What is the difference of pretrain method in 20-minutes Tour to MMPose?

[seon-creator]

Hello, I'm using mmpose for detection keypoint project.
I'm studying to understand more about mmpose usage.
When I read the description about model Backbone part, I checked there are two method of using pretrain model.

So I tried both method on same model with my custom data.
My custom data has 5-keypoint, and I used 2d body model.
This two plots are the train and validation result of loss and AP.

usign method 1

using method 2

The first plot is the result of applying this method.

init_cfg=dict(
    type='Pretrained',
    checkpoint='PATH/TO/YOUR_MODEL_WEIGHTS.pth'),

The second plot is the result of applying this method.

init_cfg=dict(
    type='Pretrained',
    prefix='backbone.',
    checkpoint='PATH/TO/YOUR_CHECKPOINT.pth'),

After then, I compared the accuracy of the keypoint. The result of method 2 was better than 1.
I wonder why the method 2's performance was better than 1.
Is there any idea of this?

[Tau-J]

Hi @seon-creator , prefix='backbone.', means only load pretrained backbone's weights into the model for training, which is the standard usage for transfer learning.

[seon-creator]

Thanks for the quick response!
So, Is that mean that the first method loads the entire model including the backbone?

init_cfg=dict(
    type='Pretrained',
    checkpoint='PATH/TO/YOUR_MODEL_WEIGHTS.pth'),

[Tau-J]

Yes