This project includes the following folders:
- code: Contains all the code.(forked and modified from here)
- Compiled Results: Contains the test accuracies for the model trained on CIFAR and SVHN dataset.
- logs: Contains all the logs generated during the training.
- plots: Contains the plots for the datapoints gathered in compiled results.
To run the code, follow these steps:
- You can download the CIFAR-10 dataset from here
- You can download the SVHN dataset from these links: train test
- CUB-200-2011 dataset can be downloaded from here. You can preprocess the data using the
split.pyprovided in the code folder. - After downloading the data place the downloaded files in a data directory of your choice.
Pretrain the ResNet-50 model on Imagenet or download the Imagenet pretrained models from pytorch.models in your checkpoint folder, e.g. ./ckpt/. Then rename the pretrained checkpoint as resnet_50_1.pth.
This can be used for CUB-200-2011 datset.(Note: Training on this dataset requires around 27Gb of GPU memory.)
To train the model, run main.py given in code folder.
Example of CIFAR-10 using AKC and ARC as both 0.5.(lambda_kd represents AKC weight and lambda_mmd represents the ARC weight)
$data_root is the path to your downloaded data.
$dataset is the dataset you want to use.(cifar10 or svhn)
$num_labels is the number of labelled data you want for training
$pretrain_path is the path of the directory where pretrained models are downloaded.
python -u main.py --data_root $data_root --dataset $dataset --num_labels $num_labels --pretrained_weight_path $pretrain_path --model wideresnetleaky --depth 28 --widen_factor 2 --l_batch_size 64 --ul_batch_size 448 --lambda_mmd 0.5 --lambda_kd 0.5 --lr 0.0001 --weight_decay 0.0005 --epochs 10 --coef 1.0 --alg pl --strong_aug true --threshold 0.95 --ema_teacher true --ema_teacher_factor 0.999 --bn_momentum 0.1 --interleave 0 --seed 10