SimCLR

PyTorch implementation of SimCLR: A Simple Framework for Contrastive Learning of Visual Representations by T. Chen et al. With support for the LARS (Layer-wise Adaptive Rate Scaling) optimizer.

Link to paper

Open SimCLR in Google Colab Notebook (with TPU support)

Open SimCLR results comparison on tensorboard.dev:

Quickstart

This downloads a pre-trained model and trains the linear classifier, which should receive an accuracy of ±82.9% on the STL-10 test set.

git clone https://github.com/spijkervet/SimCLR.git && cd SimCLR
wget https://github.com/Spijkervet/SimCLR/releases/download/1.2/checkpoint_100.tar
sh setup.sh || python3 -m pip install -r requirements.txt || exit 1
conda activate simclr
python -m testing.logistic_regression with dataset=STL10 model_path=. epoch_num=100

CPU

wget https://github.com/Spijkervet/SimCLR/releases/download/1.1/checkpoint_100.tar -O checkpoint_100.tar
python -m testing.logistic_regression with model_path=. epoch_num=100 resnet=resnet18 logistic_batch_size=32

Results

These are the top-1 accuracy of linear classifiers trained on the (frozen) representations learned by SimCLR:

Method	Batch Size	ResNet	Projection output dimensionality	Epochs	Optimizer	STL-10	CIFAR-10
SimCLR + Linear eval.	256	ResNet50	64	100	Adam	0.829	0.833
SimCLR + Linear eval.	256	ResNet50	64	100	LARS	0.783	-
SimCLR + Linear eval.	256	ResNet18	64	100	Adam	0.765	-
SimCLR + Linear eval.	256	ResNet18	64	40	Adam	0.719	-
SimCLR + Linear eval.	512	ResNet18	64	40	Adam	0.71	-
Logistic Regression	-	-	-	40	Adam	0.358	0.389

Pre-trained models

ResNet (batch_size, epochs)	Optimizer	STL-10 Top-1
ResNet50 (256, 100)	Adam	0.829
ResNet18 (256, 100)	Adam	0.765
ResNet18 (256, 40)	Adam	0.719

python -m testing.logistic_regression with model_path=. epoch_num=100

Mixed-precision training

I am still evaluating the results, but using mixed-precision training allows you to train SimCLR on CIFAR-10 with ResNet50 and a batch size of 512 on a single 2080Ti (allocating ±11.2G). Use fp16: True in the config/config.yaml file to use mixed-precision training. This will yield slightly worse results.

MP results

ResNet50, 512 batch_size, O1: 0.7862

ResNet50, 512 batch_size, O2: 0.7797

LARS optimizer

The LARS optimizer is implemented in modules/lars.py. It can be activated by adjusting the config/config.yaml optimizer setting to: optimizer: "LARS". It is still experimental and has not been thoroughly tested.

What is SimCLR?

SimCLR is a "simple framework for contrastive learning of visual representations". The contrastive prediction task is defined on pairs of augmented examples, resulting in 2N examples per minibatch. Two augmented versions of an image are considered as a correlated, "positive" pair (x_i and x_j). The remaining 2(N - 1) augmented examples are considered negative examples. The contrastive prediction task aims to identify x_j in the set of negative examples for a given x_i.

Usage

Run the following command to setup a conda environment:

sh setup.sh
conda activate simclr

Or alternatively with pip:

pip install -r requirements.txt

Then, simply run:

python main.py

Testing

To test a trained model, make sure to set the model_path variable in the config/config.yaml to the log ID of the training (e.g. logs/0). Set the epoch_num to the epoch number you want to load the checkpoints from (e.g. 40).

python -m testing.logistic_regression

or in place:

python -m testing.logistic_regression with model_path=./logs/0 epoch_num=40

Configuration

The configuration of training can be found in: config/config.yaml. I personally prefer to use files instead of long strings of arguments when configuring a run. An example config.yaml file:

# train options
batch_size: 256
workers: 16
start_epoch: 0
epochs: 40

# model options
resnet: "resnet18"
normalize: True
projection_dim: 64

# loss options
temperature: 0.5

# reload options
model_path: "logs/0" # set to the directory containing `checkpoint_##.tar` 
epoch_num: 40 # set to checkpoint number

# logistic regression options
logistic_batch_size: 256
logistic_epochs: 100

Logging and TensorBoard

The sacred package is used to log all experiments into the logs directory. To view results in TensorBoard, run:

tensorboard --logdir logs

Optimizers and learning rate schedule

This implementation features the Adam optimizer and the LARS optimizer, with the option to decay the learning rate using a cosine decay schedule. The optimizer and weight decay can be configured in the config/config.yaml file.

Dependencies

torch
torchvision
tensorboard
sacred
pyyaml