Lecture 14 - Implementing Convolutions

[@ZhangzhiPeng-x8r]

asstride gives us a efficient way to apply any odd ops to a tensor. Super cool and useful !!

[@nitinnilesh]

at 12:20 - The lecturer says that he is going to convert NHWC to NCWH. However, pytorch uses NCHW and the instructor also wrote the code to make it NCHW when he used the permute. So let's understand it - input is NHWC (index 0 - N, index 1 - H, index 2 - W, index 3 - C). When he permute it to (0, 3, 1, 2), the output becomes NCHW. All I want to say is that instructor wants to make it NCHW but wrote it as NCWH.

[@jinliangli3773]

yes， this is a typo

[@yimingsun6638]

where can we find the notebook you are writing alive?

[@cleyang]

img2col reshape operation for blocked overlap image `[4,4,3,3].reshape(16,9)` will expanding data, which is meaning we copy data in the same location, not reuse it? Then when performing gradient descent, we will have two gradients for same location. But shouldn't convolution accumulate gradients at the same location when performing gradient descent?