clip-models-for-distillation

This repo contains the source code for original and distilled models for CLIP. It also contains a simple API to use the original CLIP model.

Most of the code is taken from the original implementation by OpenAI. It is just organized in a different manner to make it easy to understand.

Related article

https://tech.pic-collage.com/distillation-of-clip-model-and-other-experiments-f8394b7321ce?gi=1d6a8777178a

Things to do before using this repo

1. Download the trained weight for the original CLIP model from this link and put it into the folder weight.
2. Create a new environment and install dependencies using pip install -r requirements.txt.

How to perform distillation

A rough outline is (refer to the article for more details):

from models import papa_clip_visual, baby_clip_visual

optimizer = AdamW(baby_clip_visual.parameters())

for x in loader:
    y_s = baby_clip_visual(x)
    y_t = papa_clip_visual(x)
    loss = loss_function(y_s, y_t)
    optimizer.zero_grad()
    loss.backward()
    ...

How to use the original CLIP model

image = Image.open('path/to/image.jpg')

from models import clip
image_vector = clip.image_to_vector(image)
text_vector = clip.text_to_vector('my text')