def find_best_frame(source, driving, cpu=False):
import face_alignment
def normalize_kp(kp):
kp = kp - kp.mean(axis=0, keepdims=True)
area = ConvexHull(kp[:, :2]).volume
area = np.sqrt(area)
kp[:, :2] = kp[:, :2] / area
return kp
fa = face_alignment.FaceAlignment(face_alignment.LandmarksType._2D,
flip_input=True,
device='cpu' if cpu else 'cuda')
kp_source = fa.get_landmarks(255 * source)[0]
kp_source = normalize_kp(kp_source)
norm = float('inf')
frame_num = 0
for i, image in tqdm(enumerate(driving)):
kp_driving = fa.get_landmarks(255 * image)[0]
kp_driving = normalize_kp(kp_driving)
new_norm = (np.abs(kp_source - kp_driving)**2).sum()
if new_norm < norm:
norm = new_norm
frame_num = i
return frame_num
def single_pass(source, target_image, generator, kp_detector, kp_source, kp_driving_initial, relative=True, adapt_movement_scale=True, cpu=False):
with torch.no_grad():
#predictions = []
#source = torch.tensor(source_image[np.newaxis].astype(np.float32)).permute(0, 3, 1, 2)
target_image = torch.tensor(target_image[np.newaxis].astype(np.float32)).permute(0, 3, 1, 2)
#initial_image = torch.tensor(initial_image[np.newaxis].astype(np.float32)).permute(0, 3, 1, 2)
#if not cpu:
#source = source.cuda()
#kp_source = kp_detector(source)
#kp_driving_initial = kp_detector(initial_image)
#print(source.shape)
#input('source shape')
#print(target_image.shape)
#input('target_image_shape')
driving_frame = target_image
if not cpu:
driving_frame = driving_frame.cuda()
kp_driving = kp_detector(driving_frame)
kp_norm = normalize_kp(kp_source=kp_source, kp_driving=kp_driving,
kp_driving_initial=kp_driving_initial, use_relative_movement=relative,
use_relative_jacobian=relative, adapt_movement_scale=adapt_movement_scale)
out = generator(source, kp_source=kp_source, kp_driving=kp_norm)
#prediction = np.transpose(out['prediction'].data.cpu().numpy(), [0, 2, 3, 1])[0]
#return prediction
return out['prediction'].data
def make_animation(source_image,
driving_video,
generator,
kp_detector,
relative=True,
adapt_movement_scale=True):
with torch.no_grad():
predictions = []
source = torch.tensor(source_image[np.newaxis].astype(
np.float32)).permute(0, 3, 1, 2).cuda()
driving = torch.tensor(
np.array(driving_video)[np.newaxis].astype(np.float32)).permute(
0, 4, 1, 2, 3).cuda()
kp_source = kp_detector(source)
kp_driving_initial = kp_detector(driving[:, :, 0])
for frame_idx in tqdm(range(driving.shape[2])):
driving_frame = driving[:, :, frame_idx]
kp_driving = kp_detector(driving_frame)
kp_norm = normalize_kp(kp_source=kp_source,
kp_driving=kp_driving,
kp_driving_initial=kp_driving_initial,
use_relative_movement=relative,
use_relative_jacobian=relative,
adapt_movement_scale=adapt_movement_scale)
out = generator(source, kp_source=kp_source, kp_driving=kp_norm)
predictions.append(
np.transpose(out['prediction'].data.cpu().numpy(),
[0, 2, 3, 1])[0])
return predictions
def predict(driving_frame, source_image, relative, adapt_movement_scale, fa, device='cuda'):
global start_frame
global start_frame_kp
global kp_driving_initial
global kp_source
with torch.no_grad():
source = torch.tensor(source_image[np.newaxis].astype(np.float32)).permute(0, 3, 1, 2).to(device)
driving = torch.tensor(driving_frame[np.newaxis].astype(np.float32)).permute(0, 3, 1, 2).to(device)
if kp_driving_initial is None:
kp_driving_initial = kp_detector(driving)
start_frame = driving_frame.copy()
start_frame_kp = get_frame_kp(fa, driving_frame)
if kp_source is None:
kp_source = kp_detector(source)
kp_driving = kp_detector(driving)
kp_norm = normalize_kp(kp_source=kp_source, kp_driving=kp_driving,
kp_driving_initial=kp_driving_initial, use_relative_movement=relative,
use_relative_jacobian=relative, adapt_movement_scale=adapt_movement_scale)
if opt.enc_downscale > 1:
h, w = int(source.shape[2] / opt.enc_downscale), int(source.shape[3] / opt.enc_downscale)
source_enc = torch.nn.functional.interpolate(source, size=(h, w), mode='bilinear')
else:
source_enc = None
out = generator(source, kp_source=kp_source, kp_driving=kp_norm, source_image_enc=source_enc, optim_ret=True)
out = np.transpose(out['prediction'].data.cpu().numpy(), [0, 2, 3, 1])[0]
out = (np.clip(out, 0, 1) * 255).astype(np.uint8)
return out
def predict(driving_frame, source_image, relative, adapt_movement_scale, fa,
generator, kp_detector, kp_driving_initial, device='cuda'):
global start_frame
global start_frame_kp
# global kp_driving_initial
with torch.no_grad():
source = torch.tensor(source_image[np.newaxis].astype(np.float32)).permute(0, 3, 1, 2).to(device)
driving = torch.tensor(driving_frame[np.newaxis].astype(np.float32)).permute(0, 3, 1, 2).to(device)
kp_source = kp_detector(source)
if kp_driving_initial is None:
kp_driving_initial = kp_detector(driving)
start_frame = driving_frame.copy()
start_frame_kp = get_frame_kp(fa, driving_frame)
kp_driving = kp_detector(driving)
kp_norm = normalize_kp(kp_source=kp_source, kp_driving=kp_driving,
kp_driving_initial=kp_driving_initial, use_relative_movement=relative,
use_relative_jacobian=relative, adapt_movement_scale=adapt_movement_scale)
out = generator(source, kp_source=kp_source, kp_driving=kp_norm)
out = np.transpose(out['prediction'].data.cpu().numpy(), [0, 2, 3, 1])[0]
out = (np.clip(out, 0, 1) * 255).astype(np.uint8)
return out
def kp_animation(source_image,
kp,
generator,
relative=True,
adapt_movement_scale=True,
cpu=False):
with torch.no_grad():
predictions = []
source = torch.tensor(source_image[np.newaxis].astype(
np.float32)).permute(0, 3, 1, 2)
if not cpu:
source = source.cuda()
kp_source = kp_driving_initial = kp[0]
for kp_driving in kp:
kp_norm = normalize_kp(kp_source=kp_source,
kp_driving=kp_driving,
kp_driving_initial=kp_driving_initial,
use_relative_movement=relative,
use_relative_jacobian=relative,
adapt_movement_scale=adapt_movement_scale)
out = generator(source, kp_source=kp_source, kp_driving=kp_norm)
predictions.append(
np.transpose(out['prediction'].data.cpu().numpy(),
[0, 2, 3, 1])[0])
return predictions
def generate_morphed_video(image, sub_video):
with torch.no_grad():
source = (
torch.tensor(np.float32(image), device="cuda").permute(2, 0, 1).unsqueeze(0)
)
kp_source = kp_detector(source)
for i, driving_frame in enumerate(sub_video):
driving_frame = (
torch.tensor(np.float32(driving_frame))
.cuda()
.permute(2, 0, 1)
.unsqueeze(0)
)
if i == 0:
kp_driving_initial = kp_detector(driving_frame)
kp_driving = kp_detector(driving_frame)
kp_norm = normalize_kp(
kp_source=kp_source,
kp_driving=kp_driving,
kp_driving_initial=kp_driving_initial,
use_relative_movement=True,
use_relative_jacobian=True,
adapt_movement_scale=True,
)
yield generator(source, kp_source=kp_source, kp_driving=kp_norm)[
"prediction"
].squeeze().permute(1, 2, 0).cpu().numpy()
def make_animation(source_image_true,source_image_fake, driving_video, generator, kp_detector, relative=True, adapt_movement_scale=True, cpu=False):
with torch.no_grad():
sparse_d = []
occlusion = []
source_true = torch.tensor(source_image_true[np.newaxis].astype(np.float32)).permute(0, 3, 1, 2)
source_fake = torch.tensor(source_image_fake[np.newaxis].astype(np.float32)).permute(0, 3, 1, 2)
if not cpu:
source_true = source_true.cuda()
driving = torch.tensor(np.array(driving_video)[np.newaxis].astype(np.float32)).permute(0, 4, 1, 2, 3)
kp_source = kp_detector(source_true)
kp_driving_initial = kp_detector(driving[:, :, 0])
for frame_idx in tqdm(range(driving.shape[2])):
driving_frame = driving[:, :, frame_idx]
if not cpu:
driving_frame = driving_frame.cuda()
kp_driving = kp_detector(driving_frame)
kp_norm = normalize_kp(kp_source=kp_source, kp_driving=kp_driving,
kp_driving_initial=kp_driving_initial, use_relative_movement=relative,
use_relative_jacobian=relative, adapt_movement_scale=adapt_movement_scale)
out = generator(source_true, source_fake,kp_source=kp_source, kp_driving=kp_norm)
print(out['sparse_deformed'].data.cpu().numpy().size())
sparse_d.append(np.transpose(out['sparse_deformed'].data.cpu().numpy(), [0, 1, 3, 4, 2])[0][0])
occlusion.append(np.transpose(out['occlusion_map'].data.cpu().numpy(), [0, 2, 3, 1])[0])
return sparse_d,occlusion
def predict(self, driving_frame):
assert self.kp_source is not None, "call set_source_image()"
with torch.no_grad():
driving = to_tensor(driving_frame).to(self.device)
if self.kp_driving_initial is None:
self.kp_driving_initial = self.kp_detector(driving)
self.start_frame = driving_frame.copy()
self.start_frame_kp = self.get_frame_kp(driving_frame)
kp_driving = self.kp_detector(driving)
kp_norm = normalize_kp(
kp_source=self.kp_source,
kp_driving=kp_driving,
kp_driving_initial=self.kp_driving_initial,
use_relative_movement=self.relative,
use_relative_jacobian=self.relative,
adapt_movement_scale=self.adapt_movement_scale)
out = self.generator(self.source,
kp_source=self.kp_source,
kp_driving=kp_norm)
out = np.transpose(out['prediction'].data.cpu().numpy(),
[0, 2, 3, 1])[0]
out = (np.clip(out, 0, 1) * 255).astype(np.uint8)
return out
def next(self, image):
_, kp_image = self._tensor_image(image)
result = []
for tgt, kp in zip(self.targets, self.kp_targets):
with torch.no_grad():
norm = normalize_kp(kp, kp_image, self.kp_start, self.adapt_movement_scale, self.relative, self.relative)
out = self.generator(tgt, kp_source=kp, kp_driving=norm)
result.append(np.transpose(out['prediction'].data.cpu().numpy(), [0, 2, 3, 1])[0])
return result
def make_animation(source_image,
driving_video,
generator,
kp_detector,
relative=True,
adapt_movement_scale=True,
cpu=False):
with torch.no_grad():
predictions = []
source = torch.tensor(source_image[np.newaxis].astype(
np.float32)).permute(0, 3, 1, 2)
if not cpu:
source = source.cuda()
driving = torch.tensor(
np.array(driving_video)[np.newaxis].astype(np.float32)).permute(
0, 4, 1, 2, 3)
kp_source = kp_detector(source)
kp_driving_initial = kp_detector(driving[:, :, 0])
# pprint('source', source) # [1, 3, 256, 256]
# pprint('driving', driving) # [1, 3, 30, 256, 256]
# pprint('kp_source_value', kp_source['value'])
# pprint('kp_source_jacobian', kp_source['jacobian'])
# pprint('kp_driving_initial', kp_driving_initial)
flag = True
for frame_idx in tqdm(range(driving.shape[2])):
driving_frame = driving[:, :, frame_idx]
if not cpu:
driving_frame = driving_frame.cuda()
kp_driving = kp_detector(driving_frame)
kp_norm = normalize_kp(kp_source=kp_source,
kp_driving=kp_driving,
kp_driving_initial=kp_driving_initial,
use_relative_movement=relative,
use_relative_jacobian=relative,
adapt_movement_scale=adapt_movement_scale)
out = generator(source, kp_source=kp_source, kp_driving=kp_norm)
# if flag:
# flag = False
# pprint('kp_driving', kp_driving)
# pprint('kp_norm', kp_norm)
# pprint('out', out)
predictions.append(
np.transpose(out['prediction'].data.cpu().numpy(),
[0, 2, 3, 1])[0])
return predictions
def make_animation(source_image,
driving_video,
generator,
kp_detector,
relative=True,
adapt_movement_scale=True,
cpu=False,
progress_var=None,
progress_label=None):
with torch.no_grad():
predictions = []
source = torch.tensor(source_image[np.newaxis].astype(
np.float32)).permute(0, 3, 1, 2)
if not cpu:
source = source.cuda()
driving = torch.tensor(
np.array(driving_video)[np.newaxis].astype(np.float32)).permute(
0, 4, 1, 2, 3)
kp_source = kp_detector(source)
kp_driving_initial = kp_detector(driving[:, :, 0])
#tqdm(range(driving.shape[2]))
#tqdm_gui(range(driving.shape[2]))
for frame_idx in range(driving.shape[2]):
percent = frame_idx / driving.shape[2] * 100
if progress_var:
progress_var.set(int(percent))
if progress_label:
progress_label.config(text=f'{percent:.3}%')
driving_frame = driving[:, :, frame_idx]
if not cpu:
driving_frame = driving_frame.cuda()
kp_driving = kp_detector(driving_frame)
kp_norm = normalize_kp(kp_source=kp_source,
kp_driving=kp_driving,
kp_driving_initial=kp_driving_initial,
use_relative_movement=relative,
use_relative_jacobian=relative,
adapt_movement_scale=adapt_movement_scale)
out = generator(source, kp_source=kp_source, kp_driving=kp_norm)
predictions.append(
np.transpose(out['prediction'].data.cpu().numpy(),
[0, 2, 3, 1])[0])
progress_label.config(text="Done!")
return predictions
def process(opt,img_orig, generator, kp_detector):
img = resize(img_orig, (256, 256))[..., :3]
kp_driving_initial = None
with torch.no_grad():
spm = torch.tensor(img[np.newaxis].astype(np.float32)).permute(0, 3, 1, 2)
source = spm.cpu() if opt.cpu else spm.cuda()
kp_source = kp_detector(source)
video = cv2.VideoCapture(opt.video)
fps = video.get(cv2.CAP_PROP_FPS)
fourcc = cv2.VideoWriter_fourcc(*'avc1')
vout = cv2.VideoWriter(opt.out, fourcc, fps, (256, 256))
while True:
frame_img = video.read()
if isinstance(frame_img, tuple):
frame_img = frame_img[1]
if frame_img is None:
print("Oops frame is None. Possibly camera or display does not work")
break
#frame_img = cv2.rotate(frame_img,cv2.ROTATE_180)
frame_img = cv2.cvtColor(frame_img, cv2.COLOR_BGR2RGB)
y, x, _ = frame_img.shape
min_dim = min(y, x)
startx = x // 2 - (min_dim // 2)
starty = y // 2 - (min_dim // 2)
frame_img = frame_img[starty:starty + min_dim, startx:startx + min_dim, :]
frame_img = resize(frame_img, (256, 256))[..., :3]
frame = torch.tensor(frame_img[np.newaxis].astype(np.float32)).permute(0, 3, 1, 2)
kp_driving = kp_detector(frame)
if kp_driving_initial is None:
kp_driving_initial = kp_driving
kp_norm = normalize_kp(kp_source=kp_source, kp_driving=kp_driving,
kp_driving_initial=kp_driving_initial, use_relative_movement=opt.relative,
use_relative_jacobian=opt.relative, adapt_movement_scale=opt.adapt_scale)
out = generator(source, kp_source=kp_source, kp_driving=kp_norm)
p = np.transpose(out['prediction'].data.cpu().numpy(), [0, 2, 3, 1])[0]
p = p*255
p = p.astype(np.uint8)
p = cv2.cvtColor(p, cv2.COLOR_BGR2RGB)
vout.write(p)
video.release()
vout.release()
def make_animation_video(source_video,
driving_video,
generator,
kp_detector,
relative=True,
adapt_movement_scale=True,
cpu=False):
with torch.no_grad():
predictions = []
#source = torch.tensor(source_image[np.newaxis].astype(np.float32)).permute(0, 3, 1, 2)
#if not cpu:
# source = source.cuda()
driving = torch.tensor(
np.array(driving_video)[np.newaxis].astype(np.float32)).permute(
0, 4, 1, 2, 3)
source_v = torch.tensor(
np.array(source_video)[np.newaxis].astype(np.float32)).permute(
0, 4, 1, 2, 3)
#kp_source = kp_detector(source)
kp_driving_initial = kp_detector(driving[:, :, 0])
# first try use source video as main counter of frames
# source_video must have more frames than drivng video
for frame_idx in tqdm(range(source_v.shape[2])):
driving_frame = driving[:, :, frame_idx]
source_frame = source_v[:, :, frame_idx]
if not cpu:
driving_frame = driving_frame.cuda()
source_frame = source_frame.cuda()
kp_driving = kp_detector(driving_frame)
kp_source = kp_detector(source_frame)
kp_norm = normalize_kp(kp_source=kp_source,
kp_driving=kp_driving,
kp_driving_initial=kp_driving_initial,
use_relative_movement=relative,
use_relative_jacobian=relative,
adapt_movement_scale=adapt_movement_scale)
out = generator(source_frame,
kp_source=kp_source,
kp_driving=kp_norm)
predictions.append(
np.transpose(out['prediction'].data.cpu().numpy(),
[0, 2, 3, 1])[0])
return predictions
def predict(self, driving_frame):
with torch.no_grad():
driving = to_tensor(driving_frame).to(self.device)
if self.kp_driving_initial is None:
self.kp_driving_initial = self.kp_detector(driving)
self.start_frame = driving_frame.copy()
self.start_frame_kp = self.get_frame_kp(driving_frame)
kp_driving = self.kp_detector(driving)
kp_norm = normalize_kp(
kp_source=self.kp_source,
kp_driving=kp_driving,
kp_driving_initial=self.kp_driving_initial,
use_relative_movement=self.relative,
use_relative_jacobian=self.relative,
adapt_movement_scale=self.adapt_movement_scale)
if self.enc_downscale > 1:
h, w = int(source.shape[2] / self.enc_downscale), int(
source.shape[3] / self.enc_downscale)
source_enc = torch.nn.functional.interpolate(source,
size=(h, w),
mode='bilinear')
else:
source_enc = None
try:
out = self.generator(self.source,
kp_source=self.kp_source,
kp_driving=kp_norm,
source_image_enc=source_enc,
optim_ret=True)
except TypeError:
Once('\n*** Please update FOMM:\ncd fomm\ngit pull\n')
out = self.generator(self.source,
kp_source=self.kp_source,
kp_driving=kp_norm)
out = np.transpose(out['prediction'].data.cpu().numpy(),
[0, 2, 3, 1])[0]
out = (np.clip(out, 0, 1) * 255).astype(np.uint8)
return out
def process(self, vframe):
with torch.no_grad():
y, x, _ = vframe.shape
min_dim = min(y, x)
startx = x // 2 - (min_dim // 2)
starty = y // 2 - (min_dim // 2)
vframe = vframe[starty:starty + min_dim,
startx:startx + min_dim, :]
vframe = resize(vframe, (256, 256))[..., :3]
if self.count < 60:
self.count += 1
p = np.concatenate([vframe[:, :, ::-1], self.img], axis=1)
p = p * 255
p = p.astype(np.uint8)
p = cv2.cvtColor(p, cv2.COLOR_BGR2RGB)
return p
frame = torch.tensor(vframe[np.newaxis].astype(
np.float32)).permute(0, 3, 1, 2)
kp_driving = kp_detector(frame)
if self.kp_driving_initial is None:
self.kp_driving_initial = kp_driving
kp_norm = normalize_kp(kp_source=self.kp_source,
kp_driving=kp_driving,
kp_driving_initial=self.kp_driving_initial,
use_relative_movement=True,
use_relative_jacobian=True,
adapt_movement_scale=True)
out = generator(self.source,
kp_source=self.kp_source,
kp_driving=kp_norm)
p = np.transpose(out['prediction'].data.cpu().numpy(),
[0, 2, 3, 1])[0]
p = np.concatenate([vframe[:, :, ::-1], p], axis=1)
p = p * 255
p = p.astype(np.uint8)
p = cv2.cvtColor(p, cv2.COLOR_BGR2RGB)
self.count += 1
return p
def do_feeding(self, task_meta: TaskMeta, generator):
source_img = cv2.imread(task_meta.source_img_path)
source_img = resize_image(source_img)
source_faces = self.get_source_faces(source_img)
if len(source_faces) == 0:
return process_code.NO_AVAILABLE_FACE
self.update_kp_sources(source_faces)
resource, frame_data_list = self.resource_manager.get_cache_data(
task_meta.material_file_path, task_meta.ppd_file_path)
predictions = []
for source_face in source_faces:
best_frame_id = self.get_best_frame(source_face, frame_data_list)
best_frame = resource.material[best_frame_id]
best_frame = best_frame[:, :, ::-1].copy()
frame = torch.tensor(best_frame[np.newaxis].astype(np.float32)) \
.permute(0, 3, 1, 2).cuda() / 255.
kp_driving_initial = self.kp_detector(frame)
relative = True
adapt_movement_scale = True
for frame_data in frame_data_list:
kp_norm = normalize_kp(
kp_source=source_face.kp_source,
kp_driving=frame_data.kp_by_detector,
kp_driving_initial=kp_driving_initial,
use_relative_movement=relative,
use_relative_jacobian=relative,
adapt_movement_scale=adapt_movement_scale)
out = self.generator(source_face.face_img,
kp_source=source_face.kp_source,
kp_driving=kp_norm)
predictions.append(
np.transpose(out['prediction'].data.cpu().numpy(),
[0, 2, 3, 1])[0])
return predictions
def make_animation(source_image,
driving_video,
relative=True,
adapt_movement_scale=True,
cpu=False):
generator, kp_detector = load_checkpoints.load_checkpoints(
config_path='config/vox-256.yaml',
checkpoint_path='data/vox-cpk.pth.tar',
cpu=True)
with torch.no_grad():
predictions = []
source = torch.tensor(source_image[np.newaxis].astype(
np.float32)).permute(0, 3, 1, 2)
if not cpu:
source = source.cuda()
driving = torch.tensor(
np.array(driving_video)[np.newaxis].astype(np.float32)).permute(
0, 4, 1, 2, 3)
kp_source = kp_detector(source)
kp_driving_initial = kp_detector(driving[:, :, 0])
for frame_idx in range(driving.shape[2]):
progress.progress(frame_idx / float(driving.shape[2]))
driving_frame = driving[:, :, frame_idx]
if not cpu:
driving_frame = driving_frame.cuda()
kp_driving = kp_detector(driving_frame)
kp_norm = normalize_kp(kp_source=kp_source,
kp_driving=kp_driving,
kp_driving_initial=kp_driving_initial,
use_relative_movement=relative,
use_relative_jacobian=relative,
adapt_movement_scale=adapt_movement_scale)
out = generator(source, kp_source=kp_source, kp_driving=kp_norm)
predictions.append(
np.transpose(out['prediction'].data.cpu().numpy(),
[0, 2, 3, 1])[0])
return predictions
def animate_image(source_image,
video_image,
orig_frame,
generator,
kp_detector,
relative=True,
adapt_movement_scale=True,
cpu=False):
with torch.no_grad():
initial_frame = torch.tensor(orig_frame[np.newaxis].astype(
np.float32)).permute(0, 3, 1, 2)
driving_frame = torch.tensor(video_image[np.newaxis].astype(
np.float32)).permute(0, 3, 1, 2)
source = torch.tensor(source_image[np.newaxis].astype(
np.float32)).permute(0, 3, 1, 2)
if not cpu:
source = source.cuda()
driving_frame = driving_frame.cuda()
initial_frame = initial_frame.cuda()
#print(driving_frame.shape,source.shape)
#driving = torch.tensor(np.array(driving_video)[np.newaxis].astype(np.float32)).permute(0, 4, 1, 2, 3)
kp_source = kp_detector(source)
kp_driving_initial = kp_detector(initial_frame)
kp_driving = kp_detector(driving_frame)
kp_norm = normalize_kp(kp_source=kp_source,
kp_driving=kp_driving,
kp_driving_initial=kp_driving_initial,
use_relative_movement=relative,
use_relative_jacobian=relative,
adapt_movement_scale=adapt_movement_scale)
out = generator(source, kp_source=kp_source, kp_driving=kp_norm)
out = np.transpose(out['prediction'].data.cpu().numpy(),
[0, 2, 3, 1])[0]
#print(np.shape(out))
return out
def process(input):
print("[INFO] loading source image and checkpoint...")
source_path = input
checkpoint_path = args['checkpoint']
if args['input_video']:
video_path = args['input_video']
else:
video_path = None
source_image = imageio.imread(source_path)
source_image = resize(source_image, (256, 256))[..., :3]
generator, kp_detector = load_checkpoints(
config_path='config/vox-256.yaml', checkpoint_path=checkpoint_path)
# Load the cascade
face_cascade = cv2.CascadeClassifier('haarcascade_frontalface_default.xml')
if not os.path.exists('output'):
os.mkdir('output')
relative = True
adapt_movement_scale = True
if args['cpu']:
cpu = True
else:
cpu = False
if video_path:
cap = cv2.VideoCapture(video_path)
print("[INFO] Loading video from the given path")
else:
cap = cv2.VideoCapture(0)
print("[INFO] Initializing front camera...")
# get vcap property
width = cap.get(cv2.CAP_PROP_FRAME_WIDTH) # float `width`
height = cap.get(cv2.CAP_PROP_FRAME_HEIGHT) # float `height`
fps = cap.get(cv2.CAP_PROP_FPS)
frame_count = cap.get(cv2.CAP_PROP_FRAME_COUNT)
print('resolution : {} x {}'.format(width, height))
print('frame rate : {} \nframe count : {}'.format(fps, frame_count))
fourcc = cv2.VideoWriter_fourcc(*'MJPG')
out1 = cv2.VideoWriter('output/test.avi', fourcc, 12, (256 * 3, 256), True)
cv2_source = cv2.cvtColor(source_image.astype('float32'),
cv2.COLOR_BGR2RGB)
cv2_source2 = (source_image * 255).astype(np.uint8)
if args['vc']:
camera = pyfakewebcam.FakeWebcam('/dev/video7', 640, 360)
camera._settings.fmt.pix.width = 640
camera._settings.fmt.pix.height = 360
img = np.zeros((360, 640, 3), dtype=np.uint8)
yoff = round((360 - 256) / 2)
xoff = round((640 - 256) / 2)
img_im = img.copy()
img_cv2_source = img.copy()
img_im[:, :, 2] = 255
img_cv2_source[:, :, 2] = 255
with torch.no_grad():
predictions = []
source = torch.tensor(source_image[np.newaxis].astype(
np.float32)).permute(0, 3, 1, 2)
if not cpu:
source = source.cuda()
kp_source = kp_detector(source)
count = 0
fps = []
if args['csv']:
line1 = []
size = 10
x_vec = np.linspace(0, 1, size + 1)[0:-1]
y_vec = np.random.randn(len(x_vec))
while (True):
start = time.time()
ret, frame = cap.read()
gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
# Detect the faces
faces = face_cascade.detectMultiScale(gray, 1.1, 4)
frame = cv2.flip(frame, 1)
if ret == True:
if not video_path:
x = 143
y = 87
w = 322
h = 322
frame = frame[y:y + h, x:x + w]
frame1 = resize(frame, (256, 256))[..., :3]
if count == 0:
source_image1 = frame1
source1 = torch.tensor(source_image1[np.newaxis].astype(
np.float32)).permute(0, 3, 1, 2)
kp_driving_initial = kp_detector(source1)
frame_test = torch.tensor(frame1[np.newaxis].astype(
np.float32)).permute(0, 3, 1, 2)
driving_frame = frame_test
if not cpu:
driving_frame = driving_frame.cuda()
kp_driving = kp_detector(driving_frame)
kp_norm = normalize_kp(
kp_source=kp_source,
kp_driving=kp_driving,
kp_driving_initial=kp_driving_initial,
use_relative_movement=relative,
use_relative_jacobian=relative,
adapt_movement_scale=adapt_movement_scale)
out = generator(source,
kp_source=kp_source,
kp_driving=kp_norm)
predictions.append(
np.transpose(out['prediction'].data.cpu().numpy(),
[0, 2, 3, 1])[0])
im = np.transpose(out['prediction'].data.cpu().numpy(),
[0, 2, 3, 1])[0]
#im = cv2.cvtColor(im,cv2.COLOR_RGB2BGR)
#cv2_source = cv2.cvtColor(cv2_source,cv2.COLOR_RGB2BGR)
im = (np.array(im) * 255).astype(np.uint8)
#cv2_source = (np.array(cv2_source)*255).astype(np.uint8)
img_im[yoff:yoff + 256, xoff:xoff + 256] = im
img_cv2_source[yoff:yoff + 256, xoff:xoff + 256] = cv2_source2
#print(faces)
#print(type(im))
if args['debug']:
#print("[DEBUG] FPS : ",1.0 / (time.time()-start))
fps.append(1.0 / (time.time() - start))
if args['cpu']:
print("[DEBUG] Avg. of FPS using CPU : ", mean(fps))
else:
print("[DEBUG] Avg. of FPS using GPU : ", mean(fps))
if args['csv']:
y_vec[-1] = mean(fps)
line1 = live_plotter(x_vec, y_vec, line1)
y_vec = np.append(y_vec[1:], 0.0)
if args['vc']:
if np.array(faces).any():
#joinedFrame = np.concatenate((cv2_source,im,frame1),axis=1)
camera.schedule_frame(img_im)
else:
#joinedFrame = np.concatenate((cv2_source,cv2_source,frame1),axis=1)
camera.schedule_frame(img_cv2_source)
#cv2.imshow('Test',joinedFrame)
#out1.write(img_as_ubyte(np.array(im)))
count += 1
else:
break
cap.release()
out1.release()
cv2.destroyAllWindows()
if count == 0:
source_image1 = frame1
source1 = torch.tensor(source_image1[np.newaxis].astype(
np.float32)).permute(0, 3, 1, 2)
kp_driving_initial = kp_detector(source1)
frame_test = torch.tensor(frame1[np.newaxis].astype(
np.float32)).permute(0, 3, 1, 2)
driving_frame = frame_test
if not cpu:
driving_frame = driving_frame.cuda()
kp_driving = kp_detector(driving_frame)
kp_norm = normalize_kp(kp_source=kp_source,
kp_driving=kp_driving,
kp_driving_initial=kp_driving_initial,
use_relative_movement=relative,
use_relative_jacobian=relative,
adapt_movement_scale=adapt_movement_scale)
out = generator(source, kp_source=kp_source, kp_driving=kp_norm)
predictions.append(
np.transpose(out['prediction'].data.cpu().numpy(),
[0, 2, 3, 1])[0])
im = np.transpose(out['prediction'].data.cpu().numpy(),
[0, 2, 3, 1])[0]
im = cv2.cvtColor(im, cv2.COLOR_RGB2BGR)
joinedFrame = np.concatenate((cv2_source, im, frame1), axis=1)
cv2.imshow('Test', joinedFrame)
out1.write(img_as_ubyte(joinedFrame))
count += 1
if cv2.waitKey(20) & 0xFF == ord('q'):
def make_animation(source_images,
driving_video,
generator,
kp_detector,
relative=True,
adapt_movement_scale=True,
cpu=False):
with torch.no_grad():
predictions = []
source = [
torch.tensor(s[np.newaxis].astype(np.float32)).permute(0, 3, 1, 2)
for s in source_images
]
driving = torch.tensor(
np.array(driving_video)[np.newaxis].astype(np.float32)).permute(
0, 4, 1, 2, 3)
if not cpu:
source = [s.cuda() for s in source]
kp_source = [kp_detector(s) for s in source]
kp_source_value = [
kp_s['value'][0].detach().cpu().numpy() for kp_s in kp_source
]
kp_driving_initial = kp_detector(driving[:, :, 0])
distance = lambda y: lambda x: np.sum(np.sum((x - y)**2, axis=1)**0.5)
kp_frame_value = kp_driving_initial['value'][0].detach().cpu().numpy()
i_prev = np.argmin(list(map(distance(kp_frame_value),
kp_source_value)))
kp_source_prev, source_prev = kp_source[i_prev], source[i_prev]
diff = 20
alpha = 0
n = len(source_images)
for frame_idx in tqdm(range(driving.shape[2])):
driving_frame = driving[:, :, frame_idx]
if not cpu:
driving_frame = driving_frame.cuda()
kp_driving = kp_detector(driving_frame)
kp_frame_value = kp_driving['value'][0].detach().cpu().numpy()
i = np.argmin(
list(
map(
distance(kp_frame_value),
kp_source_value[max(0, i_prev -
diff):min(n, i_prev + diff)])))
i += max(0, i_prev - diff)
if i != i_prev:
kp_source_prev['value'] = (kp_source_prev['value'] +
kp_source[i]['value']) / 2
kp_source_prev['jacobian'] = (kp_source_prev['jacobian'] +
kp_source[i]['jacobian']) / 2
source_prev = (source_prev + source[i]) / 2
i_prev = i
else:
kp_source_prev['value'] = alpha * kp_source_prev['value'] + (
1 - alpha) * kp_source[i]['value']
kp_source_prev['jacobian'] = alpha * kp_source_prev[
'jacobian'] + (1 - alpha) * kp_source[i]['jacobian']
source_prev = alpha * source_prev + (1 - alpha) * source[i]
kp_norm = normalize_kp(kp_source=kp_source_prev,
kp_driving=kp_driving,
kp_driving_initial=kp_driving_initial,
use_relative_movement=relative,
use_relative_jacobian=relative,
adapt_movement_scale=adapt_movement_scale)
out = generator(source_prev,
kp_source=kp_source_prev,
kp_driving=kp_norm)
predictions.append(
np.transpose(out['prediction'].data.cpu().numpy(),
[0, 2, 3, 1])[0])
return predictions
def process_task(task_id, opt, img_orig, video_file, out_file, generator, kp_detector):
LOG.error(f"Processing task {task_id}...")
img = resize(img_orig, (256, 256))[..., :3]
kp_driving_initial = None
with torch.no_grad():
spm = torch.tensor(img[np.newaxis].astype(np.float32)).permute(0, 3, 1, 2)
source = spm.cpu() if opt.cpu else spm.cuda()
kp_source = kp_detector(source)
video = cv2.VideoCapture(video_file)
fps = video.get(cv2.CAP_PROP_FPS)
fourcc = cv2.VideoWriter_fourcc(*"avc1")
vout = cv2.VideoWriter(out_file, fourcc, fps, (256, 256))
frames_count = int(cv2.VideoCapture.get(video, cv2.CAP_PROP_FRAME_COUNT))
frame_counter = 0
last_percent = 0
last_time = time.time()
while True:
frame_img = video.read()
if isinstance(frame_img, tuple):
frame_img = frame_img[1]
if frame_img is None:
print("Oops frame is None. Possibly camera or display does not work")
break
# frame_img = cv2.rotate(frame_img,cv2.ROTATE_180)
frame_img = cv2.cvtColor(frame_img, cv2.COLOR_BGR2RGB)
y, x, _ = frame_img.shape
min_dim = min(y, x)
startx = x // 2 - (min_dim // 2)
starty = y // 2 - (min_dim // 2)
frame_img = frame_img[
starty : starty + min_dim, startx : startx + min_dim, :
]
frame_img = resize(frame_img, (256, 256))[..., :3]
frame = torch.tensor(frame_img[np.newaxis].astype(np.float32)).permute(
0, 3, 1, 2
)
kp_driving = kp_detector(frame)
if kp_driving_initial is None:
kp_driving_initial = kp_driving
kp_norm = normalize_kp(
kp_source=kp_source,
kp_driving=kp_driving,
kp_driving_initial=kp_driving_initial,
use_relative_movement=opt.relative,
use_relative_jacobian=opt.relative,
adapt_movement_scale=opt.adapt_scale,
)
out = generator(source, kp_source=kp_source, kp_driving=kp_norm)
p = np.transpose(out["prediction"].data.cpu().numpy(), [0, 2, 3, 1])[0]
p = p * 255
p = p.astype(np.uint8)
p = cv2.cvtColor(p, cv2.COLOR_BGR2RGB)
vout.write(p)
frame_counter += 1
percent = int(frame_counter / frames_count * 100)
now = time.time()
if percent != last_percent and now - last_time > 1:
last_percent = percent
last_time = now
send_status(opt, task_id, percent=min(100, percent))
LOG.info(f"processed {frame_counter}/{frames_count} frames")
if last_percent != 100:
send_status(opt, task_id, percent=100)
video.release()
vout.release()
LOG.info(f"Task {task_id} done, file {out_file} written!")
def overfit(config,
source_image,
driving_video,
generator,
kp_detector,
lowres_video,
relative=True,
adapt_movement_scale=True,
cpu=False):
overfit_epochs = 10
train_params = config['train_params']
optimizer_generator = torch.optim.Adam(generator.parameters(),
lr=train_params['lr_generator'],
betas=(0.5, 0.999))
optimizer_kp_detector = torch.optim.Adam(kp_detector.parameters(),
lr=train_params['lr_kp_detector'],
betas=(0.5, 0.999))
scheduler_generator = MultiStepLR(optimizer_generator,
train_params['epoch_milestones'],
gamma=0.1,
last_epoch=-1)
scheduler_kp_detector = MultiStepLR(optimizer_kp_detector,
train_params['epoch_milestones'],
gamma=0.1,
last_epoch=-1)
for epoch in trange(0, overfit_epochs):
predictions = []
source = torch.tensor(source_image[np.newaxis].astype(
np.float32)).permute(0, 3, 1, 2)
if not cpu:
source = source.cuda()
driving = torch.tensor(
np.array(driving_video)[np.newaxis].astype(np.float32)).permute(
0, 4, 1, 2, 3)
lowres = torch.tensor(
np.array(lowres_video)[np.newaxis].astype(np.float32)).permute(
0, 4, 1, 2, 3)
kp_source = kp_detector(source)
kp_driving_initial = kp_detector(driving[:, :, 0])
for frame_idx in tqdm(range(driving.shape[2])):
driving_frame = driving[:, :, frame_idx]
lowres_frame = lowres[:, :, frame_idx]
if not cpu:
driving_frame = driving_frame.cuda()
lowres_frame = lowres_frame.cuda()
kp_driving = kp_detector(driving_frame)
kp_norm = normalize_kp(kp_source=kp_source,
kp_driving=kp_driving,
kp_driving_initial=kp_driving_initial,
use_relative_movement=relative,
use_relative_jacobian=relative,
adapt_movement_scale=adapt_movement_scale)
out = generator(source, kp_source=kp_source, kp_driving=kp_norm)
resized_prediction = transforms.Resize(
(256, 256))(out['prediction'])[None]
loss = F.mse_loss(torch.squeeze(resized_prediction, 0),
lowres_frame.detach())
loss.backward()
optimizer_generator.step()
optimizer_generator.zero_grad()
optimizer_kp_detector.step()
optimizer_kp_detector.zero_grad()
predictions.append(
np.transpose(out['prediction'].data.cpu().numpy(),
[0, 2, 3, 1])[0])
scheduler_generator.step()
scheduler_kp_detector.step()
return predictions