I think i've just found out what happened thanks to your reply. Basically my DataLoader is computing some preprocessing on the raw data on the first epoch and then saves the resulting tensors for them to be used in the further epochs... As i'm using torch.Save to save those tensors, it triggers the trains api.
Hi MiniatureShells8
The torch.save triggers the model creation.
If you are using the same filename, then the same model in the system will be used.
New filenames will create new models.
What exactly is your use case ?
You can disable it with:
Task.init('example', 'train', auto_connect_frameworks={'pytorch': False})
Okay thank you very much for your help, much appreciated ❤
Okay good to know, doing that would mean i'll not have my models saved at all (they would be locally of course, but not visible on the trains app), only my tensorboard calls would be saved, is that correct ?