Visualisation of predictions during validation steps
We want the same kind of work done in #290 (closed), but during validation steps while training a model.
We need to
- select 5 random image_idx that will be displayed in Tensorboard (the same will be used at every epoch)
- look for the predictions from batch_values
- Return them to be able to use them in train
- Create a new function that generates a similar images (image on the left, prediction displayed on the right, no GIF this time)
- try to reuse as much code as possible
- use
add_image
to write the image with its prediction with the TensorboardWriter. - we don't have the position of the lines, we simply divide the height of the page by the number of lines to write to determine where they should be displayed.
This could be done in two steps (and two Merge Requests):
- fully support line-level, partially support page-level (do not care about line positioning, simply display the text on one line in the middle of the image)
- fully support page-level by splitting the heights into rectangles with the same dimensions.
Edited by Yoann Schneider