Joel Akeret

No worries. This looks quite interesting. I had implemented a very similar version of the dice coefficient as [you ](https://github.com/jis478/Tensorflow/blob/master/Unet_modified/modeling.py#L213) did. However in my case the network struggled to learn...

I was just using the toy example of the package (https://github.com/jakeret/tf_unet/blob/master/scripts/launcher.py) its not imbalance but in principle the network should be able to learn the segmentation

Wow, the results in your notebook look very good. I transfered [your implementation](https://github.com/jis478/Tensorflow/blob/master/Unet_modified/example/membrane/code/modeling.py#L214) into the tf_unet codebase and ran the toy example with it. However, then the network is not...

- Have you tried to train it for many more epochs? 10 is most likely not enough. - Then you could experiment with the number of filters and/or layers. -...

Maybe you could logarithmize the input to reduce the dynamic range. The labels have to be one-hot-encoded. That's why you have 2 channels It doesn't seem to learn anything. Have...

hard to tell what is going on, sorry No it should be dtype=float32

Well the error `logits and labels must be same size: logits_size=[1048576,2] labels_size=[0,2]` is saying that the prediction and the label shapes are no longer in line after the code changes....

It's correct (and a flaw of the current implementation) that in the default implementation the validation is not optimal. It's possible that this might lead to overfitting or at least...

I was thinking about an approach similiar to Keras where one has to provide two data_provider, one für the training and one for validation. This would allow for a clean...

Yes you're right this could be refactored