Display predictions during training

We need to develop a way to display predictions during training. The idea is to:

• select 5 images from the validation set at random
• at each validation step, keep the prediction from these 5 images
• for each of them, use the TensorboardWriter to export a 3-plot figure (image + labels + prediction)

To display the text, you can either:

• try things with matplotlib (place the text regularly etc...)
• try to generate some clean HTML and render it with weasyprint
• anything else?