Natasha

You'd need to write a custom dataset class. The important functions for custom dataset classes are __len__ and __getitem__, so make sure you've got those in there and you can...

Came here to say the same thing! Only issue is there's no option to donate money to the developer!

@cir7 Furthermore, it would be helpful to include a conversion script from COCO format outputs obtained with MMPose to this MMAction pkl format.

> I don't think that we need a loss during this step. The loss information is useful to monitor overfitting.

What would the 18th heatmap represent? If it's a joint not defined in the annotations you may need to make a custom [HeatmapGenerator](https://github.com/HRNet/DEKR/blob/7a303139e92bdf3eab8d899415ccac37374285a4/lib/dataset/target_generators/target_generators.py#L14C7-L14C23)