PyTorch Lecture 04: Back-propagation and Autograd

[debanjansengupta]

Thanks a lot!!! Your videos are not just about the framework. You explain every bit in detail.

[khoiphan]

Great explanation of Backprop! I really like the way you visualise the math. Keep it up! :)

[MucciciBandz]

Actually now torch.Tensor and torch.autograd.Variable are now the same class. This means instead of calling w = Variable(torch.Tensor([1]), requires_grad=True) you can call w= <3 great tutorials btw

[crvnse]

Here maths meet the codes. Kim this is awesome that you take out time and explained us the gradient wonders of Deep learning .Video is evidence that pytorch is awesome and kim know how to teach Back Propagation using pytorch.

[VIVEKPANDEYIITB]

Thank you for clarifying the backpropogation using computational graphs

Very good video. Wonderfully explained. Thanks a lot my friend.
Keep up the good work

[nagong]

Thanks a Sung Kim! This is the most straightforward explanation I have ever seen

[bayesianlee]

The very excellent explanation of Backprop. I appreciate it. Thank you professor :)

[reinvanlennep]

Great explanation. Helped me a lot! Good visualisation also.

[alexeilazarev]

It's the best explanation I've ever seen! Thank you!

[taibn]

13:32 i do, I see the results of two different table decimals behind?, from numeric gradient computation only 2 decimals?

[pleung]

May i know at 12:58, why w.data should minus learning*w.grad.data, instead adding to it? thanks

[bluelight]

Great series! I have a few inquiries: 1. How is it that l gets the method backward available to it? Is this the case for any value whose computation required a Variable? 2. is w.grad.data a 1 dimension tensor requiring use of index [0] to get at the one value within it?
3. I realize this is asked already but it's not clear to me why we have to zero the gradients.. Like what happens if we don't?
Thank you

[DanielWeikert]

Thanks great work. One question. What is the exact difference between .grad and .grad.data?

[bosoninfo]

Awesome! Superb explanation!! Keep it up.

[akramsystems]

if you guys are getting objecst printing like *
3.1854)" and want to get just the value without the word tensor add '.item()' to the end of what ever it is you are doing
i.e. 'l.data[0].item()'

[cuenta]

Hi, thanks for the video. Just a question: why is it necessary to initialize each time the gradients? (w.grad.data.zero_())

[KabyanilTalukdar]

why is dy^ - y / dy^ = 1 and not -1? can someone simplify it?

[sudharsunkaleeswaran]

What is the difference between w.grad and w.grad.data?

[charlesenglebert]

what is the difference between w and w.data ?