geohot / ai-notebooks

Some ipython notebooks implementing AI algorithms
1.27k stars 208 forks source link

Loss was off by factor 10 #9

Open Popeyef5 opened 3 years ago

Popeyef5 commented 3 years ago

The loss in the MNIST from scratch notebook is off by a factor of 10 due to the averaging. The averaging in NLLLoss is done over the batch, like in https://pytorch.org/docs/stable/generated/torch.nn.NLLLoss.html

No meaningful result changes, but nice for completeness.

out variable not deleted because it is used for backprop.