How to use a pretrained model in a 2d human keypoint custom dataset with only 8 points?

[Icksir]

I have a custom dataset that is only composed by 8 keypoints:
head neck left_shoulder right_shoulder left_elbow right_elbow right_hand left_hand

An example is this image

And I created the JSON with COCO format with only a few images just to try training any model with Coco Annotator. This is an example for the first image.

'images': {'annotated': True,
 'annotating': [],
 'category_ids': [1],
 'dataset_id': 12,
 'events': [{'_cls': 'SessionEvent',
             'milliseconds': 14842,
             'tools_used': [],
             'user': 'ricky'}],
 'file_name': 'frame_2.jpg',
 'height': 1080,
 'id': 8,
 'is_modified': False,
 'metadata': {},
 'milliseconds': 14842,
 'num_annotations': 1,
 'path': '/datasets/train/frame_2.jpg',
 'regenerate_thumbnail': False,
 'width': 960}

'categories': {'color': '#dd7e6a',
 'creator': 'ricky',
 'id': 1,
 'keypoint_colors': ['#9c73bf', '#ff1a38', '#736322', '#33fff1', '#3369ff',
                     '#9d13bf', '#733941', '#ffb499'],
 'keypoints': ['head', 'neck', 'left_shoulder', 'right_shoulder', 'left_elbow',
               'right_elbow', 'left_hand', 'right_hand'],
 'metadata': {},
 'name': 'person',
 'skeleton': [[1, 2], [2, 3], [2, 4], [3, 5], [4, 6], [5, 7], [6, 8]],
 'supercategory': ''}

'annotation': {'area': 0,
 'bbox': [0, 0, 0, 0],
 'category_id': 1,
 'color': '#eb4f6c',
 'creator': 'ricky',
 'dataset_id': 12,
 'events': [{'_cls': 'SessionEvent',
             'created_at': {'$date': 1710105025694},
             'milliseconds': 11469,
             'tools_used': ['Polygon', 'Keypoints', 'Select'],
             'user': 'ricky'}],
 'height': 1080,
 'id': 14,
 'image_id': 8,
 'isbbox': False,
 'iscrowd': False,
 'keypoints': [409, 326, 2, 406, 450, 2, 368, 572, 2, 418, 469, 2, 407, 799, 2,
               445, 577, 2, 547, 879, 2, 540, 577, 2],
 'metadata': {},
 'milliseconds': 11469,
 'num_keypoints': 8,
 'segmentation': [],
 'width': 960}

The problem is that I have no idea how to make it work to train the head of a pretrained model that only outputs this 8 points.

What I did

I followed MMPose_Tutorial.ipynb located in configs/demo. I thought it would work if I made some changes with the config located in configs/body_2d_keypoint/topdown_heatmap/coco/td-hm_hrnet-w32_8xb64-210e_coco-256x192.py.

The things I have changed are:

# Change the root for the json, because they are located in /home/icksir/Documentos/safetymind/coco-annotator/train.json 
# and /home/icksir/Documentos/safetymind/coco-annotator/test.json
cfg.data_root = '/home/icksir/Documentos/safetymind/coco-annotator/'

# I only have like 6 images so I used a really small batch size
cfg.train_dataloader.batch_size = 2
cfg.val_dataloader.batch_size = 2
cfg.test_dataloader.batch_size = 2

# I changed the dataset configs to match my path and indicated that I use CocoDataset
# The images path are in /home/icksir/Documentos/safetymind/coco-annotator/datasets/train/*.jpg
cfg.dataset_type = 'CocoDataset'
cfg.train_dataloader.dataset.type = cfg.dataset_type
cfg.train_dataloader.dataset.ann_file = 'train.json'
cfg.train_dataloader.dataset.data_root = cfg.data_root
cfg.train_dataloader.dataset.data_prefix = dict(img='datasets/train/')

# The images path are in /home/icksir/Documentos/safetymind/coco-annotator/datasets/test/*.jpg
cfg.val_dataloader.dataset.type = cfg.dataset_type
cfg.val_dataloader.dataset.bbox_file = None
cfg.val_dataloader.dataset.ann_file = 'test.json'
cfg.val_dataloader.dataset.data_root = cfg.data_root
cfg.val_dataloader.dataset.data_prefix = dict(img='datasets/test/')

cfg.test_dataloader.dataset.type = cfg.dataset_type
cfg.test_dataloader.dataset.bbox_file = None
cfg.test_dataloader.dataset.ann_file = 'test.json'
cfg.test_dataloader.dataset.data_root = cfg.data_root
cfg.test_dataloader.dataset.data_prefix = dict(img='datasets/test/')

# I indicated that the head must only output 8 channels and the backbone has to have pretrained weights
cfg.model.head.out_channels = 8
cfg.model.backbone.init_cfg['prefix'] = 'backbone.'

What is my error

In the runner.train() function that is created by Runner.from_cfg(cfg), I have one error:

   in BaseDataset._serialize_data(self)
      np.ndarray = np.cumsum(address_list)
      # TODO Check if np.concatenate is necessary
-->  data_bytes = np.concatenate(data_list)
      # Empty cache for preventing making multiple copies of
      # `self.data_info` when loading data multi-processes.
      self.data_list.clear()

File <__array_function__ internals>:200, in concatenate(*args, **kwargs)

ValueError: need at least one array to concatenate

I would appreciate any help or how could I approach this dataset.