LUS COVID demo

[JulienVig]

Bug fixes to set up the LUS COVID demo for next week

• Add test cases lus_covid (image_loader, validator, end to end training)
• Fix node image_loader default channel error (default to 1 while it should be 3)
• Add lus_covid task to CLI
• Fine-tune hyper parameters
• Image dataset shuffling is done by hand while tfjs already implements it
• Fix label encoding mistake (using tf.oneHot on 'positive' and 'negative' yields the same label for both class)
• "Train Collaboratively" raises an error
• Training loss and other metrics stay constant after the first few epochs
• Model always predicts positive
• Training collaboratively with two web-client results in timeouts (server drop client contributions)
  • Addressed by small lus covid fixes #661
• UI number of participants doesn't get updated as they join a task
  • Addressed by small lus covid fixes #661
• web-client training performance stays constant while training via node.js works well

Model used to only predict the majority class (~0.5 accuracy), by normalizing image values from 0-255 to 0-1 the training now yields >0.9 on both training and validation sets

[tharvik]

génial, thanks for the hard work!

[martinjaggi]

super nice, congrats!