mtr+20_percent_data.yaml

DATA_CONFIG:
    DATASET: WaymoDataset
    OBJECT_TYPE: &object_type ['TYPE_VEHICLE', 'TYPE_PEDESTRIAN', 'TYPE_CYCLIST']
    
    DATA_ROOT: 'data/waymo'
    SPLIT_DIR: {
        'train': 'processed_scenarios_training', 
        'test': 'processed_scenarios_validation'
    }

    INFO_FILE: {
        'train': 'processed_scenarios_training_infos.pkl', 
        'test': 'processed_scenarios_val_infos.pkl'
    }

    SAMPLE_INTERVAL: {
        'train': 5, 
        'test': 5
    }

    INFO_FILTER_DICT: 
        filter_info_by_object_type: *object_type

   # for map feature encoding
    POINT_SAMPLED_INTERVAL: 1
    NUM_POINTS_EACH_POLYLINE: 20
    VECTOR_BREAK_DIST_THRESH: 0.5

    NUM_OF_SRC_POLYLINES: 768
    CENTER_OFFSET_OF_MAP: &center_offset [30.0, 0]


MODEL:
    CONTEXT_ENCODER:
        NAME: MTREncoder

        NUM_OF_ATTN_NEIGHBORS: 16
        NUM_INPUT_ATTR_AGENT: 29
        NUM_INPUT_ATTR_MAP: 9
        
        NUM_CHANNEL_IN_MLP_AGENT: 256
        NUM_CHANNEL_IN_MLP_MAP: 64
        NUM_LAYER_IN_MLP_AGENT: 3
        NUM_LAYER_IN_MLP_MAP: 5
        NUM_LAYER_IN_PRE_MLP_MAP: 3

        D_MODEL: 256
        NUM_ATTN_LAYERS: 6
        NUM_ATTN_HEAD: 8 
        DROPOUT_OF_ATTN: 0.1 

        USE_LOCAL_ATTN: True
        USE_PERFORMER: True
        USE_LIDAR: True

    MOTION_DECODER:
        NAME: MTRDecoder

        OBJECT_TYPE: *object_type 
        CENTER_OFFSET_OF_MAP: *center_offset

        NUM_FUTURE_FRAMES: 80
        NUM_MOTION_MODES: 6

        INTENTION_POINTS_FILE: data/waymo/cluster_64_center_dict.pkl

        D_MODEL: 512
        NUM_DECODER_LAYERS: 6 
        NUM_ATTN_HEAD: 8
        MAP_D_MODEL: 256
        DROPOUT_OF_ATTN: 0.1 

        NUM_BASE_MAP_POLYLINES: 256
        NUM_WAYPOINT_MAP_POLYLINES: 128

        LOSS_WEIGHTS: {
            'cls': 1.0, 
            'reg': 1.0, 
            'vel': 0.5
        }

        NMS_DIST_THRESH: 2.5


OPTIMIZATION:
    BATCH_SIZE_PER_GPU: 10
    NUM_EPOCHS: 30

    OPTIMIZER: AdamW
    LR: 0.0001
    WEIGHT_DECAY: 0.01

    SCHEDULER: lambdaLR
    DECAY_STEP_LIST: [22, 24, 26, 28]
    LR_DECAY: 0.5
    LR_CLIP: 0.000001

    GRAD_NORM_CLIP: 1000.0