We evaluate the uncertainty of the model with different sampling techniques ['ratio', 'margin', 'least', 'entropy', 'random'].
CNN and ResnetX have been implemented. We use wanb to monitor the training. You can use CIFAR-10, EMNIST or MNIST data.
We train an initial model on 50% of the original dataset. Pretending that 50% of the dataset is hidden with the parameter split_train_val_ratio=0.5.
python3 training/run_experiment.py --model_class=CNN --data_class=MNIST --max_epochs=5 --gpus=1
python3 training/run_experiment.py --model_class=Resnet18 --data_class=CIFAR10DataModule --max_epochs=15 --gpus=1
(if you use CNN you need to set manully the image size as global variable IMAGE_SIZE to 32 for CIFAR10 and 28 for EMNIST)
We apply the uncertain scoring techniques and save the highest uncertainty images. Change the arguments main.
python3 training/create_new_set.py --model_class=Resnet18 --data_class=CIFAR10DataModule
python3 training/create_new_set.py --model_class=Resnet18 --data_class=CIFAR10DataModule