IJDAR2020_code3

Documentation

Note, all the scripts have a -h for help on their arguments.

First generate a dataset

In make_dataset.py, check the dataset parameters:

• nb_papyri : number of papyri to use in the dataset
• patch_size : size of the patches in pixels (64)
• data_path : path to the directory containing the full images (of fragments)
• test_proportion : proportion of papyri that will be put aside in the test set (for example, with nb_papyri = 100 and test_proportion= 0.1, 10 papyri will go in the test set) (0.1)
• max_patches_per_fragments : maximum number of patches that will be extracted for each fragment image (5)
• fragmentAsClass : if set to True, the training and validation set will be prepared in self_supervised_learning mode. If set to False, the ground truth will be used to prepare the training and validation sets (normal mode). The Test set is always in normal mode, so if no ground truth is available, test_proportion should be set to 0
• resize : if set, resizes all the images by this value. (None)

Depending on the name formatting of your own files, you should change the content of the function get_papyrus_id() such that given an image name, the class id is returned (when the fragmentAsClass parameter is set to False)

Once all the parameters are correctly set, run :

$ python3 make_dataset.py -d <datasetDir> -n <datasetName>

Then start training

In train.py, check the training hyperparameters, then run :

$ python3 train.py -g <gpu_id> -da <path_to_dataset> -d <save_directory>

Using Tensorboard to monitor training

$ tensorboard --logdir="<path_to_logs_directory>"

Finally, evaluate the model

Check the evaluation hyperparameters, then run :

$ python3 evaluate.py -g <gpu_id> -sa <save_directory> -m <model_directory> -da <path_to_dataset> -e <experiment_name> -s <epoch_step>

Depending on the number of papyri in the test set, this can take a while