README.md

CLIP (Contrastive Language-Image Pretraining)

Installation

sudo apt-get install python3-venv
python3 -m venv clip-env
source clip-env/bin/activate
pip3 install -U pip
pip3 install -r requirements.txt

Dataset

This project uses the Flickr30k dataset from Kaggle which has become a standard benchmark for sentence-based image description. You can download the dataset as follow:

python3 download.py

Make sure you have kaggle.json file with your credentials in the given folder or you can manually provide them after running the script.

Training

After downloading the dataset, run the following script to train the CLIP model. Tweak the hyperparameters for the model in config.py. You can also change the Image and Text encoder as required.

python3 train.py

Inference pipeline is given in demo.ipynb. After training the model, feel free to test the model with your own images and texts.

How the training works?

Assume we have a batch of N images paired with their respective descriptions e.g. <image1, text1>, <image2, text2>, <imageN, textN>.

Contrastive Pre-training aims to jointly train an Image and a Text Encoder that produce image embeddings [I1, I2,...,IN] and text embeddings [T1, T2,..., TN], in a way that:

• The cosine similarities of the correct <image-text> embedding pairs <I1,T1>, <I2,T2> (where i=j) are maximized.

• In a contrastive fashion, the cosine similarities of dissimilar pairs <I1,T2>, <I1,T3>… <Ii,Tj> (where i≠j) are minimized.

Here’s a step-by-step breakdown of how CLIP works:

1. The model receives a batch of (image,text) pairs.

2. The text encoder is a standard Transformer model, say BERT. The image encoder can be either a ResNet or a Vision Transformer.

3. For every image in the batch, the Image Encoder computes an image vector. The first image corresponds to the I1 vector, the second to I2, and so on. Each vector is of size de, where de is the size of the latent dimension. Hence, the output of this step is $N \times$ de matrix.

4. Similarly, the textual descriptions are squashed into text embeddings [T1, T2,...,TN], producing a $N \times$ de matrix.

5. Finally, we multiply those matrices and calculate the pairwise cosine similarities between every image and text description. This produces an $N \times N$ matrix.

6. The goal is to maximize the cosine similarity along the diagonal - these are the correct (image, text) pairs. In a contrastive fashion, off-diagonal elements should have their similarities minimized (e.g., I1 image is described by T1 and not by T2, T3, T4, etc).

A few extra remarks:

• The model uses the symmetric cross-entropy loss as its optimization objective. This type of loss minimizes both the image-to-text direction as well as the text-to-image direction (remember, our contrastive loss matrix keeps both the (I1, T2) and (I2, T1) cosine similarities).

• Contrastive pre-training is not entirely new. It was introduced in previous models and was simply adapted by CLIP.

Results

Text to Image:

The following results are the most relevant images a from the validation set related to the text query.

Image to Text:

The following results are the most relevant texts a from the validation set related to the image query.