{
    "meta": {
        "model_name": "pose-centric-multimodal-har",
        "version": "v1.0",
        "description": "Pose-centric multimodal temporal HAR model"
    },
    "pose_model": {
        "backbone": "PoseFormerFactorized",
        "joints": 17,
        "in_channels": 3,
        "hidden_dim": 128,
        "layers": 4,
        "heads": 8,
        "max_seq_len": 30,
        "use_sinusoidal_pe": false,
        "enable_temporal_attention": true,
        "return_tokens": true
    },
    "long_term_model": {
        "enabled": true,
        "type": "transformer_encoder",
        "layers": 2,
        "heads": 8,
        "dropout": 0.1
    },
    "image_model": {
        "enabled": true,
        "backbone": "resnet18",
        "feature_dim": 128,
        "pretrained": true
    },
    "fusion_model": {
        "enabled": true,
        "mode": "concat",
        "output_dim": 128
    },
    "classifier": {
        "num_classes": 500
    },
    "regularization": {
        "short_term_dropout": 0.3,
        "long_term_dropout": 0.1
    },
    "precision": "fp32"
}