_search.py

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

egs/librispeech/ASR/pruned_transducer_stateless4/conformer.py

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

egs/librispeech/ASR/pruned_transducer_stateless4/decode.py

@@ -0,0 +1,603 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 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_stateless4/decode.py \
+        --epoch 30 \
+        --avg 15 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --max-duration 100 \
+        --decoding-method greedy_search
+
+(2) beam search
+./pruned_transducer_stateless4/decode.py \
+        --epoch 30 \
+        --avg 15 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --max-duration 100 \
+        --decoding-method beam_search \
+        --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless4/decode.py \
+        --epoch 30 \
+        --avg 15 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --max-duration 100 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+(4) fast beam search
+./pruned_transducer_stateless4/decode.py \
+        --epoch 30 \
+        --avg 15 \
+        --exp-dir ./pruned_transducer_stateless2/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 LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    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=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_transducer_stateless4/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""",
+    )
+
+    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 = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(
+        x=feature, x_lens=feature_lens
+    )
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for 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
+        }
+    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()
+    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> is 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:
+        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
+
+    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_transducer_stateless4/decoder.py

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

egs/librispeech/ASR/pruned_transducer_stateless4/encoder_interface.py

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

egs/librispeech/ASR/pruned_transducer_stateless4/export.py

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

egs/librispeech/ASR/pruned_transducer_stateless4/joiner.py

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

egs/librispeech/ASR/pruned_transducer_stateless4/model.py

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

egs/librispeech/ASR/pruned_transducer_stateless4/optim.py

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

egs/librispeech/ASR/pruned_transducer_stateless4/scaling.py

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

icefall/checkpoint.py

@@ -1,4 +1,5 @@
-# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Zengwei Yao)
 #
 # See ../../LICENSE for clarification regarding multiple authors
 #
@@ -25,6 +26,7 @@ from typing import Any, Dict, List, Optional, Union
 import torch
 import torch.nn as nn
 from lhotse.dataset.sampling.base import CutSampler
+from torch import Tensor
 from torch.cuda.amp import GradScaler
 from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.optim import Optimizer
@@ -37,6 +39,7 @@ LRSchedulerType = object
 def save_checkpoint(
     filename: Path,
     model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
     params: Optional[Dict[str, Any]] = None,
     optimizer: Optional[Optimizer] = None,
     scheduler: Optional[LRSchedulerType] = None,
@@ -51,6 +54,8 @@ def save_checkpoint(
         The checkpoint filename.
       model:
         The model to be saved. We only save its `state_dict()`.
+      model_avg:
+        The stored model averaged from the start of training.
       params:
         User defined parameters, e.g., epoch, loss.
       optimizer:
@@ -80,6 +85,9 @@ def save_checkpoint(
         "sampler": sampler.state_dict() if sampler is not None else None,
     }
 
+    if model_avg is not None:
+        checkpoint["model_avg"] = model_avg.state_dict()
+
     if params:
         for k, v in params.items():
             assert k not in checkpoint
@@ -91,6 +99,7 @@ def save_checkpoint(
 def load_checkpoint(
     filename: Path,
     model: nn.Module,
+    model_avg: Optional[nn.Module] = None,
     optimizer: Optional[Optimizer] = None,
     scheduler: Optional[LRSchedulerType] = None,
     scaler: Optional[GradScaler] = None,
@@ -118,6 +127,11 @@ def load_checkpoint(
 
     checkpoint.pop("model")
 
+    if model_avg is not None and "model_avg" in checkpoint:
+        logging.info("Loading averaged model")
+        model_avg.load_state_dict(checkpoint["model_avg"], strict=strict)
+        checkpoint.pop("model_avg")
+
     def load(name, obj):
         s = checkpoint.get(name, None)
         if obj and s:
@@ -181,6 +195,7 @@ def save_checkpoint_with_global_batch_idx(
     out_dir: Path,
     global_batch_idx: int,
     model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
     params: Optional[Dict[str, Any]] = None,
     optimizer: Optional[Optimizer] = None,
     scheduler: Optional[LRSchedulerType] = None,
@@ -201,6 +216,8 @@ def save_checkpoint_with_global_batch_idx(
       model:
         The neural network model whose `state_dict` will be saved in the
         checkpoint.
+      model_avg:
+        The stored model averaged from the start of training.
       params:
         A dict of training configurations to be saved.
       optimizer:
@@ -223,6 +240,7 @@ def save_checkpoint_with_global_batch_idx(
     save_checkpoint(
         filename=filename,
         model=model,
+        model_avg=model_avg,
         params=params,
         optimizer=optimizer,
         scheduler=scheduler,
@@ -327,3 +345,127 @@ def remove_checkpoints(
     to_remove = checkpoints[topk:]
     for c in to_remove:
         os.remove(c)
+
+
+def update_averaged_model(
+    params: Dict[str, Tensor],
+    model_cur: Union[nn.Module, DDP],
+    model_avg: nn.Module,
+) -> None:
+    """Update the averaged model:
+    model_avg = model_cur * (average_period / batch_idx_train)
+      + model_avg * ((batch_idx_train - average_period) / batch_idx_train)
+
+    Args:
+      params:
+        User defined parameters, e.g., epoch, loss.
+      model_cur:
+        The current model.
+      model_avg:
+        The averaged model to be updated.
+    """
+    weight_cur = params.average_period / params.batch_idx_train
+    weight_avg = 1 - weight_cur
+
+    if isinstance(model_cur, DDP):
+        model_cur = model_cur.module
+
+    cur = model_cur.state_dict()
+    avg = model_avg.state_dict()
+
+    average_state_dict(
+        state_dict_1=avg,
+        state_dict_2=cur,
+        weight_1=weight_avg,
+        weight_2=weight_cur,
+    )
+
+
+def average_checkpoints_with_averaged_model(
+    filename_start: str,
+    filename_end: str,
+    device: torch.device = torch.device("cpu"),
+) -> Dict[str, Tensor]:
+    """Average model parameters over the range with given
+    start model (excluded) and end model.
+
+    Let start = batch_idx_train of model-start;
+        end = batch_idx_train of model-end;
+        interval = end - start.
+    Then the average model over range from start (excluded) to end is
+    (1) avg = (model_end * end - model_start * start) / interval.
+    It can be written as
+    (2) avg = model_end * weight_end + model_start * weight_start,
+        where weight_end = end / interval,
+              weight_start = -start / interval = 1 - weight_end.
+    Since the terms `weight_end` and `weight_start` would be large
+    if the model has been trained for lots of batches, which would cause
+    overflow when multiplying the model parameters.
+    To avoid this, we rewrite (2) as:
+    (3) avg = (model_end + model_start * (weight_start / weight_end))
+              * weight_end
+
+    The model index could be epoch number or checkpoint number.
+
+    Args:
+      filename_start:
+        Checkpoint filename of the start model. We assume it
+        is saved by :func:`save_checkpoint`.
+      filename_end:
+        Checkpoint filename of the end model. We assume it
+        is saved by :func:`save_checkpoint`.
+      device:
+        Move checkpoints to this device before averaging.
+    """
+    state_dict_start = torch.load(filename_start, map_location=device)
+    state_dict_end = torch.load(filename_end, map_location=device)
+
+    batch_idx_train_start = state_dict_start["batch_idx_train"]
+    batch_idx_train_end = state_dict_end["batch_idx_train"]
+    interval = batch_idx_train_end - batch_idx_train_start
+    assert interval > 0, interval
+    weight_end = batch_idx_train_end / interval
+    weight_start = 1 - weight_end
+
+    model_end = state_dict_end["model_avg"]
+    model_start = state_dict_start["model_avg"]
+    avg = model_end
+
+    # scale the weight to avoid overflow
+    average_state_dict(
+        state_dict_1=avg,
+        state_dict_2=model_start,
+        weight_1=1.0,
+        weight_2=weight_start / weight_end,
+        scaling_factor=weight_end,
+    )
+
+    return avg
+
+
+def average_state_dict(
+    state_dict_1: Dict[str, Tensor],
+    state_dict_2: Dict[str, Tensor],
+    weight_1: float,
+    weight_2: float,
+    scaling_factor: float = 1.0,
+) -> Dict[str, Tensor]:
+    """Average two state_dict with given weights:
+    state_dict_1 = (state_dict_1 * weight_1 + state_dict_2 * weight_2)
+      * scaling_factor
+    It is an in-place operation on state_dict_1 itself.
+    """
+    # Identify shared parameters. Two parameters are said to be shared
+    # if they have the same data_ptr
+    uniqued: Dict[int, str] = dict()
+    for k, v in state_dict_1.items():
+        v_data_ptr = v.data_ptr()
+        if v_data_ptr in uniqued:
+            continue
+        uniqued[v_data_ptr] = k
+
+    uniqued_names = list(uniqued.values())
+    for k in uniqued_names:
+        state_dict_1[k] *= weight_1
+        state_dict_1[k] += state_dict_2[k] * weight_2
+        state_dict_1[k] *= scaling_factor