|
--- |
|
language: |
|
- en |
|
tags: |
|
- zero-shot-image-classification |
|
license: mit |
|
datasets: |
|
- coco2017 |
|
--- |
|
|
|
# Tiny CLIP |
|
## Introduction |
|
This is a smaller version of CLIP trained for EN only. The training script can be found [here](https://www.kaggle.com/code/sachin/tiny-en-clip/). This model is roughly 8 times smaller than CLIP. This was achieved by using a small text model (`microsoft/xtremedistil-l6-h256-uncased`) and a small vision model (`edgenext_small`). For a in-depth guide of training CLIP see [this blog](https://sachinruk.github.io/blog/pytorch/pytorch%20lightning/loss%20function/gpu/2021/03/07/CLIP.html). |
|
|
|
## Usage |
|
For now this is the recommended way to use this model |
|
``` |
|
git lfs install |
|
git clone https://huggingface.co/sachin/tiny_clip |
|
cd tiny_clip |
|
``` |
|
Once you are in the folder you could do the following: |
|
```python |
|
import models |
|
text_encoder, tokenizer, vision_encoder, transform = models.get_model() |
|
``` |
