SMPLX missing one attribute: num_betas

[sylcito]

Hi, nice work!
But I meet one trouble when I try peoplesnapshot male-3.

HAHA/runners/train_gaussians.py

Line 178 in 46166fd

self._smplx_model = smplx.create(gender=gender, **model_params)

In this line, the variable num_betas=10 is not explicitly specified, so when the SMPLX is enabled, it automatically inherits the default num_betas=16 from SMPLH. As a result, when calculating LBS (Linear Blend Skinning), in the expression v_shaped = v_template + blend_shapes(betas, shapedirs), the execution of blend_shape = torch.einsum('bl,mkl->bmk', [betas, shape_disps]) triggers a dimension mismatch issue. This is because the dimension of shape_dirs should theoretically be [10475, 3, 20], but in this case, it becomes [10475, 3, 26].

After explicitly specifying num_betas=10, the bug is resolved.

[david-svitov]

Hi! Thanks for the note!