WARNING: conv_backward and pool_backward mistake on comments (ed: no, not really)

Course Q&A Deep Learning Specialization Convolutional Neural Networks
1

Hello, i found a mistake in the instructions of conv and pool backward. In the comments there is a line :

#da_prev_pad[vert_start:vert_end, horiz_start:horiz_end, :] += None
# dA_prev[i, vert_start: vert_end, horiz_start: horiz_end, c] += None

the right instructions is :

#da_prev_pad[vert_start:vert_end, horiz_start:horiz_end, :] += None
#dA_prev[i, horiz_start: horiz_end,vert_start: vert_end , c] += None

the bug causes wrong values on dW (conv_backward) and empty array after some iterations (pooling_backward)

2

Sorry but you are misinterpreting things. Why would the vertical and horizontal indices be reversed in the second case? Also remember that h stands for “height” not “horizontal”. The comments are correct as written.

If you are getting empty outputs, it more likely means you have made a mistake in handling the stride or in the loop index values. Remember that we are looping over the output space, but updating the input space in back prop.