Issue in reproducibility

[mateomorin]

I am unable to get the exact result after training a ttc model on my data (synthetic data), even after setting every possible seed:

torch.manual_seed(42)
random.seed(42)
np.random.seed(42)

torch.use_deterministic_algorithms(True)
if torch.cuda.is_available():
    torch.cuda.manual_seed_all(42)
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False
    
    pytorch_lightning.seed_everything(42)

For full reproducibility:

git clone -b main https://github.com/mateomorin/ttc-training.git
# uncomment last test of test_reproducibility
# potentialy change your data path if you do not have access to bucket projet-ape
uv run pytest     # should have 2 passed, 1 failed and potential warnings