初步修正了多卡情况下 evaluate_dataloader 使用定制 batch_sampler 会报错的问题,改为给出 warning

This commit is contained in:
x54-729 2022-10-12 14:15:37 +08:00
parent fe30d02f86
commit 28db704f70
3 changed files with 118 additions and 7 deletions

View File

@ -609,9 +609,15 @@ class TorchDDPDriver(TorchDriver):
# evaluator
elif dist == "unrepeatdist":
args = self.get_dataloader_args(dataloader)
if type(args.batch_sampler) != BatchSampler:
# TODO 这里的目的是判断用户的 batch_sampler 是定制的,可能需要完善
logger.warning("Note that you are using customized ``batch_sampler`` in evaluate dataloader or" \
"train dataloader while testing ``overfit_batches``, which may cause that" \
"the data for distributed evaluation is not unrepeated.")
if isinstance(args.sampler, ReproducibleSampler):
sampler = conversion_between_reproducible_and_unrepeated_sampler(args.sampler)
elif not isinstance(args.sampler, UnrepeatedSampler):
# TODO 避开 batch_sampler 的情况
_check_dataloader_args_for_distributed(args, controller='Evaluator')
sampler = UnrepeatedSequentialSampler(
dataset=args.dataset
@ -622,6 +628,7 @@ class TorchDDPDriver(TorchDriver):
num_replicas=self.world_size,
rank=self.global_rank
)
# TODO 这里暂时统一替换为 BatchSampler
batch_sampler = BatchSampler(sampler, args.batch_size, drop_last=False)
return replace_batch_sampler(dataloader, batch_sampler)
else:

View File

@ -14,7 +14,7 @@ from fastNLP.envs import (
FASTNLP_BACKEND_LAUNCH,
FASTNLP_GLOBAL_SEED,
)
from fastNLP.core.samplers import re_instantiate_sampler, ReproducibleBatchSampler
from fastNLP.core.samplers import re_instantiate_sampler, ReproducibleBatchSampler, ReproducibleSampler
from fastNLP.core.utils import auto_param_call, apply_to_collection
from fastNLP.core.log import logger
@ -308,15 +308,26 @@ def optimizer_state_to_device(state, device):
def _check_dataloader_args_for_distributed(args, controller='Trainer'):
if type(args.batch_sampler) is not TorchBatchSampler or (type(args.sampler) not in {TorchRandomSampler,
TorchSequentialSampler}):
mode = 'training' if controller == 'Trainer' else 'evaluation'
substitution = 'fastNLP.RandomSampler' if controller == 'Trainer' else 'fastNLP.UnrepeatedSequentialSampler'
"""
检查 dataloader sampler 情况如果用户替换了自己定制的 sampler 为了防止
在分布式训练中出现错误会报错
"""
error_flag = (type(args.sampler) not in {TorchRandomSampler, TorchSequentialSampler})
if controller == 'Trainer':
mode = 'training'
substitution = 'fastNLP.RandomSampler'
error_flag = (type(args.batch_sampler) != TorchBatchSampler) or error_flag
else: # Evaluator
mode = 'evaluation'
substitution = 'fastNLP.UnrepeatedSequentialSampler'
if error_flag:
raise TypeError(f"Using customized ``batch_sampler`` or ``sampler`` for distributed {mode} may cause "
f"unpredictable problems, because fastNLP will substitute the dataloader's sampler into "
f"``{substitution}``. The customized sampler should set for distributed running "
f"before initializing ``{controller}`` , and then set the "
f"parameter ``use_dist_sampler`` of ``{controller}`` to ``False``.")
f"parameter ``use_dist_sampler`` of ``{controller}`` to ``False``."
f"\n Current batch_sampler: {type(args.batch_sampler)}"
f"\n Current sampler: {type(args.sampler)}")
def _create_default_config(
zero_optimization: bool = True,

View File

@ -7,6 +7,7 @@ from dataclasses import dataclass
from typing import Any
from fastNLP.core.controllers.trainer import Trainer
from fastNLP.core.samplers import BucketedBatchSampler, RandomSampler
from tests.helpers.models.torch_model import TorchNormalModel_Classification_1
from tests.helpers.datasets.torch_data import TorchNormalDataset_Classification, TorchArgMaxDataset
from tests.helpers.callbacks.helper_callbacks import RecordLossCallback, RecordMetricCallback
@ -378,4 +379,96 @@ def test_trainer_w_evaluator_overfit_torch(
trainer.run(num_train_batch_per_epoch=num_train_batch_per_epoch)
if dist.is_initialized():
dist.destroy_process_group()
dist.destroy_process_group()
@pytest.mark.torch
@pytest.mark.parametrize("driver,device", [("torch", 1), ("torch", [0, 1])]) # ("torch", [0, 1]),("torch", 1)
@pytest.mark.parametrize("train_sampler", ["batch_sampler", "sampler"])
@pytest.mark.parametrize("eval_sampler", ["batch_sampler", "sampler"])
@pytest.mark.parametrize("overfit_batches", [-1, 0])
@magic_argv_env_context
def test_trainer_w_evaluator_w_samplers(
driver,
device,
train_sampler,
eval_sampler,
overfit_batches,
):
"""
测试使用 dataloader 时使用了定制 batch_sampler sampler 且合法的情况
"""
model = TorchNormalModel_Classification_1(
num_labels=ArgMaxDatasetConfig.num_labels,
feature_dimension=ArgMaxDatasetConfig.feature_dimension
)
optimizers = SGD(model.parameters(), lr=0.001)
metrics = {"acc": Accuracy()}
dataset = TorchArgMaxDataset(
feature_dimension=ArgMaxDatasetConfig.feature_dimension,
data_num=ArgMaxDatasetConfig.data_num,
seed=ArgMaxDatasetConfig.seed
)
if train_sampler == "batch_sampler":
train_dataloader = DataLoader(
dataset=dataset,
batch_sampler=BucketedBatchSampler(
dataset,[3] * len(dataset), ArgMaxDatasetConfig.batch_size
)
)
elif train_sampler == "sampler":
train_dataloader = DataLoader(
dataset=dataset,
batch_size=ArgMaxDatasetConfig.batch_size,
sampler=RandomSampler(dataset)
)
else:
train_dataloader = DataLoader(
dataset=dataset,
batch_size=ArgMaxDatasetConfig.batch_size,
shuffle=True,
)
if eval_sampler == "batch_sampler":
eval_dataloader = DataLoader(
dataset=dataset,
batch_sampler=BucketedBatchSampler(
dataset,[3] * len(dataset), ArgMaxDatasetConfig.batch_size
)
)
elif eval_sampler == "sampler":
eval_dataloader = DataLoader(
dataset=dataset,
sampler=RandomSampler(dataset)
)
else:
DataLoader(
dataset=dataset,
batch_size=ArgMaxDatasetConfig.batch_size,
shuffle=True,
)
trainer = Trainer(
model=model,
driver=driver,
device=device,
overfit_batches=overfit_batches,
optimizers=optimizers,
train_dataloader=train_dataloader,
evaluate_dataloaders={"dl": eval_dataloader},
metrics=metrics,
n_epochs=2,
output_from_new_proc="all",
evaluate_every=-1,
torch_kwargs={
"non_blocking": False,
"set_grad_to_none": True
}
)
trainer.run()
if dist.is_initialized():
dist.destroy_process_group()