Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks). This is the same as the DropConnect impl I created for

for single image retargeting

retargeting the lip-open ratio of each source frame

Input: - memory: bs, \sum{hw}, d_model - memory_padding_mask: bs, \sum{hw} - spatial_shapes: nlevel, 2 - learnedw

Generate random input tensors and move them to GPU

Equivalent to nn.functional.interpolate, but with support for empty batch sizes. This will eventually be supported natively by PyTorch, and t

_summary_ Args: keypoints (torch.Tensor): ..., 51

Load and compile models for inference

Main function to benchmark speed and model parameters

Measure inference times for each model

Oks loss. Computing the oks loss between a set of predicted poses and target poses. The loss is calculated as negative log of oks. Args:

parsing the 2 points according to the 106 points, which cancels the roll

parsing the 2 points according to the 203 points, which cancels the roll

parsing the 2 points according to the 5 points, which cancels the roll

parsing the 2 points according to the 68 points, which cancels the roll

parsing the 2 points according to the 9 points, which cancels the roll ['right eye right', 'right eye left', 'left eye right', 'left eye left

for single image retargeting

for video retargeting

for keeping lips in the source video silent

Print benchmark results with average and standard deviation of inference times

This function disables printing when not in master process

kp_source: BxNx3 kp_driving: BxNx3 Return: Bx(3*num_kp+2)