Interp Vis Helpers

[shahbuland]

Want to add some interpretability helpers for visualizing what different models are doing
In order of priority/value. You can probably use channel-wise PCA for most of these

1. Visualizer of latents to see activations and heatmap of values
2. Visualizer of discriminator detections for what it is seeing as fake/real (I think there is a gradient based way of doing this)
3. Visualizer of feature maps at various stages in encoder/decoder (can do hooks for this)

Code for this should ideally go under owl_vaes/interp/...
I can help add the visualizations to the trainers after they've been implemented.
Refer to train and any trainer to load models for your test cases.
Feel free to message me for a checkpoint, though you could just use some diffusers checkpoint for debuggings sake.

[shahbuland]

this still needs to be done even though i did it in the blog post. It would be nice to have it in a model agnostic format that adds onto wandb