Hi, I have created extended example based on your repository: https://github.com/mrgloom/keras-semantic-segmentation-example

It also use random colors for foreground and background (not like lighter and darker like here https://github.com/ZFTurbo/ZF_UNET_224_Pretrained_Model/blob/master/train_infinite_generator.py#L24 ), one idea behind it is that in that case network can learn 'shape of object' not just 'thresholding and separating background and foreground', also looks like using random colors make problem harder and network converges slower.

Also I have experienced some problems:

1. Netwoks not always converges on second run with fixed params even for this toy problem, looks like it depens on random seed.
2. Dice loss and jaccard loss are harder to train than binary crossentropy, any ideas why? Network architecture is the same just loss differs, I even tried to load trained weights from binary crossentropy loss network and use them in dice loss network which show high dice coef.

I know this is not an issue, but I wanted to contact you to know how did you make the network deeper in keras for the DSTL competition using this model?

When I use tensorflow-1.3.0 as backend, I get this kind of error:

builtins.ValueError: Dimension 2 in both shapes must be equal, but are 3 and 32 for 'Assign' (op: 'Assign') with input shapes: [3,3,3,32], [3,3,32,3].

Here is preprocessing for the batch (looks like 256 should be 255 ;) ) https://github.com/ZFTurbo/ZF_UNET_224_Pretrained_Model/blob/master/zf_unet_224_model.py#L27

Is it ok for real images to use code like this or it should be calculated for entire dataset?

batch=batch-np.mean(batch)
batch=batch/np.std(batch)

Also how crucial is impact of data normalization for U-net? In my tests even on this simple synthetic data network doesn't converges if input is not normalized.

You have generated pretrained weights for 224224 input size, but I have 128128. How can we use such weights in this situation, but without padding/upsampling 128*128 images. Sorry for silly question - is it worth trying in kaggle salt competition?

Traceback (most recent call last): File "train.py", line 11, in import segmentation_models as sm File "/home/melih/anaconda3/envs/ai/lib/python3.6/site-packages/segmentation_models/init.py", line 98, in set_framework(_framework) File "/home/melih/anaconda3/envs/ai/lib/python3.6/site-packages/segmentation_models/init.py", line 68, in set_framework import efficientnet.keras # init custom objects File "/home/melih/anaconda3/envs/ai/lib/python3.6/site-packages/efficientnet/keras.py", line 17, in init_keras_custom_objects() File "/home/melih/anaconda3/envs/ai/lib/python3.6/site-packages/efficientnet/init.py", line 71, in init_keras_custom_objects keras.utils.generic_utils.get_custom_objects().update(custom_objects) AttributeError: module 'keras.utils' has no attribute 'generic_utils'

when I run the code, I got the result below but don't know why there is no generic_utils attribute in the library since there is in the keras.

Is it possible to run the model for 6 input channels? Three inputs in that are RGB values and the other three are metrics I want to pass on into the architecture for my use case.

Thanks for the repo. I am wondering why do you use a smoothing factor of 1.0 in both dice and jaccard coefficients? Where does this value comes from? And what about using another smaller value close to zero, e.g. K.epsilon()

Hi! I would like to know how I should perform the model.fit instruction. model.fit(trainSet, mask_trainSet, batch_size=20, nb_epoch=1, verbose=1,validation_split=0.2, shuffle=True, callbacks=[model_checkpoint])¿? What I write in callback??

And how should I use the weights if I wan't to use pretained weights??

Thank you very much and sorry for the inconvenience!

I run your code and then find that the img batch has a shape(16,224,224,3),but mask batch has a shape(16,1,224,224). I don't understand it.Can you explain it to me?I use my dataset to train unet and then the dice coef is high，but the real effect is bad.