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_no_val_end_module(monkeypatch, tmpdir, tmpdir_server, url_ckpt):
"""Tests use case where trainer saves the model, and user loads it from tags independently."""
# set $TORCH_HOME, which determines torch hub's cache path, to tmpdir
monkeypatch.setenv('TORCH_HOME', tmpdir)
model = EvalModelTemplate()
# logger file to get meta
logger = tutils.get_default_logger(tmpdir)
# fit model
trainer = Trainer(
default_root_dir=tmpdir,
max_epochs=1,
logger=logger,
checkpoint_callback=ModelCheckpoint(tmpdir),
)
result = trainer.fit(model)
# traning complete
assert result == 1, 'amp + ddp model failed to complete'
# save model
new_weights_path = os.path.join(tmpdir, 'save_test.ckpt')
trainer.save_checkpoint(new_weights_path)
# load new model
hparams_path = tutils.get_data_path(logger, path_dir=tmpdir)
hparams_path = os.path.join(hparams_path, 'hparams.yaml')
ckpt_path = f'http://{tmpdir_server[0]}:{tmpdir_server[1]}/{os.path.basename(new_weights_path)}' if url_ckpt else new_weights_path
model_2 = EvalModelTemplate.load_from_checkpoint(checkpoint_path=ckpt_path,
hparams_file=hparams_path)
model_2.eval()
def test_loading_meta_tags(tmpdir):
""" test for backward compatibility to meta_tags.csv """
tutils.reset_seed()
hparams = EvalModelTemplate.get_default_hparams()
# save tags
logger = tutils.get_default_logger(tmpdir)
logger.log_hyperparams(Namespace(some_str='a_str', an_int=1, a_float=2.0))
logger.log_hyperparams(hparams)
logger.save()
# load hparams
path_expt_dir = tutils.get_data_path(logger, path_dir=tmpdir)
hparams_path = os.path.join(path_expt_dir,
TensorBoardLogger.NAME_HPARAMS_FILE)
hparams = load_hparams_from_yaml(hparams_path)
# save as legacy meta_tags.csv
tags_path = os.path.join(path_expt_dir, 'meta_tags.csv')
save_hparams_to_tags_csv(tags_path, hparams)
tags = load_hparams_from_tags_csv(tags_path)
assert hparams == tags
def test_strict_model_load(monkeypatch, tmpdir, tmpdir_server, url_ckpt):
"""Tests use case where trainer saves the model, and user loads it from tags independently."""
# set $TORCH_HOME, which determines torch hub's cache path, to tmpdir
monkeypatch.setenv("TORCH_HOME", tmpdir)
model = EvalModelTemplate()
# Extra layer
model.c_d3 = torch.nn.Linear(model.hidden_dim, model.hidden_dim)
# logger file to get meta
logger = tutils.get_default_logger(tmpdir)
# fit model
trainer = Trainer(
default_root_dir=tmpdir,
max_epochs=1,
logger=logger,
checkpoint_callback=ModelCheckpoint(dirpath=tmpdir),
)
result = trainer.fit(model)
# traning complete
assert result == 1
# save model
new_weights_path = os.path.join(tmpdir, "save_test.ckpt")
trainer.save_checkpoint(new_weights_path)
# load new model
hparams_path = tutils.get_data_path(logger, path_dir=tmpdir)
hparams_path = os.path.join(hparams_path, "hparams.yaml")
ckpt_path = (
f"http://{tmpdir_server[0]}:{tmpdir_server[1]}/{os.path.basename(new_weights_path)}"
if url_ckpt
else new_weights_path
)
try:
EvalModelTemplate.load_from_checkpoint(
checkpoint_path=ckpt_path,
hparams_file=hparams_path,
)
except Exception:
failed = True
else:
failed = False
assert failed, "Model should not been loaded since the extra layer added."
failed = False
try:
EvalModelTemplate.load_from_checkpoint(
checkpoint_path=ckpt_path,
hparams_file=hparams_path,
strict=False,
)
except Exception:
failed = True
assert not failed, "Model should be loaded due to strict=False."
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_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_strict_model_load_less_params(monkeypatch, tmpdir, tmpdir_server,
url_ckpt):
"""Tests use case where trainer saves the model, and user loads it from tags independently."""
# set $TORCH_HOME, which determines torch hub's cache path, to tmpdir
monkeypatch.setenv('TORCH_HOME', tmpdir)
model = EvalModelTemplate()
# logger file to get meta
logger = tutils.get_default_logger(tmpdir)
# fit model
trainer = Trainer(
default_root_dir=tmpdir,
max_epochs=1,
logger=logger,
checkpoint_callback=ModelCheckpoint(tmpdir),
)
result = trainer.fit(model)
# traning complete
assert result == 1
# save model
new_weights_path = os.path.join(tmpdir, 'save_test.ckpt')
trainer.save_checkpoint(new_weights_path)
# load new model
hparams_path = os.path.join(tutils.get_data_path(logger, path_dir=tmpdir),
'hparams.yaml')
hparams_url = f'http://{tmpdir_server[0]}:{tmpdir_server[1]}/{os.path.basename(new_weights_path)}'
ckpt_path = hparams_url if url_ckpt else new_weights_path
class CurrentModel(EvalModelTemplate):
def __init__(self):
super().__init__()
self.c_d3 = torch.nn.Linear(7, 7)
CurrentModel.load_from_checkpoint(
checkpoint_path=ckpt_path,
hparams_file=hparams_path,
strict=False,
)
with pytest.raises(
RuntimeError,
match=r'Missing key\(s\) in state_dict: "c_d3.weight", "c_d3.bias"'
):
CurrentModel.load_from_checkpoint(
checkpoint_path=ckpt_path,
hparams_file=hparams_path,
strict=True,
)
def test_model_saving_loading(tmpdir):
"""Tests use case where trainer saves the model, and user loads it from tags independently."""
model = EvalModelTemplate()
# logger file to get meta
logger = tutils.get_default_logger(tmpdir)
# fit model
trainer = Trainer(
max_epochs=1,
logger=logger,
checkpoint_callback=ModelCheckpoint(tmpdir),
default_root_dir=tmpdir,
)
result = trainer.fit(model)
# traning complete
assert result == 1, 'amp + ddp model failed to complete'
# make a prediction
dataloaders = model.test_dataloader()
if not isinstance(dataloaders, list):
dataloaders = [dataloaders]
for dataloader in dataloaders:
for batch in dataloader:
break
x, y = batch
x = x.view(x.size(0), -1)
# generate preds before saving model
model.eval()
pred_before_saving = model(x)
# save model
new_weights_path = os.path.join(tmpdir, 'save_test.ckpt')
trainer.save_checkpoint(new_weights_path)
# load new model
hparams_path = tutils.get_data_path(logger, path_dir=tmpdir)
hparams_path = os.path.join(hparams_path, 'hparams.yaml')
model_2 = EvalModelTemplate.load_from_checkpoint(
checkpoint_path=new_weights_path,
hparams_file=hparams_path,
)
model_2.eval()
# make prediction
# assert that both predictions are the same
new_pred = model_2(x)
assert torch.all(torch.eq(pred_before_saving, new_pred)).item() == 1
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:
trainer_options.update(checkpoint_callback=True)
trainer = Trainer(**trainer_options)
initial_values = torch.tensor(
[torch.sum(torch.abs(x)) for x in model.parameters()])
result = trainer.fit(model)
post_train_values = torch.tensor(
[torch.sum(torch.abs(x)) for x in model.parameters()])
assert result == 1, 'trainer failed'
# Check that the model is actually changed post-training
assert torch.norm(initial_values - post_train_values) > 0.1
# test model loading
pretrained_model = load_model_from_checkpoint(
logger, trainer.checkpoint_callback.best_model_path)
# test new model accuracy
test_loaders = model.test_dataloader()
if not isinstance(test_loaders, list):
test_loaders = [test_loaders]
for dataloader in test_loaders:
run_prediction(dataloader, pretrained_model)
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.checkpoint_connector.hpc_save(save_dir, logger)
trainer.checkpoint_connector.hpc_load(save_dir, on_gpu=on_gpu)
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_default_args(tmpdir):
"""Tests default argument parser for Trainer"""
# logger file to get meta
logger = tutils.get_default_logger(tmpdir)
parser = ArgumentParser(add_help=False)
args = parser.parse_args()
args.logger = logger
args.max_epochs = 5
trainer = Trainer.from_argparse_args(args)
assert isinstance(trainer, Trainer)
assert trainer.max_epochs == 5
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_default_args(mock_argparse, tmpdir):
"""Tests default argument parser for Trainer"""
mock_argparse.return_value = Namespace(**Trainer.default_attributes())
# logger file to get meta
logger = tutils.get_default_logger(tmpdir)
parser = ArgumentParser(add_help=False)
args = parser.parse_args()
args.logger = logger
args.max_epochs = 5
trainer = Trainer.from_argparse_args(args)
assert isinstance(trainer, Trainer)
assert trainer.max_epochs == 5
def test_loading_yaml(tmpdir):
tutils.reset_seed()
hparams = EvalModelTemplate.get_default_hparams()
# save tags
logger = tutils.get_default_logger(tmpdir)
logger.log_hyperparams(Namespace(some_str='a_str', an_int=1, a_float=2.0))
logger.log_hyperparams(hparams)
logger.save()
# load hparams
path_expt_dir = tutils.get_data_path(logger, path_dir=tmpdir)
hparams_path = os.path.join(path_expt_dir, 'hparams.yaml')
tags = load_hparams_from_yaml(hparams_path)
assert tags['batch_size'] == 32 and tags['hidden_dim'] == 1000
def test_no_val_module(monkeypatch, tmpdir, tmpdir_server, url_ckpt):
"""Tests use case where trainer saves the model, and user loads it from tags independently."""
# set $TORCH_HOME, which determines torch hub's cache path, to tmpdir
monkeypatch.setenv("TORCH_HOME", str(tmpdir))
model = EvalModelTemplate()
# logger file to get meta
logger = tutils.get_default_logger(tmpdir)
trainer = Trainer(
default_root_dir=tmpdir,
max_epochs=1,
logger=logger,
checkpoint_callback=ModelCheckpoint(dirpath=tmpdir),
)
# fit model
result = trainer.fit(model)
# training complete
assert result == 1, "amp + ddp model failed to complete"
# save model
new_weights_path = os.path.join(tmpdir, "save_test.ckpt")
trainer.save_checkpoint(new_weights_path)
# assert ckpt has hparams
ckpt = torch.load(new_weights_path)
assert LightningModule.CHECKPOINT_HYPER_PARAMS_KEY in ckpt.keys(), "module_arguments missing from checkpoints"
# load new model
hparams_path = tutils.get_data_path(logger, path_dir=tmpdir)
hparams_path = os.path.join(hparams_path, "hparams.yaml")
ckpt_path = (
f"http://{tmpdir_server[0]}:{tmpdir_server[1]}/{os.path.basename(new_weights_path)}"
if url_ckpt
else new_weights_path
)
model_2 = EvalModelTemplate.load_from_checkpoint(
checkpoint_path=ckpt_path,
hparams_file=hparams_path,
)
model_2.eval()
def test_cpu_slurm_save_load(tmpdir):
"""Verify model save/load/checkpoint on CPU."""
hparams = EvalModelTemplate.get_default_hparams()
model = EvalModelTemplate(**hparams)
# logger file to get meta
logger = tutils.get_default_logger(tmpdir)
version = logger.version
# fit model
trainer = Trainer(
default_root_dir=tmpdir,
max_epochs=1,
logger=logger,
limit_train_batches=0.2,
limit_val_batches=0.2,
checkpoint_callback=ModelCheckpoint(tmpdir),
)
result = trainer.fit(model)
real_global_step = trainer.global_step
# traning complete
assert result == 1, 'cpu model failed to complete'
# predict with trained model before saving
# make a prediction
dataloaders = model.test_dataloader()
if not isinstance(dataloaders, list):
dataloaders = [dataloaders]
for dataloader in dataloaders:
for batch in dataloader:
break
x, y = batch
x = x.view(x.size(0), -1)
model.eval()
pred_before_saving = model(x)
# test HPC saving
# simulate snapshot on slurm
saved_filepath = trainer.checkpoint_connector.hpc_save(
trainer.weights_save_path, logger)
assert os.path.exists(saved_filepath)
# new logger file to get meta
logger = tutils.get_default_logger(tmpdir, version=version)
trainer = Trainer(
default_root_dir=tmpdir,
max_epochs=1,
logger=logger,
checkpoint_callback=ModelCheckpoint(tmpdir),
)
model = EvalModelTemplate(**hparams)
# set the epoch start hook so we can predict before the model does the full training
def assert_pred_same():
assert trainer.global_step == real_global_step and trainer.global_step > 0
# predict with loaded model to make sure answers are the same
trainer.model.eval()
new_pred = trainer.model(x)
assert torch.all(torch.eq(pred_before_saving, new_pred)).item() == 1
model.on_epoch_start = assert_pred_same
# by calling fit again, we trigger training, loading weights from the cluster
# and our hook to predict using current model before any more weight updates
trainer.fit(model)
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()
