def test_amp_gpu_ddp_slurm_managed(tmpdir):
"""Make sure DDP + AMP work."""
# simulate setting slurm flags
tutils.set_random_master_port()
os.environ['SLURM_LOCALID'] = str(0)
model = EvalModelTemplate()
# 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,
gpus=[0],
distributed_backend='ddp_spawn',
precision=16,
checkpoint_callback=checkpoint,
logger=logger,
)
trainer.is_slurm_managing_tasks = True
result = trainer.fit(model)
# correct result and ok accuracy
assert result == 1, 'amp + ddp model failed to complete'
# test root model address
assert trainer.slurm_connector.resolve_root_node_address('abc') == 'abc'
assert trainer.slurm_connector.resolve_root_node_address('abc[23]') == 'abc23'
assert trainer.slurm_connector.resolve_root_node_address('abc[23-24]') == 'abc23'
assert trainer.slurm_connector.resolve_root_node_address('abc[23-24, 45-40, 40]') == 'abc23'
def test_running_test_pretrained_model_cpu(tmpdir):
"""Verify test() on pretrained model."""
model = EvalModelTemplate()
# logger file to get meta
logger = tutils.get_default_logger(tmpdir)
# logger file to get weights
checkpoint = tutils.init_checkpoint_callback(logger)
trainer_options = dict(
progress_bar_refresh_rate=0,
max_epochs=3,
limit_train_batches=0.4,
limit_val_batches=0.2,
checkpoint_callback=checkpoint,
logger=logger,
default_root_dir=tmpdir,
)
# fit model
trainer = Trainer(**trainer_options)
result = trainer.fit(model)
# correct result and ok accuracy
assert result == 1, 'training failed to complete'
pretrained_model = EvalModelTemplate.load_from_checkpoint(
trainer.checkpoint_callback.best_model_path)
new_trainer = Trainer(**trainer_options)
new_trainer.test(pretrained_model)
# test we have good test accuracy
tutils.assert_ok_model_acc(new_trainer)
def test_running_test_no_val(tmpdir):
"""Verify `test()` works on a model with no `val_loader`."""
model = EvalModelTemplate()
# logger file to get meta
logger = tutils.get_default_logger(tmpdir)
# logger file to get weights
checkpoint = tutils.init_checkpoint_callback(logger)
# fit model
trainer = Trainer(
default_root_dir=tmpdir,
progress_bar_refresh_rate=0,
max_epochs=1,
limit_train_batches=0.4,
limit_val_batches=0.2,
limit_test_batches=0.2,
checkpoint_callback=checkpoint,
logger=logger,
early_stop_callback=False,
)
result = trainer.fit(model)
assert result == 1, 'training failed to complete'
trainer.test()
# test we have good test accuracy
tutils.assert_ok_model_acc(trainer)
def test_running_test_after_fitting(tmpdir):
"""Verify test() on fitted model."""
model = EvalModelTemplate()
# logger file to get meta
logger = tutils.get_default_logger(tmpdir)
# logger file to get weights
checkpoint = tutils.init_checkpoint_callback(logger)
# fit model
trainer = Trainer(
default_root_dir=tmpdir,
progress_bar_refresh_rate=0,
max_epochs=2,
limit_train_batches=0.4,
limit_val_batches=0.2,
limit_test_batches=0.2,
checkpoint_callback=checkpoint,
logger=logger,
)
result = trainer.fit(model)
assert result == 1, 'training failed to complete'
trainer.test()
# test we have good test accuracy
tutils.assert_ok_model_acc(trainer, thr=0.5)
def test_running_test_pretrained_model_distrib_ddp_spawn(tmpdir):
"""Verify `test()` on pretrained model."""
tutils.set_random_master_port()
model = EvalModelTemplate()
# exp file to get meta
logger = tutils.get_default_logger(tmpdir)
# exp file to get weights
checkpoint = tutils.init_checkpoint_callback(logger)
trainer_options = dict(
progress_bar_refresh_rate=0,
max_epochs=2,
limit_train_batches=0.4,
limit_val_batches=0.2,
checkpoint_callback=checkpoint,
logger=logger,
gpus=[0, 1],
distributed_backend='ddp_spawn',
default_root_dir=tmpdir,
)
# fit model
trainer = Trainer(**trainer_options)
result = trainer.fit(model)
log.info(os.listdir(tutils.get_data_path(logger, path_dir=tmpdir)))
# correct result and ok accuracy
assert result == 1, 'training failed to complete'
pretrained_model = EvalModelTemplate.load_from_checkpoint(
trainer.checkpoint_callback.best_model_path)
# run test set
new_trainer = Trainer(**trainer_options)
results = new_trainer.test(pretrained_model)
pretrained_model.cpu()
acc = results[0]['test_acc']
assert acc > 0.5, f"Model failed to get expected {0.5} accuracy. test_acc = {acc}"
dataloaders = model.test_dataloader()
if not isinstance(dataloaders, list):
dataloaders = [dataloaders]
for dataloader in dataloaders:
tpipes.run_prediction(dataloader, pretrained_model)
def test_running_test_pretrained_model_distrib(tmpdir, backend):
"""Verify `test()` on pretrained model."""
tutils.set_random_master_port()
model = EvalModelTemplate()
# exp file to get meta
logger = tutils.get_default_logger(tmpdir)
# exp file to get weights
checkpoint = tutils.init_checkpoint_callback(logger)
trainer_options = dict(
progress_bar_refresh_rate=0,
max_epochs=2,
limit_train_batches=0.4,
limit_val_batches=0.2,
checkpoint_callback=checkpoint,
logger=logger,
gpus=[0, 1],
distributed_backend=backend,
)
# fit model
trainer = Trainer(**trainer_options)
result = trainer.fit(model)
log.info(os.listdir(tutils.get_data_path(logger, path_dir=tmpdir)))
# correct result and ok accuracy
assert result == 1, 'training failed to complete'
pretrained_model = tutils.load_model_from_checkpoint(
logger,
trainer.checkpoint_callback.dirpath,
module_class=EvalModelTemplate)
# run test set
new_trainer = Trainer(**trainer_options)
new_trainer.test(pretrained_model)
# test we have good test accuracy
tutils.assert_ok_model_acc(new_trainer)
dataloaders = model.test_dataloader()
if not isinstance(dataloaders, list):
dataloaders = [dataloaders]
for dataloader in dataloaders:
tpipes.run_prediction(dataloader, pretrained_model)
def run_model_test(trainer_options,
model,
on_gpu: bool = True,
version=None,
with_hpc: bool = True):
reset_seed()
save_dir = trainer_options['default_root_dir']
# logger file to get meta
logger = get_default_logger(save_dir, version=version)
trainer_options.update(logger=logger)
if 'checkpoint_callback' not in trainer_options:
# logger file to get weights
checkpoint = init_checkpoint_callback(logger)
trainer_options.update(checkpoint_callback=checkpoint)
# fit model
trainer = Trainer(**trainer_options)
result = trainer.fit(model)
# correct result and ok accuracy
assert result == 1, 'amp + ddp model failed to complete'
# test model loading
pretrained_model = load_model_from_checkpoint(
logger, trainer.checkpoint_callback.dirpath)
# test new model accuracy
test_loaders = model.test_dataloader()
if not isinstance(test_loaders, list):
test_loaders = [test_loaders]
[
run_prediction(dataloader, pretrained_model)
for dataloader in test_loaders
]
if with_hpc:
if trainer.use_ddp or trainer.use_ddp2:
# on hpc this would work fine... but need to hack it for the purpose of the test
trainer.model = pretrained_model
trainer.optimizers, trainer.lr_schedulers, trainer.optimizer_frequencies = \
trainer.init_optimizers(pretrained_model)
# test HPC loading / saving
trainer.hpc_save(save_dir, logger)
trainer.hpc_load(save_dir, on_gpu=on_gpu)
def test_dp_resume(tmpdir):
"""Make sure DP continues training correctly."""
hparams = EvalModelTemplate.get_default_hparams()
model = EvalModelTemplate(**hparams)
trainer_options = dict(
max_epochs=1,
gpus=2,
distributed_backend='dp',
default_root_dir=tmpdir,
)
# get logger
logger = tutils.get_default_logger(tmpdir)
# exp file to get weights
# logger file to get weights
checkpoint = tutils.init_checkpoint_callback(logger)
# add these to the trainer options
trainer_options['logger'] = logger
trainer_options['checkpoint_callback'] = checkpoint
# fit model
trainer = Trainer(**trainer_options)
trainer.is_slurm_managing_tasks = True
result = trainer.fit(model)
# track epoch before saving. Increment since we finished the current epoch, don't want to rerun
real_global_epoch = trainer.current_epoch + 1
# correct result and ok accuracy
assert result == 1, 'amp + dp model failed to complete'
# ---------------------------
# HPC LOAD/SAVE
# ---------------------------
# save
trainer.hpc_save(tmpdir, logger)
# init new trainer
new_logger = tutils.get_default_logger(tmpdir, version=logger.version)
trainer_options['logger'] = new_logger
trainer_options['checkpoint_callback'] = ModelCheckpoint(tmpdir)
trainer_options['limit_train_batches'] = 0.5
trainer_options['limit_val_batches'] = 0.2
trainer_options['max_epochs'] = 1
new_trainer = Trainer(**trainer_options)
# set the epoch start hook so we can predict before the model does the full training
def assert_good_acc():
assert new_trainer.current_epoch == real_global_epoch and new_trainer.current_epoch > 0
# if model and state loaded correctly, predictions will be good even though we
# haven't trained with the new loaded model
dp_model = new_trainer.model
dp_model.eval()
dataloader = trainer.train_dataloader
tpipes.run_prediction(dataloader, dp_model, dp=True)
# new model
model = EvalModelTemplate(**hparams)
model.on_train_start = assert_good_acc
# fit new model which should load hpc weights
new_trainer.fit(model)
# test freeze on gpu
model.freeze()
model.unfreeze()
