I am having a problem understanding this table. We have a convolutional layer then a max pool layer and then starts the bulk of this architecture. I am comparing this paper with what we did on the PA.
For conv2_x, we have 3 bottleneck blocks, then 4 blocks then 6 and so on. On the PA we split these blocks into identity and convolutional blocks. I am planning on using ResNet50 for my thesis to classify lung x-ray images into covid and normal, this is why I am trying to implement this on my own.
Have you seen this link ?
Hi, @Marios_Constantinou!
Exactly, those blocks are sequentially concatenated as you mentioned. I agree with @balaji.ambresh that there is no reason to reinvent the wheel and replicate the architecture again. For your purpose, I recommend you to import ResNet from tensorflow (even with pre-trained weights if you want) without the top fully-connected layer so you can add your own final layer for binary classification.
I would like to build it myself, mostly to gain more intuition and more practice with tensorflow/keras platforms! So on the PA of week 2 course 4, we didn’t build this actual ResNet50 model right? The one from the paper i mean. It was a mix of the 32 layer block and the 50 layer one? Because what we used on the PA was multiple identity blocks and conv blocks instead of 1 identity block paired with a bottleneck block, right?
