def process_video(
item: Tuple[pd.Index, pd.Series],
source_dir: Path,
facedestination_dir: Path,
checkpoint_folder: Path,
face_size: int,
face_extractor: FaceExtractor,
lazycheck: bool = False,
deepcheck: bool = False,
) -> (pd.DataFrame, Path, pd.DataFrame, Path, List[Tuple[Image.Image,
Path]]) or None:
idx, record = item
# Checkpoint
video_faces_checkpoint_path = checkpoint_folder.joinpath(
record['path']).with_suffix('.faces.pkl')
if not lazycheck:
if video_faces_checkpoint_path.exists():
try:
df_video_faces = pd.read_pickle(
str(video_faces_checkpoint_path))
for _, r in df_video_faces.iterrows():
face_path = facedestination_dir.joinpath(r.name)
assert (face_path.exists())
if deepcheck:
img = Image.open(face_path)
img_arr = np.asarray(img)
assert (img_arr.ndim == 3)
assert (np.prod(img_arr.shape) > 0)
except Exception as e:
print('Error while checking: {}'.format(
video_faces_checkpoint_path))
print(e)
video_faces_checkpoint_path.unlink()
if not (video_faces_checkpoint_path.exists()):
try:
video_face_dict_list = []
# Load faces
frames = face_extractor.process_video(
source_dir.joinpath(record['path']))
if len(frames) == 0:
return
face_extractor.keep_only_best_face(frames)
for frame_idx, frame in enumerate(frames):
frames[frame_idx]['subjects'] = [0] * len(
frames[frame_idx]['detections'])
# Extract and save faces, bounding boxes, keypoints
images_to_save: List[Tuple[Image.Image, Path]] = []
for frame_idx, frame in enumerate(frames):
if len(frames[frame_idx]['detections']):
fullframe = Image.fromarray(frames[frame_idx]['frame'])
# Preserve the only found face even if not a good one, otherwise preserve only clusters > -1
subjects = np.unique(frames[frame_idx]['subjects'])
if len(subjects) > 1:
subjects = np.asarray([s for s in subjects if s > -1])
for face_idx, _ in enumerate(frame['faces']):
subj_id = frames[frame_idx]['subjects'][face_idx]
if subj_id in subjects: # Exclude outliers if other faces detected
face_path = facedestination_dir.joinpath(
record['path'],
'fr{:03d}_subj{:1d}.jpg'.format(
frames[frame_idx]['frame_idx'], subj_id))
face_dict = {
'facepath':
str(face_path.relative_to(
facedestination_dir)),
'video':
idx,
'label':
record['label'],
'videosubject':
subj_id
}
for field_idx, key in enumerate(
blazeface.BlazeFace.detection_keys):
face_dict[key] = frames[frame_idx][
'detections'][face_idx][field_idx]
cropping_bb = adapt_bb(
frame_height=fullframe.height,
frame_width=fullframe.width,
bb_height=face_size,
bb_width=face_size,
left=face_dict['xmin'],
top=face_dict['ymin'],
right=face_dict['xmax'],
bottom=face_dict['ymax'])
face = fullframe.crop(cropping_bb)
for key in blazeface.BlazeFace.detection_keys:
if (key[0] == 'k'
and key[-1] == 'x') or (key[0] == 'x'):
face_dict[key] -= cropping_bb[0]
elif (key[0] == 'k'
and key[-1] == 'y') or (key[0] == 'y'):
face_dict[key] -= cropping_bb[1]
face_dict['left'] = face_dict.pop('xmin')
face_dict['top'] = face_dict.pop('ymin')
face_dict['right'] = face_dict.pop('xmax')
face_dict['bottom'] = face_dict.pop('ymax')
face_path.parent.mkdir(parents=True, exist_ok=True)
images_to_save.append((face, face_path))
video_face_dict_list.append(face_dict)
if len(video_face_dict_list) > 0:
df_video_faces = pd.DataFrame(video_face_dict_list)
df_video_faces.index = df_video_faces['facepath']
del df_video_faces['facepath']
# type conversions
for key in [
'kp1x', 'kp1y', 'kp2x', 'kp2y', 'kp3x', 'kp3y', 'kp4x',
'kp4y', 'kp5x', 'kp5y', 'kp6x', 'kp6y', 'left', 'top',
'right', 'bottom'
]:
df_video_faces[key] = df_video_faces[key].astype(np.int16)
df_video_faces['conf'] = df_video_faces['conf'].astype(
np.float32)
df_video_faces['video'] = df_video_faces['video'].astype(
'category')
video_faces_checkpoint_path.parent.mkdir(parents=True,
exist_ok=True)
else:
print('No faces extracted for video {}'.format(record['path']))
df_video_faces = pd.DataFrame()
return df_video_faces, video_faces_checkpoint_path, images_to_save
except Exception as e:
print('Error while processing: {}'.format(record['path']))
print("-" * 60)
traceback.print_exc(file=sys.stdout, limit=5)
print("-" * 60)
return
print('Model loaded!')
transf = utils.get_transformer(face_policy,
face_size,
net.get_normalizer(),
train=False)
facedet = BlazeFace().to(device)
facedet.load_weights("blazeface/blazeface.pth")
facedet.load_anchors("blazeface/anchors.npy")
videoreader = VideoReader(verbose=False)
video_read_fn = lambda x: videoreader.read_frames(x,
num_frames=frames_per_video)
face_extractor = FaceExtractor(video_read_fn=video_read_fn, facedet=facedet)
vid_real_faces = face_extractor.process_video(
'samples/490868123550446422495477631417.mp4')
vid_fake_faces = face_extractor.process_video(
'samples/284649338838012868101332189709.mp4')
## Predict scores for each frame
# For each frame, we consider the face with the highest confidence score found by BlazeFace (= frame['faces'][0])
faces_real_t = torch.stack([
transf(image=frame['faces'][0])['image'] for frame in vid_real_faces
if len(frame['faces'])
])
faces_fake_t = torch.stack([
transf(image=frame['faces'][0])['image'] for frame in vid_fake_faces
if len(frame['faces'])
])
# In[9]:
torch.cuda.device(0)
# In[10]:
torch.cuda.device_count()
# In[11]:
torch.cuda.get_device_name(0)
# In[12]:
vid_real_faces = face_extractor.process_video('samples/lynaeydofd.mp4')
vid_fake_faces = face_extractor.process_video('samples/mqzvfufzoq.mp4')
# In[13]:
im_real_face = vid_real_faces[0]['faces'][0]
im_fake_face = vid_fake_faces[0]['faces'][0]
# In[14]:
fig, ax = plt.subplots(1, 2, figsize=(8, 4))
ax[0].imshow(im_real_face)
ax[0].set_title('REAL')
ax[1].imshow(im_fake_face)
def run_nb(modelname):
# ## Parameters
# In[2]:
"""
Choose an architecture between
- EfficientNetB4
- EfficientNetB4ST
- EfficientNetAutoAttB4
- EfficientNetAutoAttB4ST
- Xception
"""
net_model = modelname
"""
Choose a training dataset between
- DFDC
- FFPP
"""
train_db = 'DFDC'
# In[3]:
device = torch.device(
'cuda:0') if torch.cuda.is_available() else torch.device('cpu')
face_policy = 'scale'
face_size = 224
frames_per_video = 32
# ## Initialization
# In[4]:
print('=' * 20)
model_url = weights.weight_url['{:s}_{:s}'.format(net_model, train_db)]
print('=' * 20)
net = getattr(fornet, net_model)().eval().to(device)
print('=' * 20)
net.load_state_dict(
load_url(model_url, map_location=device, check_hash=True))
# In[5]:
transf = utils.get_transformer(face_policy,
face_size,
net.get_normalizer(),
train=False)
# In[6]:
facedet = BlazeFace().to(device)
facedet.load_weights("../blazeface/blazeface.pth")
facedet.load_anchors("../blazeface/anchors.npy")
videoreader = VideoReader(verbose=False)
video_read_fn = lambda x: videoreader.read_frames(
x, num_frames=frames_per_video)
face_extractor = FaceExtractor(video_read_fn=video_read_fn,
facedet=facedet)
# ## Detect faces
# In[7]:
torch.cuda.is_available()
# In[8]:
torch.cuda.current_device()
# In[9]:
torch.cuda.device(0)
# In[10]:
torch.cuda.device_count()
# In[11]:
torch.cuda.get_device_name(0)
# In[12]:
vid_real_faces = face_extractor.process_video('samples/lynaeydofd.mp4')
vid_fake_faces = face_extractor.process_video('samples/mqzvfufzoq.mp4')
# In[13]:
im_real_face = vid_real_faces[0]['faces'][0]
im_fake_face = vid_fake_faces[0]['faces'][0]
# In[14]:
fig, ax = plt.subplots(1, 2, figsize=(8, 4))
ax[0].imshow(im_real_face)
ax[0].set_title('REAL')
ax[1].imshow(im_fake_face)
ax[1].set_title('FAKE')
# ## Predict scores for each frame
# In[15]:
# For each frame, we consider the face with the highest confidence score found by BlazeFace (= frame['faces'][0])
faces_real_t = torch.stack([
transf(image=frame['faces'][0])['image'] for frame in vid_real_faces
if len(frame['faces'])
])
faces_fake_t = torch.stack([
transf(image=frame['faces'][0])['image'] for frame in vid_fake_faces
if len(frame['faces'])
])
with torch.no_grad():
faces_real_pred = net(faces_real_t.to(device)).cpu().numpy().flatten()
faces_fake_pred = net(faces_fake_t.to(device)).cpu().numpy().flatten()
# In[16]:
fig, ax = plt.subplots(1, 2, figsize=(12, 4))
ax[0].stem([f['frame_idx'] for f in vid_real_faces if len(f['faces'])],
expit(faces_real_pred),
use_line_collection=True)
ax[0].set_title('REAL')
ax[0].set_xlabel('Frame')
ax[0].set_ylabel('Score')
ax[0].set_ylim([0, 1])
ax[0].grid(True)
ax[1].stem([f['frame_idx'] for f in vid_fake_faces if len(f['faces'])],
expit(faces_fake_pred),
use_line_collection=True)
ax[1].set_title('FAKE')
ax[1].set_xlabel('Frame')
ax[1].set_ylabel('Score')
ax[1].set_ylim([0, 1])
ax[1].set_yticks([0, 1], ['REAL', 'FAKE'])
# In[17]:
"""
Print average scores.
An average score close to 0 predicts REAL. An average score close to 1 predicts FAKE.
"""
print('Average score for REAL video: {:.4f}'.format(
expit(faces_real_pred.mean())))
print('Average score for FAKE face: {:.4f}'.format(
expit(faces_fake_pred.mean())))
