finetune.md

Tutorial 1: Finetuning Models

Detectors pre-trained on the COCO dataset can serve as a good pre-trained model for other datasets, e.g., PoseTrack Dataset. This tutorial provides instruction for users to use the models provided in the Model Zoo for other datasets to obatin better performance.

There are two steps to finetune a model on a new dataset.

• Add support for the new dataset following Tutorial 2: Adding New Dataset.
• Modify the configs as will be discussed in this tutorial.

To finetune on the custom datasets, the users need to modify four parts in the config.

Modify Model

Then the new config needs to modify the model according to the keypoint numbers of the new datasets. By only changing out_channels in the keypoint_head.

# model settings
model = dict(
    type='TopDown',
    pretrained=None,
    backbone=dict(type='ResNet', depth=18),
    keypoint_head=dict(
        type='TopDownSimpleHead',
        in_channels=512,
        out_channels=17,
    ),
    train_cfg=dict(),
    test_cfg=dict(
        flip_test=True,
        post_process='unbiased',
        shift_heatmap=True,
        modulate_kernel=11),
    loss_pose=dict(type='JointsMSELoss', use_target_weight=False))

Modify dataset

The users may also need to prepare the dataset and write the configs about dataset. MMPose already support COCO and MPII-TRB Dataset.

Modify training schedule

The finetuning hyperparameters vary from the default schedule. It usually requires smaller learning rate and less training epochs

# optimizer
optimizer = dict(
    type='Adam',
    lr=1e-4,
)
optimizer_config = dict(grad_clip=None)
# learning policy
lr_config = dict(
    policy='step',
    warmup='linear',
    warmup_iters=500,
    warmup_ratio=0.001,
    step=[10, 15])
total_epochs = 20

Use pre-trained model

To use the pre-trained model, the new config add the link of pre-trained models in the load_from. The users might need to download the model weights before training to avoid the download time during training.

load_from = 'resnet50_coco.pth'  # noqa