A short explanation about the config file:
{
"name": the name of the config
"dataset": the name of the dataset
"categories": A magic number for visual embeddings categories, to include colors and shapes
"max_vis_features": number of visual features per input (~ nubmer of visual objects appear in the figure)
"max_seq_len": number of text tokens per input (~questions length)
"binary_answers": A flag for a binary answers dataset
"main_folder": the FULL PATH to the CRCT folder (../)
"model_config": ViLBERT parameters,
"save_path": path to folder where to save experiments,
"figure_feat_path": a path to the figure features that was extracted from the detector.,
"qa_parent_dir": a path to the Q&As files folder
"tensorboard": a path to where to save tensorboard informations,
"checkpoints_dir": pretrained checkpoints to start from
"dataset_files_divisions": this is a dictionary that contains the number of images appear in each figure features file.
"splits": the name of the dataset's splits
}