(Collaboratively done by Shweta Shaw and myself)
Implements the ideas presented in Supervised Contrastive Learning by Khosla et al. The authors propose a two-stage framework to enhance the performance of image classifiers and also achieves SoTA results.
(Figures gathered from the paper)
A detailed discussion of the paper and the results of our experiments are available here in this report.
This repository consists of the notebooks (runnable on Colab) showing the experiments we have done.
- Contrastive loss for supervised classification.
- Prannay Khosla for sharing his comments on our work.
├── Flowers
│ ├── Contrastive_Training_Flowers.ipynb
│ ├── Contrastive_Training_Flowers_Augmentation.ipynb
│ ├── Fully_Supervised_Training_Flowers.ipynb
│ └── Fully_Supervised_Training_Flowers_Augmentation.ipynb
├── ImageNet_Subset
│ ├── Contrastive_Training_Imagenet_subset_Adam.ipynb
│ ├── Contrastive_Training_Imagenet_subset_RMSprop.ipynb
│ ├── Contrastive_Training_Imagenet_subset_SGD.ipynb
│ ├── Fully_Supervised_Training_IMGNET_subset_Adam.ipynb
│ ├── Fully_Supervised_Training_IMGNET_subset_RMSprop.ipynb
│ └── Fully_Supervised_Training_IMGNET_subset_SGD.ipynb
├── Pets
│ ├── Contrastive_Training_Pets.ipynb
│ └── Fully_Supervised_Training_Pets.ipynb
├── Visualization_ImageNet_subset.ipynb
├── Visualization_Pets.ipynb
Contrastive_Training_*.ipynbnotebooks show the supervised contrastive framework proposed in the paper.Fully_Supervised_Training_*.ipynbnotebooks show the typical fully supervised training with different datasets.Visualization_ImageNet_*.ipynbnotebooks show the visualizations of the embeddings learned by the supervised contrastive learning framework.
- Flowers
- Cats vs. Dogs
- ImageNet Subset (https://github.com/thunderInfy/imagenet-5-categories)
- The authors used AutoAugment in the paper. However, we used simple augmnetation operations which worked for the datasets we tried. Note that, there's no augmentation for the Pets dataset as we got pretty good results on that one even without any data augmentation.
- LARS optimizer was used in the paper, however we used Adam. We have also shown the effect of different optimizers like SGD and RMSProp along with learning rate schedules.
The above plots are from the experiments conducted on the Pets dataset. More results from the other two datasets have been discussed in the above-mentioned report and can be found here: https://app.wandb.ai/authors/scl.
If you go to any of the notebooks listed in the repository and use an extension like "Open notebook in Google Colab" to open it, you should be able to run the experiments right off the bat.
At the time of performing the experiments, we used TensorFlow 2.2. We specifically did not denote the versions of the other libraries. All of our experiments were performed on Google Colab.
Via GitHub issues