data: dict {title: {motion:xxx, foot_contact:xxx}}

returns a list of [pad, conv] => should be += to some list, then apply sequential

create path by first checking its existence, :param paths: path :return:

rotations: [T, J, 4] root_positions: [T, 3]

data: {"train": dict_train, "test": dict_test} dict_train: {"style2d_code": blabla, etc.} separate_compute: compute t-SNE for 2D & 3D sep

Find parent rotations

foot_contact: [T, 4] -> take joints 0, 2 as standards phase = left foot in contact (0) --> right foot in contact (pi) --> left foot in contac

Forward function used in test time. This function wraps <forward> function in no_grad() so we don't save intermediate steps for backprop

Blank Transforms Parameters ---------- anim : Animation Input animation Returns -------

raw: - nrot: N * [J * 4 + 4, T] - rfree: N * [J * 3 + 4, T] - proj: N * [V, J * 2, T]

positions: positions [T, J, 3], trimmed (only "chosen_joints") hips, sdrs: left/right hip joints, left/right shoulder joints output: loca

The targeted armature could be a reduced one, e.g. no fingers. So move the skinning weight of each reduced armature to its nearest ancestor.

Calculate GAN loss for the discriminator Parameters: netD (network) -- the discriminator D real (tensor array) --