Decoupled-Contrastive-Learning

This repository is an implementation for the loss function proposed in Decoupled Contrastive Loss paper.

Requirements

Pytorch
Numpy

Usage Example

import torch
import torchvision.models as models

from loss import dcl

resnet18 = models.resnet18()
random_input = torch.rand((10, 3, 244, 244))
output = resnet18(random_input)

# for DCL
loss_fn = dcl.DCL(temperature=0.5)
loss = loss_fn(output, output)  # loss = tensor(-0.2726, grad_fn=<AddBackward0>

# for DCLW
loss_fn = dcl.DCLW(temperature=0.5, sigma=0.5)
loss = loss_fn(output, output)  # loss = tensor(38.8402, grad_fn=<AddBackward0>)

How to Run

You can simply run_simclr.sh to train and test the SimCLR.

You can select different loss function by changing the LOSS variable in the same file. The valid value for this variable include ce for Cross-Entropy, dcl for Decoupled Contrastive Loss and dclw for Weighted Decoupled Contrastive Loss.

The dataset can be selected by changing the DATASET variable as well. Valid values include cifar10, cifar100, and stl10.

The final results can be seen in the output stream.

Also, if you operate under Slurm, you can submit a job using sbatch run_simclr.sh and find the output file in the same directory.

Results

Below are the SimCLR results of the Resnet18 on the CIFAR10 dataset. The temperature is set to 0.1 and sigma to 0.5. The model is also trained for 100 epochs on SimCLR pretraining and 100 epochs on linear probing.

Loss	32 Batch Size	64 Batch Size	128 Batch Size	256 Batch Size
Cross Entropy	78.3	81.47	83.09	83.26
DCL	84.6	85.57	85.63	85.36
DCLW	83.32	82.68	83.5	82.7

Below are the SimCLR results of the Resnet18 on the CIFAR100 dataset. The temperature is set to 0.1 and sigma to 0.5. The model is also trained for 100 epochs on SimCLR pretraining and 100 epochs on linear probing.

Loss	32 Batch Size	64 Batch Size	128 Batch Size	256 Batch Size
Cross Entropy	52.2	57.76	58	59.6
DCL	60.1	60.33	61.73	61.92
DCLW	60.16	59.7	59.05	57.3

Credits

The SimCLR implementation is taken from https://github.com/leftthomas/SimCLR.

Contribution

Any contributions would be appreciated!

MaginaDai/Decoupled-Contrastive-Learning