Dataset was downloaded from cvwc2019challenge.
Run below command to preprocess the data. The following preprocessing function, flips the images and keypoints and saves the results to image folder, csv and json files. And then extracts the parts for all the images and stores in the designated parts folder. Finally it creates 4-folds of the dataset.
python preprocessing.py
All the paths to save and retrieve the files can be changed in config.ini file
We submitted our training jobs on gaivi server. Bash script is available in trainer.sh. We used 4-fold cross validation for training. To run for a respective fold, change the --fold argument in trainer.sh as shown below.
srun python -u ppbm_fold.py --fold=0 > fold0.out
srun python -u ppbm_fold.py --fold=1 > fold1.out
srun python -u ppbm_fold.py --fold=2 > fold2.out
srun python -u ppbm_fold.py --fold=3 > fold3.out
Submit the job by running below command:
#sbatch trainer.sh
Weights can be downloaded from the following link weights.
Run the below command to evaluate the model
python testing.py
python evaluation_script.py