SadTalker / src /facerender /modules /make_animation.py
vinthony's picture
init
a22eb82
raw
history blame
6.3 kB
from scipy.spatial import ConvexHull
import torch
import torch.nn.functional as F
import numpy as np
from tqdm import tqdm
def normalize_kp(kp_source, kp_driving, kp_driving_initial, adapt_movement_scale=False,
use_relative_movement=False, use_relative_jacobian=False):
if adapt_movement_scale:
source_area = ConvexHull(kp_source['value'][0].data.cpu().numpy()).volume
driving_area = ConvexHull(kp_driving_initial['value'][0].data.cpu().numpy()).volume
adapt_movement_scale = np.sqrt(source_area) / np.sqrt(driving_area)
else:
adapt_movement_scale = 1
kp_new = {k: v for k, v in kp_driving.items()}
if use_relative_movement:
kp_value_diff = (kp_driving['value'] - kp_driving_initial['value'])
kp_value_diff *= adapt_movement_scale
kp_new['value'] = kp_value_diff + kp_source['value']
if use_relative_jacobian:
jacobian_diff = torch.matmul(kp_driving['jacobian'], torch.inverse(kp_driving_initial['jacobian']))
kp_new['jacobian'] = torch.matmul(jacobian_diff, kp_source['jacobian'])
return kp_new
def headpose_pred_to_degree(pred):
device = pred.device
idx_tensor = [idx for idx in range(66)]
idx_tensor = torch.FloatTensor(idx_tensor).to(device)
pred = F.softmax(pred)
degree = torch.sum(pred*idx_tensor, 1) * 3 - 99
return degree
def get_rotation_matrix(yaw, pitch, roll):
yaw = yaw / 180 * 3.14
pitch = pitch / 180 * 3.14
roll = roll / 180 * 3.14
roll = roll.unsqueeze(1)
pitch = pitch.unsqueeze(1)
yaw = yaw.unsqueeze(1)
pitch_mat = torch.cat([torch.ones_like(pitch), torch.zeros_like(pitch), torch.zeros_like(pitch),
torch.zeros_like(pitch), torch.cos(pitch), -torch.sin(pitch),
torch.zeros_like(pitch), torch.sin(pitch), torch.cos(pitch)], dim=1)
pitch_mat = pitch_mat.view(pitch_mat.shape[0], 3, 3)
yaw_mat = torch.cat([torch.cos(yaw), torch.zeros_like(yaw), torch.sin(yaw),
torch.zeros_like(yaw), torch.ones_like(yaw), torch.zeros_like(yaw),
-torch.sin(yaw), torch.zeros_like(yaw), torch.cos(yaw)], dim=1)
yaw_mat = yaw_mat.view(yaw_mat.shape[0], 3, 3)
roll_mat = torch.cat([torch.cos(roll), -torch.sin(roll), torch.zeros_like(roll),
torch.sin(roll), torch.cos(roll), torch.zeros_like(roll),
torch.zeros_like(roll), torch.zeros_like(roll), torch.ones_like(roll)], dim=1)
roll_mat = roll_mat.view(roll_mat.shape[0], 3, 3)
rot_mat = torch.einsum('bij,bjk,bkm->bim', pitch_mat, yaw_mat, roll_mat)
return rot_mat
def keypoint_transformation(kp_canonical, he):
kp = kp_canonical['value'] # (bs, k, 3)
yaw, pitch, roll= he['yaw'], he['pitch'], he['roll']
yaw = headpose_pred_to_degree(yaw)
pitch = headpose_pred_to_degree(pitch)
roll = headpose_pred_to_degree(roll)
if 'yaw_c' in he:
yaw = yaw + he['yaw_c']
if 'pitch_c' in he:
pitch = pitch + he['pitch_c']
if 'roll_c' in he:
roll = roll + he['roll_c']
rot_mat = get_rotation_matrix(yaw, pitch, roll) # (bs, 3, 3)
t, exp = he['t'], he['exp']
# keypoint rotation
kp_rotated = torch.einsum('bmp,bkp->bkm', rot_mat, kp)
# keypoint translation
t = t.unsqueeze_(1).repeat(1, kp.shape[1], 1)
kp_t = kp_rotated + t
# add expression deviation
exp = exp.view(exp.shape[0], -1, 3)
kp_transformed = kp_t + exp
return {'value': kp_transformed}
def make_animation(source_image, source_semantics, target_semantics,
generator, kp_detector, mapping,
yaw_c_seq=None, pitch_c_seq=None, roll_c_seq=None,
use_exp=True):
with torch.no_grad():
predictions = []
kp_canonical = kp_detector(source_image)
he_source = mapping(source_semantics)
kp_source = keypoint_transformation(kp_canonical, he_source)
for frame_idx in tqdm(range(target_semantics.shape[1]), 'Face Renderer:'):
target_semantics_frame = target_semantics[:, frame_idx]
he_driving = mapping(target_semantics_frame)
if not use_exp:
he_driving['exp'] = he_driving['exp']*0
if yaw_c_seq is not None:
he_driving['yaw_c'] = yaw_c_seq[:, frame_idx]
if pitch_c_seq is not None:
he_driving['pitch_c'] = pitch_c_seq[:, frame_idx]
if roll_c_seq is not None:
he_driving['roll_c'] = roll_c_seq[:, frame_idx]
kp_driving = keypoint_transformation(kp_canonical, he_driving)
#kp_norm = normalize_kp(kp_source=kp_source, kp_driving=kp_driving,
#kp_driving_initial=kp_driving_initial)
kp_norm = kp_driving
out = generator(source_image, kp_source=kp_source, kp_driving=kp_norm)
predictions.append(out['prediction'])
predictions_ts = torch.stack(predictions, dim=1)
return predictions_ts
class AnimateModel(torch.nn.Module):
"""
Merge all generator related updates into single model for better multi-gpu usage
"""
def __init__(self, generator, kp_extractor, mapping):
super(AnimateModel, self).__init__()
self.kp_extractor = kp_extractor
self.generator = generator
self.mapping = mapping
self.kp_extractor.eval()
self.generator.eval()
self.mapping.eval()
def forward(self, x):
source_image = x['source_image']
source_semantics = x['source_semantics']
target_semantics = x['target_semantics']
yaw_c_seq = x['yaw_c_seq']
pitch_c_seq = x['pitch_c_seq']
roll_c_seq = x['roll_c_seq']
predictions_video = make_animation(source_image, source_semantics, target_semantics,
self.generator, self.kp_extractor,
self.mapping, use_exp = True,
yaw_c_seq=yaw_c_seq, pitch_c_seq=pitch_c_seq, roll_c_seq=roll_c_seq)
return predictions_video