Problem when setting Residual=True about unet-keras

[pietz]

Truth be told I only tested this with the tensorflow backend and now that theano has reached its end of life there's another reason to use tf.

Try changing both concatenations in the code to axis=1. There's one in the convblock and another in the levelblock.

Also, residual connections are usually an addition and not a concatenation, but this will lead to similar shape problems on the first convblock. The difference should be minor.

[bernardohenz]

I already tested this and the problem was similar. But I changed to tensorflow backend and it worked.
I think the shape mismatch may be a combination of matching errors among input, convs and concatenate channel-axis, but now it is working perfectly.

Btw, nice code, very clean and easy to use. I would suggest to add the possibility of choosing Batch/Instance Normalization, but Keras does not have Instance normalization yet (only keras-contrib).

Congrats for the work.

[pietz]

Thanks for the nice words!

[dscarmo]

Hello, i have the same problem with the code as is on the repository, except it occurs anytime i use a input_shape not multiple of 2.

(128, 128, 1) works, (512, 512, 1) works, but other shapes, even the one used on the original paper (572, 572, 1) will output the same error.

For 181, 181, 1, i have the following output:

ValueError: A Concatenate layer requires inputs with matching shapes except for the concat axis. Got inputs shapes: [(None, 45, 45, 256), (None, 44, 44, 256)]

I tried using axis=1 in concatenates, the error changes to:

I also tried changing start_ch, depth and inc_rate to no avail. output:

ValueError: A Concatenate layer requires inputs with matching shapes except for the concat axis. Got inputs shapes: [(None, 45, 45, 256), (None, 44, 44, 256)]

The code is very nice and clean, good work!

For now i will try using center patches of 128x128.

[sakvaua]

@dscarmo Original paper uses valid convolutions while this implementation uses same. With your shapes
572/2=286/2=143 which is not divisible by two. And that's why it gives you an error. Your images should have shapes that divide by two as many times as the depth of your network.

[karanpathak]

Very well written code!