Gaige Wang

When call pngquant.quant_image() in multiprocessing mode, the resulting images are disordered , say, two png images may be overwritten by each other.

some questions

3

why is it correct to accumulate losses of several frames and compute gradient just based on last frame?

https://github.com/wuhaozhe/pytorch-nicp/blob/099f189b749154e97d0f6fa5a1e16e8fb885cce0/bfm_model.py#L23C60-L23C60 > lm_index = bfm_meta_data['keypoints'] > faces = torch.from_numpy(faces).long().to(device) - 1 > lm_index = torch.from_numpy(lm_index).long().to(device) According to this line from 3dmm-fitting-pytorch repo, self.kp_inds = torch.tensor(model_dict['keypoints'] - 1).squeeze().long()](https://github.com/ascust/3DMM-Fitting-Pytorch/blob/master/core/BFM09Model.py#L26) , whether lm_index...