Regarding reproducibility of the results.

[AdityaJainTD]

Hi,

I have been training with different strategies (EWC, Replay, Hybrid). Often times the evaluation accuracy ranges from 88.xx - 92.xx for the same code without any changes applied to it. I have provided the seed value for the nc_benchmark function. Do I need to seed anything else to generate reproducible results?

[AndreaCossu]

In addition, try calling this function at the beginning of your main

def set_seed(seed):
    if seed is None:
        return
    torch.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    np.random.seed(seed)
    random.seed(seed)
    if torch.cuda.is_available():
        torch.backends.cudnn.deterministic = True
        torch.backends.cudnn.enabled = True
        torch.backends.cudnn.benchmark = False

[AdityaJainTD]

I tried adding this function into my code. I ran it quite a few times, my score was still fluctuating between 85.xx - 90.xx.

[AdityaJainTD]

Hi,
I am still facing the same issue. Is there anything else that I need to add seed value to?

[AntonioCarta]

Avalanche doesn't really add any randomization mechanism. They all reuse underlying libraries (numpy, python random module, pytorch). You can check the pytorch randomness guide to see if there is anything missing.