relic-pytorch 0.1.2

Simple self-supervised contrastive based on based on the ReLIC method

ReLIC

A PyTorch implementation of a computer vision self-supervised learning method based on Representation Learning via Invariant Causal Mechanisms (ReLIC).

This simple approach is very similar to BYOL and SimCLR. The training technique uses a online and target encoder (EMA) with a simple critic MLP projector, while the instance discrimination loss function resembles the contrastive loss used in SimCLR. The other half of the loss function acts as a regularizer - it includes an invariance penalty, which forces the representations to stay invariant under data augmentations and amplifies intra-class distances.

Results

First, we learned features using SimCLR on the STL10 unsupervised set. Then, we train a linear classifier on top of the frozen features from SimCLR. The linear model is trained on features extracted from the STL10 train set and evaluated on the STL10 test set.

Models are first trained on training subsets - for CIFAR10 50,000 and for STL10 100,000 images. For evaluation, I trained and tested LogisticRegression on:

1. CIFAR10 - 50,000 train images on 10,000 test images.
2. STL10 - features were learned on 100k unlabeled images. LogReg was trained on 5k train images and evaluated on 8k test images.

Linear probing were evaluated on features extracted from encoders using the scikit LogisticRegression model.

More detailed evaluation steps and results for CIFAR10 and STL10 can be found in the notebooks directory.

Evaulation model	Dataset	Feature Extractor	Encoder	Feature dim	Projection Head dim	Epochs	Top1 %
LogisticRegression	CIFAR10	ReLIC	ResNet-18	512	64	100	71.07
LogisticRegression	STL10	ReLIC	ResNet-18	512	64	100	76.10
LogisticRegression	STL10	ReLIC	ResNet-50	2048	64	100	80.40

Usage

Instalation

$ pip install relic-pytorch

Code currently supports ResNet18, ResNet50 and an experimental version of the EfficientNet model. Supported datasets are STL10 and CIFAR10.

All training is done from scratch.

Examples

CIFAR10 ResNet-18 model was trained with this command:

relic_train --dataset_name "cifar10" --encoder_model_name resnet18 --fp16_precision --tau 5 --gamma 0.99 --alpha 1.0

STL10 ResNet-50 model was trained with this command:

relic_train --dataset_name "stl10" --encoder_model_name resnet50 --fp16_precision

Detailed options

Once the code is setup, run the following command with optinos listed below: relic_train [args...]⬇️

ReLIC

options:
  -h, --help            show this help message and exit
  --dataset_path DATASET_PATH
                        Path where datasets will be saved
  --dataset_name {stl10,cifar10}
                        Dataset name
  -m {resnet18,resnet50,efficientnet}, --encoder_model_name {resnet18,resnet50,efficientnet}
                        model architecture: resnet18, resnet50 or efficientnet (default: resnet18)
  -save_model_dir SAVE_MODEL_DIR
                        Path where models
  --num_epochs NUM_EPOCHS
                        Number of epochs for training
  -b BATCH_SIZE, --batch_size BATCH_SIZE
                        Batch size
  -lr LEARNING_RATE, --learning_rate LEARNING_RATE
  -wd WEIGHT_DECAY, --weight_decay WEIGHT_DECAY
  --fp16_precision      Whether to use 16-bit precision GPU training.
  --proj_out_dim PROJ_OUT_DIM
                        Projector MLP out dimension
  --proj_hidden_dim PROJ_HIDDEN_DIM
                        Projector MLP hidden dimension
  --log_every_n_steps LOG_EVERY_N_STEPS
                        Log every n steps
  --gamma GAMMA         Initial EMA coefficient
  --tau TAU             Softmax temperature
  --alpha ALPHA         Regularization loss factor
  --update_gamma_after_step UPDATE_GAMMA_AFTER_STEP
                        Update EMA gamma after this step
  --update_gamma_every_n_steps UPDATE_GAMMA_EVERY_N_STEPS
                        Update EMA gamma after this many steps

Citation

@misc{mitrovic2020representation,
      title={Representation Learning via Invariant Causal Mechanisms}, 
      author={Jovana Mitrovic and Brian McWilliams and Jacob Walker and Lars Buesing and Charles Blundell},
      year={2020},
      eprint={2010.07922},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}