I have a few questions from the class –
net = nn.Sequential( nn.Linear(28*28, 10), nn.LogSoftmax() )
In last non-linear layer, why did we use logsoftmax, not softmax? Weren’t we exponentiating outputs from 2nd last layer so as to make them all +ve ? Why back to log after doing [exp]/[sum of exp].
n.Parameter(torch.randn(*dims)/dims[0])
What is the reason of dividing by dims[0]. I tried and it doesn’t work if we don’t divide by dims[0]. By it doesn’t work, I mean fit() gives loss = nans and very bad accuracy.
Thanks 
I just posted the video.
1 Like
The loss functions in pytorch generally assume you have LogSoftmax, for computational efficiency reasons: https://discuss.pytorch.org/t/does-nllloss-handle-log-softmax-and-softmax-in-the-same-way/8835
This is He initialization (http://www.jefkine.com/deep/2016/08/08/initialization-of-deep-networks-case-of-rectifiers/) . Although I may have forgotten a sqrt there…
Without careful initialization you’ll get gradient explosion. We discuss this in the DL course.
1 Like
Helpful links. Thanks
hi there,
I have a question pertaining to optimizer.zero_grad() . I have gone couple of times over the section where it is explained why do we have to call this function.
I still don’t understand it .
From pytorch forum , i understand unless for the special cases where one wants to simulate bigger batches by accumulating the gradients , one has to invoke optimizer.zero_grad() to clear the grandients for the next batch.
Would like to understand Jeremy explanation though.
The title changed “with” for “in” and the link got broken.
New one.
https://jidindinesh.github.io/2019-02-08-Broadcasting-In-PyTorch
On the next video (lesson 10) around minute 25-26 he explains again why we need to clear the weights and bias.
Hope it helps.
