Follow the below instructions to replicate the results obtained by us.

Install the libraries and dependencies from requirements.txt by running the following command.

pip install -r requirements.txt

Run the below commands to replicate the results obtained by us.

Run the below command to train a Wide Resnet to learn image representations. This makes use of image augmentations and loss as described in the Barlow Twins Paper. To resume training from a checkpoint, set train-from-start = 0 and provide a valid path for the parameter checkpoint-path. Specify wide = 0 to train it on Resnet-18 instead of Wide Resnet-50.

python train_barlow.py \
--checkpoint-path $SCRATCH/checkpoints/model_barlow.pth.tar \
--num-epochs 500 \
--batch-size 512 \
--train-from-start 1 \
--dataset-folder /dataset \
--learning-rate 0.05 \
--lambd 0.05 \
--weight-decay 0.000015 \
--wide 1

Run the below command to train a classifier on top of the barlow twins backbone. Provide the path of the previously trained model to the parameter transfer-path. The output of this model will be saved in the path provided in checkpoint-path and the model with best accuracy on validation set will be saved in the path provided in best-path.

python train_transfer.py \
--checkpoint-path $SCRATCH/checkpoints/model_transfer.pth.tar \
--transfer-path $SCRATCH/checkpoints/model_barlow.pth.tar \
--best-path $SCRATCH/checkpoints/model_transfer_best.pth.tar \
--num-epochs 160 \
--batch-size 32 \
--fine-tune 1 \
--dataset-folder /dataset \
--learning-rate-classifier 0.00001 \
--learning-rate-barlow 0.00001 \
--weight-decay 0.001 \
--wide 1 \
--model-name barlow \
--dropout 0.1 \
--seed 10

Run the below command to fine tune the model trained by the previous method using consistency regularization. As done previously, provide the path of the previously trained model to the parameter transfer-path. The output of this model will be saved in the path provided in checkpoint-path and the model with best accuracy on validation set will be saved in the path provided in best-path.

python train_fixmatch_transfer.py \
--checkpoint-path $SCRATCH/checkpoints/model_fm_transfer.pth.tar \
--transfer-path $SCRATCH/checkpoints/model_transfer_best.pth.tar \
--best-path $SCRATCH/checkpoints/model_transfer_fm_best.pth.tar \
--num-epochs 400 \
--num-steps 200 \
--train-from-start 1 \
--batch-size 64 \
--fine-tune 1 \
--dataset-folder /dataset \
--learning-rate 0.00001 \
--mu 7 \
--threshold 0.9 \
--lambd 1 \
--momentum 0.9 \
--weight-decay 0.001

The model saved at the best-path will be the final model.

Run the below code to get file names of the images we need to request new labels for. For this, the following steps have to followed.

Run the below command to get image representations for both labelled and unlabeled dataset. Provide the necessary output path to the parameter out-path.

python label_request/get_img_representations.py \
--checkpoint-path $SCRATCH/checkpoints/model_transfer_barlow_best.pth.tar \
--out-path $SCRATCH/representations/ \
--batch-size 512 \
--dataset-folder /dataset \
--wide 1 

Run the below command to get the image ids using the processes defined in the paper. Provide the previously generated representations path to representations-path and the label outputs will be saved in dest-path.

python label_request/get_label_samples.py \
--representations-path $SCRATCH/representations/ \
--dest-path $SCRATCH/label_request/

Run the below command to generate the new dataset.

python label_request/create_new_dataset.py \
--label-request-path $SCRATCH/label_request/ \
--dataset /dataset/