README.md

Setting up the Configurations

Here we describe the different parameters set in each configuration file:

• frame_dir: Directory where frames are stored.
• save_dir: Directory to save model checkpoints, predictions, processed datasets...
• store_mode: store if it's the first time running the script to prepare and store dataset information, or load to load previously stored information.
• batch_size: Batch size.
• clip_len: Length of the clips in number of frames.
• crop_dim: Integer value specifying the resolution (in pixels) to crop the frame into a square. Use -1 to indicate no cropping.
• dataset: Name of the dataset (soccernetball).
• event_team: Boolean indicating if detecting also team side in addition to spotting.
• radi_displacement: Radius of displacement used.
• epoch_num_frames: Number of frames used per epoch.
• feature_arch: Feature extractor architecture (rny002_gsf or rny008_gsf).
• learning_rate: Learning rate.
• mixup: Boolean indicating whether to use mixup or not.
• modality: Input modality used (rgb).
• num_classes: Number of classes for the current dataset.
• num_epochs: Number of epochs for training.
• warm_up_epochs: Number of warm-up epochs.
• start_val_epoch: Epoch where validation evaluation starts.
• temporal_arch: Temporal architecture used (ed_sgp_mixer).
• n_layers: Number of blocks/layers used for the temporal architecture.
• sgp_ks: Kernel size of the SGP and SGP-Mixer layers.
• sgp_r: $r$ factor in SGP and SGP-Mixer layers.
• only_test: Boolean indicating if only inference is performed or training + inference.
• criterion: Criterion used for validation evaluation (map, loss).
• num_workers: Number of workers.
• joint_train: Additional dataset information for joint training (SoccerNet Action Spotting -soccernet- in our case).
  • frame_dir: Directory where frames are stored (for additional dataset - snas).
  • dataset: Additional dataset used (soccernet).
  • num_classes: Nº of classes of the additional dataset.