update

嘉渊

2023-04-23 887039e9d335f9964ebacb14b0205ee891e6819b

update

 funasr/build_utils/build_trainer.py

@@ -3,13 +3,15 @@

"""Trainer module."""
import argparse
from contextlib import contextmanager
import dataclasses
import logging
import os
import time
from contextlib import contextmanager
from dataclasses import is_dataclass
from distutils.version import LooseVersion
import logging
from io import BytesIO
from pathlib import Path
import time
from typing import Dict
from typing import Iterable
from typing import List
@@ -20,9 +22,6 @@

import humanfriendly
import oss2
from io import BytesIO
import os
import numpy as np
import torch
import torch.nn
import torch.optim
@@ -30,7 +29,7 @@

from funasr.iterators.abs_iter_factory import AbsIterFactory
from funasr.main_funcs.average_nbest_models import average_nbest_models
from funasr.main_funcs.calculate_all_attentions import calculate_all_attentions
from funasr.models.base_model import FunASRModel
from funasr.schedulers.abs_scheduler import AbsBatchStepScheduler
from funasr.schedulers.abs_scheduler import AbsEpochStepScheduler
from funasr.schedulers.abs_scheduler import AbsScheduler
@@ -39,7 +38,6 @@
from funasr.torch_utils.device_funcs import to_device
from funasr.torch_utils.recursive_op import recursive_average
from funasr.torch_utils.set_all_random_seed import set_all_random_seed
from funasr.models.base_model import FunASRModel
from funasr.train.distributed_utils import DistributedOption
from funasr.train.reporter import Reporter
from funasr.train.reporter import SubReporter
@@ -95,37 +93,32 @@
    wandb_model_log_interval: int
    use_pai: bool
    oss_bucket: Union[oss2.Bucket, None]
    batch_interval: int


class Trainer:
    """Trainer having a optimizer.

    If you'd like to use multiple optimizers, then inherit this class
    and override the methods if necessary - at least "train_one_epoch()"

    >>> class TwoOptimizerTrainer(Trainer):
    ...     @classmethod
    ...     def add_arguments(cls, parser):
    ...         ...
    ...
    ...     @classmethod
    ...     def train_one_epoch(cls, model, optimizers, ...):
    ...         loss1 = model.model1(...)
    ...         loss1.backward()
    ...         optimizers[0].step()
    ...
    ...         loss2 = model.model2(...)
    ...         loss2.backward()
    ...         optimizers[1].step()
    """Trainer

    """

    def __init__(self):
        raise RuntimeError("This class can't be instantiated.")
    def __init__(self,
                 args,
                 model: FunASRModel,
                 optimizers: Sequence[torch.optim.Optimizer],
                 schedulers: Sequence[Optional[AbsScheduler]],
                 train_dataloader: AbsIterFactory,
                 valid_dataloader: AbsIterFactory,
                 trainer_options,
                 distributed_option: DistributedOption):
        self.trainer_options = self.build_options(args)
        self.model = model
        self.optimizers = optimizers
        self.schedulers = schedulers
        self.train_dataloader = train_dataloader
        self.valid_dataloader = valid_dataloader
        self.trainer_options = trainer_options
        self.distributed_option = distributed_option

    @classmethod
    def build_options(cls, args: argparse.Namespace) -> TrainerOptions:
    def build_options(self, args: argparse.Namespace) -> TrainerOptions:
        """Build options consumed by train(), eval()"""
        assert check_argument_types()
        return build_dataclass(TrainerOptions, args)
@@ -164,20 +157,17 @@

        logging.info(f"The training was resumed using {checkpoint}")

    @classmethod
    def run(
            cls,
            model: FunASRModel,
            optimizers: Sequence[torch.optim.Optimizer],
            schedulers: Sequence[Optional[AbsScheduler]],
            train_iter_factory: AbsIterFactory,
            valid_iter_factory: AbsIterFactory,
            trainer_options,
            distributed_option: DistributedOption,
    ) -> None:
    def run(self) -> None:
        """Perform training. This method performs the main process of training."""
        assert check_argument_types()
        # NOTE(kamo): Don't check the type more strictly as far trainer_options
        model = self.model
        optimizers = self.optimizers
        schedulers = self.schedulers
        train_dataloader = self.train_dataloader
        valid_dataloader = self.valid_dataloader
        trainer_options = self.trainer_options
        distributed_option = self.distributed_option
        assert is_dataclass(trainer_options), type(trainer_options)
        assert len(optimizers) == len(schedulers), (len(optimizers), len(schedulers))

@@ -188,9 +178,6 @@
                logging.warning("No keep_nbest_models is given. Change to [1]")
                trainer_options.keep_nbest_models = [1]
            keep_nbest_models = trainer_options.keep_nbest_models

        # assert batch_interval is set and >0
        assert trainer_options.batch_interval > 0

        output_dir = Path(trainer_options.output_dir)
        reporter = Reporter()
@@ -211,7 +198,7 @@
            scaler = None

        if trainer_options.resume and (output_dir / "checkpoint.pb").exists():
            cls.resume(
            self.resume(
                checkpoint=output_dir / "checkpoint.pb",
                model=model,
                optimizers=optimizers,
@@ -228,14 +215,8 @@
            )

        if distributed_option.distributed:
            if trainer_options.sharded_ddp:
                dp_model = fairscale.nn.data_parallel.ShardedDataParallel(
                    module=model,
                    sharded_optimizer=optimizers,
                )
            else:
                dp_model = torch.nn.parallel.DistributedDataParallel(
                    model, find_unused_parameters=trainer_options.unused_parameters)
            dp_model = torch.nn.parallel.DistributedDataParallel(
                model, find_unused_parameters=trainer_options.unused_parameters)
        elif distributed_option.ngpu > 1:
            dp_model = torch.nn.parallel.DataParallel(
                model,
@@ -288,11 +269,11 @@
            reporter.set_epoch(iepoch)
            # 1. Train and validation for one-epoch
            with reporter.observe("train") as sub_reporter:
                all_steps_are_invalid, max_update_stop = cls.train_one_epoch(
                all_steps_are_invalid, max_update_stop = self.train_one_epoch(
                    model=dp_model,
                    optimizers=optimizers,
                    schedulers=schedulers,
                    iterator=train_iter_factory.build_iter(iepoch),
                    iterator=train_dataloader.build_iter(iepoch),
                    reporter=sub_reporter,
                    scaler=scaler,
                    summary_writer=train_summary_writer,
@@ -301,9 +282,9 @@
                )

            with reporter.observe("valid") as sub_reporter:
                cls.validate_one_epoch(
                self.validate_one_epoch(
                    model=dp_model,
                    iterator=valid_iter_factory.build_iter(iepoch),
                    iterator=valid_dataloader.build_iter(iepoch),
                    reporter=sub_reporter,
                    options=trainer_options,
                    distributed_option=distributed_option,
@@ -532,9 +513,8 @@
                pai_output_dir=trainer_options.output_dir,
            )

    @classmethod
    def train_one_epoch(
            cls,
            self,
            model: torch.nn.Module,
            iterator: Iterable[Tuple[List[str], Dict[str, torch.Tensor]]],
            optimizers: Sequence[torch.optim.Optimizer],
@@ -570,31 +550,11 @@
        # processes, send stop-flag to the other processes if iterator is finished
        iterator_stop = torch.tensor(0).to("cuda" if ngpu > 0 else "cpu")

        # get the rank
        rank = distributed_option.dist_rank
        # get the num batch updates
        num_batch_updates = 0
        # ouput dir
        output_dir = Path(options.output_dir)
        # batch interval
        batch_interval = options.batch_interval
        assert batch_interval > 0

        start_time = time.perf_counter()
        for iiter, (_, batch) in enumerate(
                reporter.measure_iter_time(iterator, "iter_time"), 1
        ):
            assert isinstance(batch, dict), type(batch)

            if rank == 0:
                if hasattr(model, "num_updates") or (hasattr(model, "module") and hasattr(model.module, "num_updates")):
                    num_batch_updates = model.get_num_updates() if hasattr(model,
                                                                           "num_updates") else model.module.get_num_updates()
                if (num_batch_updates % batch_interval == 0) and (options.oss_bucket is not None) and options.use_pai:
                    buffer = BytesIO()
                    torch.save(model.state_dict(), buffer)
                    options.oss_bucket.put_object(os.path.join(output_dir, f"{num_batch_updates}batch.pth"),
                                                  buffer.getvalue())

            if distributed:
                torch.distributed.all_reduce(iterator_stop, ReduceOp.SUM)
@@ -792,10 +752,9 @@
                torch.distributed.all_reduce(iterator_stop, ReduceOp.SUM)
        return all_steps_are_invalid, max_update_stop

    @classmethod
    @torch.no_grad()
    def validate_one_epoch(
            cls,
            self,
            model: torch.nn.Module,
            iterator: Iterable[Dict[str, torch.Tensor]],
            reporter: SubReporter,