explainability tools for representation

[mattersoflight]

We should extend the explainability tools developed for robust virtual staining to representation learning. I suggest the following:

• Export a movie of features (input, PC of features) for a chosen track.
• Effective receptive field of a chosen embedding dimension.

[ziw-liu]

Effective receptive field of a chosen embedding dimension.

The embedding is the global average pooling of the deepest feature map, so by definition its ERF is the entire input image.

[ziw-liu]

Related: #170, #162.