Is there an existing issue for this?

• I have searched the existing issues

Operating System

Mac OS Sequoia 15.6.1

DeepLabCut version

3.0.0rc10

What engine are you using?

pytorch

DeepLabCut mode

single animal

Device type

Builtin GPU (metal)

Bug description 🐛

Confidence values are not constrained between 0 and 1. For example, on one model the confidence values are mostly between 2 and 4, and on another they mostly stay below 1 but spike to over 12. My expectation is that these values represent a range from 0 (not at all confident) to 1 (100% confident), as is the case for the standard resnet model.

Here are some likelihood plots showing the issue:

Steps To Reproduce

1. Create a new environment with deeplabcut==3.0.0rc10
2. Create a new project and label data
3. Create the training dataset with an rtmpose model, like deeplabcut.create_training_dataset(config_path, net_type="rtmpose_x")
4. Train the model and analyze the data

Relevant log output

Anything else?

The mm pose implementation has a similar issue (with one potential solution in the comments): open-mmlab/mmpose#2755
This article explains where the confidence scores are derived from in the rtmpose architecture: https://deepwiki.com/Tau-J/rtmlib/4.2.1-rtmpose#simcc-postprocessing

The softmax solution in the first file seems correct, but I haven't dug into the guts of the deeplabcut implementation to see where to apply it.

EDIT:
It appears there is a bool to apply the softmax function in the SimCCPredictor, and an option to normalize. Looking at my config, it's possible the normalize: false line is the issue here:

  heads:
    bodypart:
      type: RTMCCHead
      weight_init: RTMPose
      target_generator:
        type: SimCCGenerator
        input_size:
        - 384
        - 384
        smoothing_type: gaussian
        sigma:
        - 6.93
        - 6.93
        simcc_split_ratio: 2.0
        label_smooth_weight: 0.0
        normalize: false
      criterion:
        x:
          type: KLDiscreteLoss
          use_target_weight: true
          beta: 10.0
          label_softmax: true
        y:
          type: KLDiscreteLoss
          use_target_weight: true
          beta: 10.0
          label_softmax: true
      predictor:
        type: SimCCPredictor
        simcc_split_ratio: 2.0

Code of Conduct

• I agree to follow this project's Code of Conduct