Different error rates on the learn.fit_one_cycle,although with same values

Part 1 (2019)
1

Although I set the np.raandom.seed() value to a constant value I get different error rates everytime I run the learn.fit_one_cycle(5,max_lr=slice(1e-03,1e-02)).Isn’t this suppose to give the same error rate everytime I run this function .Please help
Thanks

3

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

gpu = 0
torch.cuda.set_device(gpu)
cudnn.benchmark = True
cudnn.enabled=True
torch.cuda.manual_seed_all(seed)

Have you tried add those above?

1 Like
4

Will surely try this.Thanks.
But can you please touch on why were the error_rates different without typing the codes mentioned by you?

5

np.random.seed only deals with numpy. There is also randomness in pytorch (which fastai is built on). You can read more at https://pytorch.org/docs/stable/notes/randomness.html

1 Like