C4 - W3 - Image Segmentation with U-Net

Hello !

I have a question regarding a conv_block
We are implementing a conv_block which will be used four times later in the Lab.
Why the n_filters is set to 32, whereas the first two operations in U-net are convolutions
yielding the dimension of 64 ?

Thanks to anyone for clarifying this.

That is a “keyword” argument in python, so you have to declare a default value for it in the function declaration. But you then can supply whatever value you want when you invoke it.

My concern was just that it is not consistent with the U-Net schema provided in the description, where the 5th conv_block has 1024 layers, and in our implementation - 512 (namely (None, 6, 8, 512)) when not specifying n_filters by hand.

Yepp true! You can try to implement on your own after passing all the tests. Not looks like huge concern. You may achieve new SOTA in deeplearning.ai community :wink:


LOL. I tried to recalculate with n_filters = 64 it takes much more time to compute and accuracy is improved slightly :wink: