Different results when loading model

vision
1

Hello

Yet another question regarding loading models with seed.

I have followed every post about the subject and trying to achieve deterministic results by using the proposed solution:

def seed_torch(seed=1029):
    random.seed(seed)
    os.environ['PYTHONHASHSEED'] = str(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    torch.cuda.manual_seed_all(seed) # if you are using multi-GPU.
    torch.backends.cudnn.benchmark = False
    torch.backends.cudnn.deterministic = True

seed_torch()

But without success, each time I get a different result when evaluating a single data point.

My model has been trained on another computer, but should that matter?

I have checked the values for each layer and they are the same on every load, it is just the output that differs.

Any ideas?

2

Can it be because I’m using nn.MaxPool1d that might be mentioned under Reproducibility — PyTorch 2.1 documentation

… and many forms of pooling, padding, and sampling. There currently is no simple way of avoiding non-determinism in these functions. ?