about the training strategy

[lvZic]

1. In the training, I found the camera params (s, tx, ty) seems not converge, but the mano params could converge normaly. I checked the dataset (freihand and interhand), the camera's focal length differs with each image, and it range from 400mm to 800mm. I wonder if the diversity of focal length in the dataset makes it converge worse?

2. Also, the solution may be adjusting the loss weight in the training process? just as u do /src/strategies/opt_default.py. I wonder how the converge order, what i mean is which parmas should converge first or last?

3. The camera params GT and camera loss isnot found in your project, should I remove the camera loss too?