posec3d_kp_img.py

Go to the documentation of this file.

model = dict(
    type='ActionRecognitionRunner',
    backbone=dict(
        action_feat = dict(
            type='ResNet3d',
            in_channels = 17,
            base_channels = 64,
            stage_blocks = (4, 6, 3),
            out_indices = (2, ),
            spatial_strides = (2, 2, 2),
            temporal_strides = (1, 1, 2),
            dilations = (1, 1, 1),
            conv1_kernel = (1, 7, 7),
            conv1_stride_s = 1,
            conv1_stride_t = 1,
            pool1_stride_s = 1,
            pool1_stride_t = 1,
            inflate = (0, 1, 1),
            inflate_style = '3x1x1',
            input_key = 'pose_heatmap_for_action'
        ),
        pose_feat = dict(
            type='ResNet3d',
            in_channels = 17,
            base_channels = 32,
            stage_blocks = (4, 6, 3),
            out_indices = (2, ),
            spatial_strides = (2, 2, 2),
            temporal_strides = (1, 1, 2),
            dilations = (1, 1, 1),
            conv1_kernel = (1, 7, 7),
            conv1_stride_s = 1,
            conv1_stride_t = 1,
            pool1_stride_s = 1,
            pool1_stride_t = 1,
            inflate = (0, 1, 1),
            inflate_style = '3x1x1',
            input_key = 'pose_heatmap_for_pose'
        ),
        img_feat = dict(
            type = 'ResNet2d',
            in_channels = 768,
            stage_blocks = [2, 2, 2, 2],
            input_key = 'image_feature',
        ),
    ),
    fusion=dict(
        type='CatLayer',
        input_key = ['action_feat', 'pose_feat', 'img_feat'],
        input_type = ['3d', '3d', '2d'],
    ),
    head=dict(
        action_upper = dict(
            type='LinearHead',
            in_channels=1024+512+768,
            num_classes=6,
            dropout_ratio=0.5,
            input_key = 'fusion',
        ),
        action_lower = dict(
            type='LinearHead',
            in_channels=1024+512+768,
            num_classes=12,
            dropout_ratio=0.5,
            input_key = 'fusion',
        ),
        pose = dict(
            type='MLPHead',
            in_channels=1024+512+768,
            num_classes=8,
            layer_channels=[2048],
            dropout_ratio=0.5,
            input_key = 'fusion',
        ),
        hand = dict(
            type='MLPHead',
            in_channels=1024+512+768,
            num_classes=4,
            layer_channels=[2048],
            dropout_ratio=0.5,
            input_key = 'fusion',
        ),
        foot = dict(
            type='MLPHead',
            in_channels=1024+512+768,
            num_classes=2,
            layer_channels=[2048],
            dropout_ratio=0.5,
            input_key = 'fusion',
        ),
        aux_action_upper = dict(
            type='I3DHead',
            in_channels=1024,
            num_classes=6,
            dropout_ratio=0.5,
            input_key = 'action_feat',
        ),
        aux_action_lower = dict(
            type='I3DHead',
            in_channels=1024,
            num_classes=12,
            dropout_ratio=0.5,
            input_key = 'action_feat',
        ),
        aux_pose = dict(
            type='MLPHead',
            in_channels=512,
            num_classes=8,
            layer_channels=[1024],
            dropout_ratio=0.5,
            input_key = 'pose_feat',
            input_type = '3d',
        ),
        aux_hand = dict(
            type='MLPHead',
            in_channels=512,
            num_classes=4,
            layer_channels=[1024],
            dropout_ratio=0.5,
            input_key = 'pose_feat',
            input_type = '3d',
        ),
        aux_foot = dict(
            type='MLPHead',
            in_channels=512,
            num_classes=2,
            layer_channels=[1024],
            dropout_ratio=0.5,
            input_key = 'pose_feat',
            input_type = '3d',
        ),
    ),
    predict_keys = dict(
        #예측한 라벨의 키 = 헤드의 키(스코어의 키)
        pred_action_upper = 'action_upper',
        pred_action_lower = 'action_lower',
        pred_pose = 'pose',
        pred_hand = 'hand',
        pred_foot = 'foot',
    )
)
 
score_keys = ['action_upper', 'action_lower', 'pose', 'hand', 'foot', 'aux_action_upper', 'aux_action_lower', 'aux_pose', 'aux_hand', 'aux_foot']
pred_keys = ['pred_action_upper', 'pred_action_lower', 'pred_pose', 'pred_hand', 'pred_foot']
gt_keys = ['gt_action_upper', 'gt_action_lower', 'gt_pose', 'gt_hand', 'gt_foot', 'gt_action_upper', 'gt_action_lower', 'gt_pose', 'gt_hand', 'gt_foot']
target_tasks = ['action_upper', 'action_lower', 'pose', 'hand', 'foot', 'action_upper', 'action_lower', 'pose', 'hand', 'foot']
train_tasks = ['action_upper', 'action_lower', 'pose', 'hand', 'foot']
 
 
loss = dict(
    MutiTaskSigmoidFocalLoss = dict(
        weight = 1.0,
        task_key = 'category',
        pred_keys = score_keys,
        gt_keys = gt_keys,
        target_tasks = target_tasks,
        weights = [1.0, 1.0, 1.0, 1.0, 1.0, 0.25, 0.25, 0.25, 0.25, 0.25],
        # data_num = dict(action = [100, 100, 100, 24, 100, 34, 3, 100, 100], pose = [98, 517, 11, 35])
    )
)
 
metric_args = dict(pred_key=pred_keys, 
                   gt_key=gt_keys[:5], 
                   target_tasks=target_tasks[:5],
                   task_key='category')
 
collect_keys = ['pose_heatmap_for_action', 'pose_heatmap_for_pose', 'image_feature', 'gt_action_upper', 'gt_action_lower', 'gt_pose', 'gt_hand', 'gt_foot']
 
data_loader = dict(
    type = 'ActionDatasetLoader_mtml',
    data_folder = '/media/safemotion/HDD5/pjm_test/action_train_dataset_2023/action_kp_imf_split',
    category_info = dict(action_upper = 6,
                         action_lower = 12,
                         pose = 8,
                         hand = 4,
                         foot = 2),
    clip_len_action = 20,
    clip_len_pose = 6,
)
ep_mul = 20
train = dict(
    num_workers = 8,
    init_lr = 0.1,
    batch_size = 32,
    epochs = 100*ep_mul,
    optimizer = 'SGD',
    optimizer_args = dict(momentum=0.9, nesterov=True, weight_decay=0.0001),
    scheduler = 'CosineAnnealingLR',
    # scheduler = 'StepLR',
    scheduler_args = dict(T_max=30*ep_mul, eta_min=0),
    adjust_lr_epoch = [10*ep_mul, 50*ep_mul, 100*ep_mul, 130*ep_mul],
    adjust_lr_rate = [0.5, 0.1, 0.1, 0.1],
    val_interval = 1,
 
    
    update_loss_weight = False,
    update_loss_weight_interval = 20,
    base_weight = 0.5,
 
    pretrained = None,#'/media/safemotion/HDD5/pjm_test/action_train_test/9.pth',
    save_root = '/media/safemotion/HDD5/pjm_test/action_train_result/action_kp_img',
 
)
 
test = dict(
    model_path = None,
    save_root = '',
)