This repository contains a PyTorch implementation of SimCLR based on the work A Simple Framework for Contrastive Learning of Visual Representations by Chen et al.
SimCLR consists of 4 elements:
- a random data augmentation (crops and color jitter)
- a base encoder (ResNet50)
- a projection head (2 layer MLP)
- a contrastive loss (NT-Xent)
Each input image is randomly tranformed twice. The task is to identify the second transformation among a batch of other source images. For efficiency, the other images in the batch are used as negative examples and the loss is calculated based on pairwise cosine similarities.
- LARS instead of Adam
- Same LR schedule
- Latent dimension of 128
Linear evalutation using modified ResNet50, t=0.5, bs=256.
| Source | Epoch 100 | Epoch 200 | Epoch 300 | Epoch 400 | Epoch 500 |
|---|---|---|---|---|---|
| Paper | 83.9% | 89.2% | 91.5% | 92.1% | 93.0% |
| This repo | 79.5% | 82.8% |
As shown in Fugure B.6. These numbers aren't final yet.