lightning/tests/plugins/test_amp_plugin.py

98 lines
2.5 KiB
Python

import os
from unittest import mock
import pytest
import torch
from pytorch_lightning import Trainer
from pytorch_lightning.plugins import NativeMixedPrecisionPlugin
from tests.helpers.boring_model import BoringModel
from tests.helpers.runif import RunIf
@RunIf(amp_native=True)
@mock.patch.dict(
os.environ, {
"CUDA_VISIBLE_DEVICES": "0,1",
"SLURM_NTASKS": "2",
"SLURM_JOB_NAME": "SOME_NAME",
"SLURM_NODEID": "0",
"LOCAL_RANK": "0",
"SLURM_LOCALID": "0"
}
)
@mock.patch('torch.cuda.device_count', return_value=2)
@pytest.mark.parametrize(
['ddp_backend', 'gpus'],
[('ddp', 2), ('ddp2', 2), ('ddp_spawn', 2)],
)
def test_amp_choice_custom_ddp_cpu(device_count_mock, ddp_backend, gpus):
class MyNativeAMP(NativeMixedPrecisionPlugin):
pass
trainer = Trainer(
precision=16,
amp_backend='native',
accelerator=ddp_backend,
plugins=[MyNativeAMP()],
)
assert isinstance(trainer.precision_plugin, MyNativeAMP)
class GradientUnscaleBoringModel(BoringModel):
def on_after_backward(self):
norm = torch.nn.utils.clip_grad_norm_(self.parameters(), 2)
if not (torch.isinf(norm) or torch.isnan(norm)):
assert norm.item() < 15.
@RunIf(min_gpus=2, amp_native=True)
def test_amp_gradient_unscale(tmpdir):
model = GradientUnscaleBoringModel()
trainer = Trainer(
max_epochs=2,
default_root_dir=os.getcwd(),
limit_train_batches=2,
limit_test_batches=2,
limit_val_batches=2,
amp_backend='native',
accelerator='ddp_spawn',
gpus=2,
precision=16,
track_grad_norm=2,
log_every_n_steps=1,
)
trainer.fit(model)
class UnscaleAccumulateGradBatchesBoringModel(BoringModel):
def on_after_backward(self):
norm = torch.nn.utils.clip_grad_norm_(self.parameters(), 2)
if not (torch.isinf(norm) or torch.isnan(norm)):
assert norm.item() < 15.
@RunIf(min_gpus=2, amp_native=True)
def test_amp_gradient_unscale_accumulate_grad_batches(tmpdir):
model = UnscaleAccumulateGradBatchesBoringModel()
trainer = Trainer(
max_epochs=2,
default_root_dir=os.getcwd(),
limit_train_batches=2,
limit_test_batches=2,
limit_val_batches=2,
amp_backend='native',
accelerator='ddp_spawn',
gpus=2,
precision=16,
track_grad_norm=2,
log_every_n_steps=1,
accumulate_grad_batches=2,
)
trainer.fit(model)