https://github.com/ovshake/cobra

Code for COBRA: Contrastive Bi-Modal Representation Algorithm (https://arxiv.org/abs/2005.03687)
https://github.com/ovshake/cobra

contrastive-learning cross-modal machine-learning pytorch representation-learning

Last synced: about 15 hours ago
JSON representation

Code for COBRA: Contrastive Bi-Modal Representation Algorithm (https://arxiv.org/abs/2005.03687)

• Host: GitHub
• URL: https://github.com/ovshake/cobra
• Owner: ovshake
• License: mit
• Created: 2020-05-10T19:02:28.000Z (over 4 years ago)
• Default Branch: master
• Last Pushed: 2023-07-06T21:54:54.000Z (over 1 year ago)
• Last Synced: 2024-04-21T11:26:19.668Z (7 months ago)
• Topics: contrastive-learning, cross-modal, machine-learning, pytorch, representation-learning
• Language: Python
• Homepage:
• Size: 477 KB
• Stars: 14
• Watchers: 2
• Forks: 3
• Open Issues: 5
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# COBRA: Contrastive Bi-Modal Representation Algorithm

This repository contains the code for the paper COBRA: Contrastive Bi-Modal Representation Algorithm [(ArXiv)](https://arxiv.org/abs/2005.03687) by [Vishaal Udandarao](https://vishaal27.github.io/), [Abhishek Maiti](http://ovshake.me), Deepak Srivatsav, Suryatej Reddy, Yifang Yin and [Rajiv Ratn Shah](http://faculty.iiitd.ac.in/~rajivratn/). 

## Methodology 

We present a novel framework COBRA that aims to train two modalities (image and text) in a joint fashion inspired by the **Contrastive Predictive Coding (CPC)** and **Noise Contrastive Estimation (NCE)** paradigms which preserve both inter and intra-class relationships. We empirically show that this framework reducesthe modality gap significantly and generates a robust and task agnostic joint-embedding space. We outperform existing work on four diverse downstream tasks spanning across seven benchmark cross-modal datasets.


**A visualisation of the loss function**:







![Supervised Contrastive Loss](https://github.com/ovshake/cobra/blob/master/images/CPCLoss.JPG)

## Architecture

![](https://github.com/ovshake/cobra/blob/master/images/Architecture.JPG)

## Datasets

The 7 datasets used to empirically prove our results are:
 

1. PKU-XMedia 

2. MS-COCO 

3. NUS-Wide 10k 

4. Wikipedia

5. FakeNewsNet

6. MeTooMA

7. CrisisMMD

## t-SNEs

![](https://github.com/ovshake/cobra/blob/master/images/t-SNE_1.JPG)

![](https://github.com/ovshake/cobra/blob/master/images/t-SNE_2.JPG)

## Results

![Wikipedia](https://github.com/ovshake/cobra/blob/master/images/Wiki.JPG)

![MS-COCO](https://github.com/ovshake/cobra/blob/master/images/MSCoco.JPG)

![PKU-XMedia](https://github.com/ovshake/cobra/blob/master/images/PKU-XMedia.JPG)

![MS-COCO](https://github.com/ovshake/cobra/blob/master/images/NUS-Wide.JPG)

![MS-COCO](https://github.com/ovshake/cobra/blob/master/images/FakeNewsNet.JPG)

![MS-COCO](https://github.com/ovshake/cobra/blob/master/images/MeTooMA.JPG)

![MS-COCO](https://github.com/ovshake/cobra/blob/master/images/CrisisMMD.JPG)

## Instructions for running

The code has been tested on Python 3.6.8 and PyTorch 1.5.1. 

- Install all the dependencies using the following command:

```

pip install -r requirements.txt

```

- Create a folder `features` to save the trained models

- To train COBRA, use the following command:

```

python main.py

```

- To switch between NCE contrastive loss and softmax contrastive loss, change the `use_nce` flag. To change the number of anchor points and number of negative samples, modify the `num_anchors` and `num_negative_samples` respectively.

## Queries

In case of any queries, please open an issue. We will respond as soon as possible.