David Marx

"vanilla return value

1

rather than an ambiguous _model attribute, let's just attach an attribute that returns the objects that would be returned from the native "load" function of the particular model implementation the...

better name...

7

"mmc" is fine, "multi-modal comparators" is awful. Maybe "zero-shot classifiers?" zsc is nice and memorable

"load" perceptors that are actually just API wrappers. will probably need to take a token or something as an argument

Attach model metadata to facilitate estimating inference/memory cost

2

e.g. * Size in RAM * some model-standardized measure of inference speed

Feature brainstorming

6

- [ ] loss scoring - [x] multi-perceptor - [x] weighted multi-perceptor - [ ] cutout methods? + augs? make that an independent library maybe? - [ ] perceptor weight...

I'm thinking this would be an attribute on the loader. could potentially be an arbitrary yaml or MD file that accompanies the model however

* currently defaulting to dot product * add optional normalization for cosine similarity * add great circle/spherical dists. * maybe we can use this modularization to set-up preferred losses for...