Dev

[ctr26]

This is me trying to fix all the models so that their latent space sizes are all correct.

Not yet ready for merge

pytest is saying a few bits are failing

Changes

• Removed depencence on **kwargs in models and moved kwargs to model factory structure

• [UNSURE] - Attempted to fix VQVAE latentspace not being exact (by averaging before VQ, in hindsight feels wrong)

• [Bug] ResNet50VAEDecoder was using resnet18 before

• Refactored ResNet18VQVAEDecoder and ResNet50VQVAEDecoder to use BaseResNetVQVAEDecoder

• Added testing across different latent spaces

• [UNSURE] - Added scale invariance using Frobenius norm

• Ground work for future cli

• Removed old bad tests

• [UNSURE] - Attempted to fix VQVAE latentspace not being exact (by averaging before VQ, in hindsight feels wrong)

• • @ctr26 needs to check the pythae code on how they deal with their latent space because I think adaptive averaging isnt the way

• [UNSURE] - Added scale invariance using Frobenius norm

• • I've tried to make the model scale invariant and then had to adjust the distance scale ranges so the model isn't dealing with very small numbers. On paper this should work but in practise it may not. Currently the scale of the shape is appending to the latent space vector, before this scaling was essentially spread across the latent space

[ctr26]

@afoix going to fix the ci and this can then be v0.0.2?

[afoix]

@ctr26 I finish to reviwed and yes this can go a as V.0.0.2