I am currently doing the first assignment of Week 1 and in the notebook it says “For this exercise, don’t worry about vectorization! Just implement everything with for-loops”. However, I would still like to know how we can reformulate the convolution forward and backward pass in a vectorized way to speed up computation. Could someone show me how to do that or point me into the right direction?
You will see it later in the course. It’s just using the tensorflow built-in functions and layers. You just have to create the network with convolutional layers and TF will do it for you
thanks for your answer. Sorry, I think I didn’t phrase my question well enough. I was wondering if there is something like a clever way to arrange the input data and use vectorization, such that we can refrain from looping over each height and width index of an image sequentially. Because theoretically, I think it should be possible to compute each single step convolution (see function
conv_single_step in the assignment), at once, since they’re all independent computations. I am pretty sure TF is doing that under the hood already, but I would like to be able to re-implement that myself with NumPy.