lightning/tests/tests_pytorch/models/test_amp.py

252 lines
8.2 KiB
Python

# Copyright The PyTorch Lightning team.
#
# 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 os
from unittest import mock
import pytest
import torch
from torch import optim
from torch.utils.data import DataLoader
import tests_pytorch.helpers.utils as tutils
from lightning_lite.plugins.environments import SLURMEnvironment
from pytorch_lightning import Trainer
from pytorch_lightning.demos.boring_classes import BoringModel, RandomDataset
from tests_pytorch.helpers.runif import RunIf
class AMPTestModel(BoringModel):
def _step(self, batch):
self._assert_autocast_enabled()
output = self(batch)
is_bfloat16 = self.trainer.precision_plugin.precision == "bf16"
assert output.dtype == torch.float16 if not is_bfloat16 else torch.bfloat16
loss = self.loss(batch, output)
return loss
def loss(self, batch, prediction):
# todo (sean): convert bfloat16 to float32 as mse loss for cpu amp is currently not supported
if self.trainer.precision_plugin.device == "cpu":
prediction = prediction.float()
return super().loss(batch, prediction)
def training_step(self, batch, batch_idx):
output = self._step(batch)
return {"loss": output}
def validation_step(self, batch, batch_idx):
output = self._step(batch)
return {"x": output}
def test_step(self, batch, batch_idx):
output = self._step(batch)
return {"y": output}
def predict_step(self, batch, batch_idx, dataloader_idx=0):
self._assert_autocast_enabled()
output = self(batch)
is_bfloat16 = self.trainer.precision_plugin.precision == "bf16"
assert output.dtype == torch.float16 if not is_bfloat16 else torch.bfloat16
return output
def _assert_autocast_enabled(self):
if self.trainer.precision_plugin.device == "cpu":
assert torch.is_autocast_cpu_enabled()
else:
assert torch.is_autocast_enabled()
@pytest.mark.flaky(reruns=3)
@pytest.mark.parametrize(
("strategy", "precision", "devices"),
(
("single_device", 16, 1),
("single_device", "bf16", 1),
("ddp_spawn", 16, 2),
("ddp_spawn", "bf16", 2),
),
)
def test_amp_cpus(tmpdir, strategy, precision, devices):
"""Make sure combinations of AMP and strategies work if supported."""
trainer = Trainer(
default_root_dir=tmpdir,
accelerator="cpu",
devices=devices,
strategy=strategy,
precision=precision,
max_epochs=1,
limit_train_batches=1,
limit_val_batches=1,
limit_test_batches=1,
limit_predict_batches=1,
logger=False,
enable_checkpointing=False,
enable_model_summary=False,
enable_progress_bar=False,
)
model = AMPTestModel()
trainer.fit(model)
trainer.test(model)
trainer.predict(model)
@RunIf(min_cuda_gpus=2)
@pytest.mark.parametrize("strategy", [None, "dp", "ddp_spawn"])
@pytest.mark.parametrize("precision", [16, pytest.param("bf16", marks=RunIf(bf16_cuda=True))])
@pytest.mark.parametrize("devices", [1, 2])
def test_amp_gpus(tmpdir, strategy, precision, devices):
"""Make sure combinations of AMP and strategies work if supported."""
trainer = Trainer(
default_root_dir=tmpdir,
max_epochs=1,
accelerator="gpu",
devices=devices,
strategy=strategy,
precision=precision,
)
model = AMPTestModel()
trainer.fit(model)
trainer.test(model)
trainer.predict(model, DataLoader(RandomDataset(32, 64)))
@RunIf(min_cuda_gpus=2)
@mock.patch.dict(
os.environ,
{
"SLURM_NTASKS": "1",
"SLURM_NTASKS_PER_NODE": "1",
"SLURM_JOB_NAME": "SOME_NAME",
"SLURM_NODEID": "0",
"LOCAL_RANK": "0",
"SLURM_LOCALID": "0",
"SLURM_PROCID": "0",
},
)
def test_amp_gpu_ddp_slurm_managed(tmpdir):
"""Make sure DDP + AMP work."""
# simulate setting slurm flags
model = AMPTestModel()
# exp file to get meta
logger = tutils.get_default_logger(tmpdir)
# exp file to get weights
checkpoint = tutils.init_checkpoint_callback(logger)
# fit model
trainer = Trainer(
default_root_dir=tmpdir,
max_epochs=1,
accelerator="gpu",
devices=[0],
strategy="ddp_spawn",
precision=16,
callbacks=[checkpoint],
logger=logger,
)
trainer.fit(model)
assert isinstance(trainer.strategy.cluster_environment, SLURMEnvironment)
@mock.patch("pytorch_lightning.plugins.precision.apex_amp.ApexMixedPrecisionPlugin.backward")
def test_amp_without_apex(bwd_mock, tmpdir):
"""Check that even with apex amp type without requesting precision=16 the amp backend is void."""
model = BoringModel()
trainer = Trainer(default_root_dir=tmpdir, amp_backend="native")
assert trainer.amp_backend is None
trainer = Trainer(default_root_dir=tmpdir, max_epochs=1, amp_backend="apex")
assert trainer.amp_backend is None
trainer.fit(model)
assert not bwd_mock.called
@RunIf(min_cuda_gpus=1, amp_apex=True)
@mock.patch("pytorch_lightning.plugins.precision.apex_amp.ApexMixedPrecisionPlugin.backward")
def test_amp_with_apex(bwd_mock, tmpdir):
"""Check calling apex scaling in training."""
class CustomModel(BoringModel):
def training_step(self, batch, batch_idx, optimizer_idx):
return super().training_step(batch, batch_idx)
def configure_optimizers(self):
optimizer1 = optim.Adam(self.parameters(), lr=0.01)
optimizer2 = optim.SGD(self.parameters(), lr=0.01)
lr_scheduler1 = optim.lr_scheduler.StepLR(optimizer1, 1, gamma=0.1)
lr_scheduler2 = optim.lr_scheduler.StepLR(optimizer2, 1, gamma=0.1)
return [optimizer1, optimizer2], [lr_scheduler1, lr_scheduler2]
model = CustomModel()
model.training_epoch_end = None
trainer = Trainer(
default_root_dir=tmpdir, max_steps=5, precision=16, amp_backend="apex", accelerator="gpu", devices=1
)
assert str(trainer.amp_backend) == "AMPType.APEX"
trainer.fit(model)
# `max_steps` is fulfilled in the third batch first optimizer, but we don't check the loop
# `done` condition until all optimizers have run, so the number of backwards is higher than `max_steps`
assert bwd_mock.call_count == 6
assert isinstance(trainer.lr_scheduler_configs[0].scheduler.optimizer, optim.Adam)
assert isinstance(trainer.lr_scheduler_configs[1].scheduler.optimizer, optim.SGD)
@RunIf(min_cuda_gpus=1, amp_apex=True)
def test_amp_with_apex_reload(tmpdir):
model = BoringModel()
trainer = Trainer(
default_root_dir=tmpdir,
max_steps=1,
limit_test_batches=1,
precision=16,
amp_backend="apex",
accelerator="gpu",
devices=1,
)
trainer.fit(model)
trainer.fit_loop.max_steps = 2
with pytest.raises(RuntimeError, match="Resuming training with APEX is currently not supported."):
trainer.fit(model, ckpt_path=trainer.checkpoint_callback.best_model_path)
trainer.test(model, ckpt_path="best")
@pytest.mark.parametrize("clip_val", [0, 10])
@mock.patch("torch.nn.utils.clip_grad_norm_")
def test_precision_16_clip_gradients(mock_clip_grad_norm, clip_val, tmpdir):
"""Ensure that clip gradients is only called if the value is greater than 0."""
model = BoringModel()
trainer = Trainer(
default_root_dir=tmpdir,
enable_progress_bar=False,
max_epochs=1,
devices=1,
precision=16,
limit_train_batches=4,
limit_val_batches=0,
gradient_clip_val=clip_val,
)
trainer.fit(model)
if clip_val > 0:
mock_clip_grad_norm.assert_called()
else:
mock_clip_grad_norm.assert_not_called()