Issues training with idname Obama

[soni-H]

Hi,

I'm trying to re-train the model for id Obama to understand the pipeline better. However, in the second iteration I get the following error -

verts_final = uv_vertices + uv_vertices_deforms
RuntimeError: The size of tensor a (12280) must match the size of tensor b (14876) at non-singleton dimension 1

While debugging it, I found the self.uvmask_flaten_idx is changing in the second iteration which in turn is leading to a uv_vertices of arbitrarily some different dimension. Can you please direct me what exactly am I missing here? I'm using the inputs (alpha, face-parsing outputs, tracking outputs) as provided by you.

Thanks,
Harshita

[soni-H]

Can one of the authors please comment on the issue? @xiangjun-xj
On further debugging, it looks like the DeformModel.uvmask_flaten_idx is changing right after render() is invoked. Can you at least tell if my understanding about DeformModel.uvmask_flaten_idx being constant throughout the entire training is correct or not?

[xiangjun-xj]

Sorry, as almost no one else has the similar issue, I think it is more likely a personal problem. What about repeating the process from the very beginning?