Problem in training with 'channels_first' mode

[biswa-iitg]

Is UNET model training using ResNet backbone is possible with 'channels_first' (3,H,W) mode? I am trying and the model is not learning. The same system trains well with 'channels_last' and input shape (H,W,3).

[qubvel]

Please provide a code example

[biswa-iitg]

@qubvel Please find the code
train_aug_py.txt

[qubvel]

I have never try to train network with channels_first
You can do next:

1. train network with channels_last and save weights
2. create same network, but with channels_first and load weights

If an output is the same from both networks, the problem might be in loss function