mt.Callbacks.SavePyvistaPoints(
on_step=1000,
properties=['curvature'],
filename=f'./fold_{fold_number}/train/data'),
mt.Callbacks.LRFind(),
mt.Callbacks.ReduceLROnValidPlateau()
]
test_callbacks = [
mt.Callbacks.SavePyvistaPoints(
on_batch=1,
properties=['curvature'],
filename=f'./fold_{fold_number}/test/data')
]
trainer = mt.Trainer(model, optimizer, criterion, **config.trainer)
start_time = time()
trainer.fit_on_loader(train_loader,
valid_dataloader=valid_loader,
cycles=[
100,
],
callbacks=fit_callbacks,
std_to_file=f'./fold_{fold_number}/train/log')
end_time = time()
train_time = end_time - start_time
trainer.load_best_model(f'./fold_{fold_number}/Checkpoints')
time_log.append(
dict(train_steps_per_epoch=len(train_loader),
test_steps_per_epoch=len(test_loader)))
fold_number += 1
if fold_number in [1]:
print(f'runing fold {fold_number}/{5}')
model = config.ml_model(*config.model.args, **config.model.kwargs)
criterion = torch.nn.L1Loss(**config.criterion)
test_callbacks = []
trainer = mt.Trainer(model, None, criterion, **config.trainer)
trainer.load_best_model(f'./fold_{fold_number}/Checkpoints')
start_time = time()
trainer.test_on_loader(
test_loader,
std_to_file=f'./fold_{fold_number}/test/log',
callbacks=test_callbacks)
end_time = time()
test_time = end_time - start_time
del fold_number
del model
del criterion