In this repository we explore more principled methods for image masking
In your environment of choice, install the necessary requirements
!pip install -r requirements.txt
Create a config file that suits your machine:
cd ./config/user
cp abizeul_biomed.yaml myusername_mymachine.yaml
Adjust the paths to point to the directory you would like to use for storage of results and for fetching the data
To launch experiments, you can find a good example for training at ./script/jobs_euler_pcmae_random.sh and ./script/jobs_euler_eval_pcmae_random.sh for evaluation. Otherwise, you can run the following command to get started with training:
EXPERIMENT="mae_clevr"
python main.py user=myusername_mymachine experiment=$EXPERIMENT run_tag=$EXPERIMENT"
The EXPERIMENT variable refers the experimental setting defined in ./config/experiment/.
To evaluate a checkpoint, you can gain inspiration from ./config/user/callen_euler.yaml where I stored my runs. Then the following command gives an overview of how to launch the evaluation
EXPERIMENT="mae_clevr"
EPOCH=800
RUN_TAG="$EXPERIMENT_eval_$EPOCH"
python main.py user=callen_euler experiment=$DATASET trainer=eval checkpoint=pretrained checkpoint.epoch=$EPOCH run_tag=$RUN_TAG"
A change in the masking strategy should be reflected in ./dataset/dataloader.py file which define the image-masking pairs. At the moment, the mask only refers to a threshold; The change should also be reflected in ./model/module.py where each batch is masked and passed through the model