Nicolas Audebert

You can add a new Python class in `custom_dataset.py`. Follow the example of the DFC2018 dataset.

@jfrob27 what sort of data normalization do you use on astro data?

I'll have to check which hyperparameters I used for this experiment.

I'll have to check which hyperparameters I used for this experiment.

Yes, the Docker registry at Inria has been closed. I haven't got the chance yet to migrate to Docker hub but you can still build the container locally with `docker...

With what parameters? Can you give some more details to reproduce this problem?

Currently training with multiple images is not supported. This is one the main limitation of this work, I am working on it but it is very low priority for me.

@snowzm you can try `lee` model which is fully convolutional IIRC. Validation accuracy is grossly incorrect in this case.

@DarrenmondZhang This is indeed troublesome. You should not have any prediction in the undefined column. On what dataset is this?

I'm looking into it.