icefall/icefall/checkpoint.py

# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
#
# See ../../LICENSE for clarification regarding multiple authors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.


import glob
import logging
import os
import re
from pathlib import Path
from typing import Any, Dict, List, Optional, Union

import torch
import torch.nn as nn
from lhotse.dataset.sampling.base import CutSampler
from torch.cuda.amp import GradScaler
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.optim import Optimizer
from torch.optim.lr_scheduler import _LRScheduler


def save_checkpoint(
    filename: Path,
    model: Union[nn.Module, DDP],
    params: Optional[Dict[str, Any]] = None,
    optimizer: Optional[Optimizer] = None,
    scheduler: Optional[_LRScheduler] = None,
    scaler: Optional[GradScaler] = None,
    sampler: Optional[CutSampler] = None,
    rank: int = 0,
) -> None:
    """Save training information to a file.

    Args:
      filename:
        The checkpoint filename.
      model:
        The model to be saved. We only save its `state_dict()`.
      params:
        User defined parameters, e.g., epoch, loss.
      optimizer:
        The optimizer to be saved. We only save its `state_dict()`.
      scheduler:
        The scheduler to be saved. We only save its `state_dict()`.
      scalar:
        The GradScaler to be saved. We only save its `state_dict()`.
      rank:
        Used in DDP. We save checkpoint only for the node whose rank is 0.
    Returns:
      Return None.
    """
    if rank != 0:
        return

    logging.info(f"Saving checkpoint to {filename}")

    if isinstance(model, DDP):
        model = model.module

    checkpoint = {
        "model": model.state_dict(),
        "optimizer": optimizer.state_dict() if optimizer is not None else None,
        "scheduler": scheduler.state_dict() if scheduler is not None else None,
        "grad_scaler": scaler.state_dict() if scaler is not None else None,
        "sampler": sampler.state_dict() if sampler is not None else None,
    }

    if params:
        for k, v in params.items():
            assert k not in checkpoint
            checkpoint[k] = v

    torch.save(checkpoint, filename)


def load_checkpoint(
    filename: Path,
    model: nn.Module,
    optimizer: Optional[Optimizer] = None,
    scheduler: Optional[_LRScheduler] = None,
    scaler: Optional[GradScaler] = None,
    sampler: Optional[CutSampler] = None,
    strict: bool = False,
) -> Dict[str, Any]:
    """
    TODO: document it
    """
    logging.info(f"Loading checkpoint from {filename}")
    checkpoint = torch.load(filename, map_location="cpu")

    if next(iter(checkpoint["model"])).startswith("module."):
        logging.info("Loading checkpoint saved by DDP")

        dst_state_dict = model.state_dict()
        src_state_dict = checkpoint["model"]
        for key in dst_state_dict.keys():
            src_key = "{}.{}".format("module", key)
            dst_state_dict[key] = src_state_dict.pop(src_key)
        assert len(src_state_dict) == 0
        model.load_state_dict(dst_state_dict, strict=strict)
    else:
        model.load_state_dict(checkpoint["model"], strict=strict)

    checkpoint.pop("model")

    def load(name, obj):
        s = checkpoint.get(name, None)
        if obj and s:
            obj.load_state_dict(s)
            checkpoint.pop(name)

    load("optimizer", optimizer)
    load("scheduler", scheduler)
    load("grad_scaler", scaler)
    load("sampler", sampler)

    return checkpoint


def average_checkpoints(
    filenames: List[Path], device: torch.device = torch.device("cpu")
) -> dict:
    """Average a list of checkpoints.

    Args:
      filenames:
        Filenames of the checkpoints to be averaged. We assume all
        checkpoints are saved by :func:`save_checkpoint`.
      device:
        Move checkpoints to this device before averaging.
    Returns:
      Return a dict (i.e., state_dict) which is the average of all
      model state dicts contained in the checkpoints.
    """
    n = len(filenames)

    avg = torch.load(filenames[0], map_location=device)["model"]
    for i in range(1, n):
        state_dict = torch.load(filenames[i], map_location=device)["model"]
        for k in avg:
            avg[k] += state_dict[k]

    for k in avg:
        if avg[k].is_floating_point():
            avg[k] /= n
        else:
            avg[k] //= n

    return avg


def save_checkpoint_with_global_batch_idx(
    out_dir: Path,
    global_batch_idx: int,
    model: Union[nn.Module, DDP],
    params: Optional[Dict[str, Any]] = None,
    optimizer: Optional[Optimizer] = None,
    scheduler: Optional[_LRScheduler] = None,
    scaler: Optional[GradScaler] = None,
    sampler: Optional[CutSampler] = None,
    rank: int = 0,
):
    """Save training info after processing given number of batches.

    Args:
      out_dir:
        The directory to save the checkpoint.
      global_batch_idx:
        The number of batches processed so far from the very start of the
        training. The saved checkpoint will have the following filename:

            f'out_dir / checkpoint-{global_batch_idx}.pt'
      model:
        The neural network model whose `state_dict` will be saved in the
        checkpoint.
      params:
        A dict of training configurations to be saved.
      optimizer:
        The optimizer used in the training. Its `state_dict` will be saved.
      scheduler:
        The learning rate scheduler used in the training. Its `state_dict` will
        be saved.
      scaler:
        The scaler used for mix precision training. Its `state_dict` will
        be saved.
      sampler:
        The sampler used in the training dataset.
      rank:
        The rank ID used in DDP training of the current node. Set it to 0
        if DDP is not used.
    """
    out_dir = Path(out_dir)
    out_dir.mkdir(parents=True, exist_ok=True)
    filename = out_dir / f"checkpoint-{global_batch_idx}.pt"
    save_checkpoint(
        filename=filename,
        model=model,
        params=params,
        optimizer=optimizer,
        scheduler=scheduler,
        scaler=scaler,
        sampler=sampler,
        rank=rank,
    )


def find_checkpoints(out_dir: Path) -> List[str]:
    """Find all available checkpoints in a directory.

    The checkpoint filenames have the form: `checkpoint-xxx.pt`
    where xxx is a numerical value.

    Args:
      out_dir:
        The directory where to search for checkpoints.
    Returns:
      Return a list of checkpoint filenames, sorted in descending
      order by the numerical value in the filename.
    """
    checkpoints = list(glob.glob(f"{out_dir}/checkpoint-[0-9]*.pt"))
    pattern = re.compile(r"checkpoint-([0-9]+).pt")
    idx_checkpoints = [
        (int(pattern.search(c).group(1)), c) for c in checkpoints
    ]

    idx_checkpoints = sorted(idx_checkpoints, reverse=True, key=lambda x: x[0])
    ans = [ic[1] for ic in idx_checkpoints]
    return ans


def remove_checkpoints(
    out_dir: Path,
    topk: int,
    rank: int = 0,
):
    """Remove checkpoints from the given directory.

    We assume that checkpoint filename has the form `checkpoint-xxx.pt`
    where xxx is a number, representing the number of processed batches
    when saving that checkpoint. We sort checkpoints by filename and keep
    only the `topk` checkpoints with the highest `xxx`.

    Args:
      out_dir:
        The directory containing checkpoints to be removed.
      topk:
        Number of checkpoints to keep.
      rank:
        If using DDP for training, it is the rank of the current node.
        Use 0 if no DDP is used for training.
    """
    assert topk >= 1, topk
    if rank != 0:
        return
    checkpoints = find_checkpoints(out_dir)

    if len(checkpoints) == 0:
        logging.warn(f"No checkpoints found in {out_dir}")
        return

    if len(checkpoints) <= topk:
        return

    to_remove = checkpoints[topk:]
    for c in to_remove:
        os.remove(c)