bert_base_all_train.yaml

data:
    modality: "pose"
    
    train_pipeline:
        dataset:
            _target_: openhands.datasets.isolated.WLASLDataset
            split_file: /training_data/wlasl_new.json
            root_dir: /path/to/poses/
            splits: "train"
            modality: "pose"

        transforms:
            - PoseSelect:
                preset: mediapipe_holistic_minimal_27
            - CenterAndScaleNormalize:
                reference_points_preset: shoulder_mediapipe_holistic_minimal_27
                scale_factor: 1
            - ShearTransform:
                shear_std: 0.1
            - RotatationTransform:
                rotation_std: 0.1

        dataloader:
            _target_: torch.utils.data.DataLoader
            batch_size: 32
            shuffle: true
            num_workers: 6
            pin_memory: true
            drop_last: true
    
    
    valid_pipeline:
        dataset:
            _target_: openhands.datasets.isolated.WLASLDataset
            split_file: /training_data/wlasl_new.json
            root_dir: /path/to/poses/
            splits: "test"
            modality: "pose"

        transforms:
            - PoseSelect:
                preset: mediapipe_holistic_minimal_27
            - CenterAndScaleNormalize:
                reference_points_preset: shoulder_mediapipe_holistic_minimal_27
                scale_factor: 1

        dataloader:
            _target_: torch.utils.data.DataLoader
            batch_size: 32
            shuffle: true
            num_workers: 3
            pin_memory: true
            drop_last: false

model:
    encoder:
        type: pose-flattener
        params:
            num_points: 27
    decoder:
        type: param_bert
        params:
            max_position_embeddings: 121
            layer_norm_eps: 1e-12
            hidden_dropout_prob: 0.1
            hidden_size: 96
            num_attention_heads: 6
            num_hidden_layers: 3
            cls_token: true
        parameters: 
            [
                # "Handshape",
                # "Selected Fingers",
                # "Flexion",
                # "Spread",
                # "Spread Change",
                # "Thumb Position",
                # "Thumb Contact",
                # "Sign Type",
                # "Path Movement",
                # "Repeated Movement",
                # "Major Location",
                # "Minor Location",
                # "Second Minor Location",
                # "Contact",
                # "Nondominant Handshape", 
                # "Wrist Twist",
                # "Handshape Morpheme 2"
            ]

optim:
    loss: 'CrossEntropyLoss'
    optimizer:
        name: Adam
        params:
            lr: 1e-4

    scheduler:
        name: CosineAnnealingLR
        params:
            last_epoch: -1
            T_max: 10

trainer:
    gpus: 1
    max_epochs: 150
    # resume_from_checkpoint: /path/to/model.ckpt

exp_manager:
    create_tensorboard_logger: true
    create_wandb_logger: true
    wandb_logger_kwargs:
        name: model_name_here
        project: project_name_here

    create_checkpoint_callback: true
    checkpoint_callback_params:
        monitor: "val_acc"
        mode: "max"
        save_top_k: 1
        dirpath: "/pretrained_models/"

    early_stopping_callback: true
    early_stopping_params:
        monitor: "val_acc"
        patience: 80
        verbose: true
        mode: "max"