Training on custom dataset

[williamhoole]

Is it possible to train/ fine tune the model on a custom dataset?

[orhir]

Of course it is. You will need to change the config file to refer to the new annotation file, by changing the 'data' var in the config.

for example:

data = dict(
    samples_per_gpu=16,
    workers_per_gpu=8,
    train=dict(
        type='TransformerPoseDataset',
        ann_file=f'NEW_ANNOTATION.json',
        img_prefix=f'{NEW_IMAGE_ROOT/',
        # img_prefix=f'{data_root}',
        data_cfg=data_cfg,
        valid_class_ids=None,
        max_kpt_num=channel_cfg['max_kpt_num'],
        num_shots=1,
        pipeline=train_pipeline),

Keep in mind the new annotation should have similar structure as the original MP100 dataset annotation.