Reproducibility in Minirocket Pytorch version

[Arderiu]

Hello, I am running some experiments with the Minirocket model (pytroch version) using the GPU. Is there a way to make the generation of the features deterministic? Every time I call Minirocket get_minirocket_features() with a fixed dataset I get slightly different features, which makes it hard to do proper benchmarking. Thanks!

[oguiza]

Hi @Arderiu,
You can try it using:

from fastai.torch_core import set_seed
set_seed(123, reproducibility=True)

before you instantiate the model or learner.

Here's the code set_seed uses:

def set_seed(s, reproducible=False):
    "Set random seed for `random`, `torch`, and `numpy` (where available)"
    try: torch.manual_seed(s)
    except NameError: pass
    try: torch.cuda.manual_seed_all(s)
    except NameError: pass
    try: np.random.seed(s%(2**32-1))
    except NameError: pass
    random.seed(s)
    if reproducible:
        torch.backends.cudnn.deterministic = True
        torch.backends.cudnn.benchmark = False