python - Tensorflow for image segmentation: Changing minibatch size stops learning

Question

I have created a net for image segmentation, in particular brain tumors. The jupyter notebook with the code is here.

When I train the CNN with minibatch size of 1, I get a fairly good result:

But when I change the size to something larger ( 2 or more) the results are terrible:

Tensorboard shows the diference in the loss. Clearly the the net with batchsize 2 is not minimizing the loss (blue)

Any ideas on why this could be the case?

Answer 1

I found the problem. I checked my graph with tensorboard and I notice that in the CONV1/S1 I was not connecting the output of the ReLu to the next layer (CONV1/S2), instead I was connecting the output of the conv2d directly.

I changed that line in the code and everything is working as expected.