Roman Solovyev

You can train without pre-trained weights then. It will work without them.

Did you train your model on some dataset? Model without training won't work on your task.

There can be many reasons. Check if you multiply your predictions with 255 before draw it.

Yes, it's possible. Just change line: https://github.com/ZFTurbo/ZF_UNET_224_Pretrained_Model/blob/master/zf_unet_224_model.py#L23 From: INPUT_CHANNELS = 3 To: INPUT_CHANNELS = 6 But I propose you ti use other models for segemntation from this repository: https://github.com/qubvel/segmentation_models

I think problem for 512x512 case can be also with learning rate. Try to decrease it.

It's just from my experience. You can check some materials about it: https://www.kdnuggets.com/2017/11/estimating-optimal-learning-rate-deep-neural-network.html **Deep learning rule**: if NN must converge but it doesn't, try to reduce learning rate. )

You can set any EPS value. As I know it won't affect result too much. 1.0 is relatively small to the mask size, so it's ok.

It's actually the part which keras do. As I know it computed over batch.

Check the example here: https://github.com/ZFTurbo/ZF_UNET_224_Pretrained_Model/blob/master/train_infinite_generator.py Generator must output image and mask. As I can see you output only mask.

You have incorrect shape for masks. It must have 5 channels (224, 224, 5) but you give RGB images (224, 224, 3).