PyTorch implementation of SimCLR: A Simple Framework for Contrastive Learning of Visual Representations

Check out the Blog post with full documentation: Exploring SimCLR: A Simple Framework for Contrastive Learning of Visual Representations

Config file

Before runing SimCLR, make sure you choose the correct running configurations on the config.yaml file.

batch_size: 256 # A batch size of N, produces 2 * (N-1) negative samples. Original implementation uses a batch size of 8192
out_dim: 64 # Output dimensionality of the embedding vector z. Original implementation uses 2048
s: 1
temperature: 0.5 # Temperature parameter for the contrastive objective
base_convnet: "resnet18" # The ConvNet base model. Choose one of: "resnet18 or resnet50". Original implementation uses resnet50
use_cosine_similarity: True # Distance metric for contrastive loss. If False, uses dot product
epochs: 40 # Number of epochs to train
num_workers: 4 # Number of workers for the data loader

Feature Evaluation

Feature evaluation is done using a linear model protocol. Feature are learnt using the STL10 unsupervised set and evaluated in the train/test splits;

Check the feature_eval/FeatureEvaluation.ipynb notebook for reproducebility.

Feature Extractor	Method	Architecture	Top 1
Logistic Regression	PCA Features	-	36.0%
KNN	PCA Features	-	31.8
Logistic Regression	SimCLR	ResNet-18	71.8%
KNN	SimCLR	ResNet-18	66.7%