Misleading example for per-sample gradient

[mingfeisun]

In the example of per-sample gradient, the following line can be misleading since the predictions of a net are logits:

tutorials/intermediate_source/per_sample_grads.py

Line 49 in 08a61b7

return F.nll_loss(predictions, targets)

The correct way should be:

return F.nll_loss(F.log_softmax(predictions, dim=-1), targets) 

Would appreciate if this can be corrected.

[jmarintur]

Hi @mingfeisun, looking at the network used in the example, it does seem there's an error:

tutorials/intermediate_source/per_sample_grads.py

Lines 44 to 46 in 08a61b7

	output = F.log_softmax(x, dim=1)
	output = x
	return output

the output is overwritten after applying the F.log_softmax in L45. If we remove that line, all should be good. Wdyt @albanD ?

[albanD]

Ho yeah that looks like a typo, can you send a PR to fix this please?

[jmarintur]

Here's the PR @albanD.