def __getitem__(self, idx):
data = copy.deepcopy(self.datalist[idx])
img_path, skeleton_path, original_shape, action_label, frame_num, start_frame_idx = data[
'img_path'], data['skeleton_path'], data['original_shape'], data[
'action_label'], data['frame_num'], data['start_frame_idx']
# video load
video, skeleton_frame_idxs = load_video(img_path, frame_num,
start_frame_idx)
resized_shape = video.shape[1:3]
# augmentation
video, img2aug_trans, aug2img_trans, do_flip = augmentation(
video, self.data_split)
video = video.transpose(0, 3, 1, 2).astype(
np.float32) / 255. # frame_num, channel_dim, height, width
# skeleton information load
pose_coords, pose_scores = self.load_skeleton(skeleton_path,
skeleton_frame_idxs,
original_shape,
resized_shape)
# process skeleton information
pose_coords, pose_scores = process_skeleton(pose_coords, pose_scores,
img2aug_trans, do_flip,
self.flip_pairs,
self.joint_num,
resized_shape)
"""
# for debug
# keypoint visualization
for i in range(cfg.frame_per_seg):
img = video[i,::-1,:,:].transpose(1,2,0) * 255
person_num = len(pose_coords[i])
for p in range(person_num):
#for j in range(self.joint_num):
#coord = (int(pose_coords[i][p][j][0]), int(pose_coords[i][p][j][1]))
#cv2.circle(img, coord, radius=3, color=(255,0,0), thickness=-1, lineType=cv2.LINE_AA)
#cv2.imwrite(str(idx) + '_' + str(action_label) + '_' + str(i) + '_' + str(j) + '.jpg', img)
coord = pose_coords[i][p].copy()
coord[:,0] = coord[:,0] / cfg.input_hm_shape[1] * cfg.input_img_shape[1]
coord[:,1] = coord[:,1] / cfg.input_hm_shape[0] * cfg.input_img_shape[0]
img = vis_keypoints(img, pose_coords[i][p] * 4, self.skeleton)
cv2.imwrite(str(idx) + '_' + str(action_label) + '_' + str(i) + '.jpg', img)
"""
inputs = {
'video': video,
'pose_coords': pose_coords,
'pose_scores': pose_scores
}
targets = {'action_label': action_label}
meta_info = {'img_id': data['img_id']}
return inputs, targets, meta_info
if len(losses) > 50:
done = True
for i in range(-51, -2):
if (losses[i] - losses[i + 1]) > loss_tolerance * losses[i]:
done = False
losses.append(loss)
return image, done
if __name__ == "__main__":
style_path = "../images/style.jpg"
style_image, _ = load_img(style_path)
images, yuvs = load_video("../videos/cat2.mp4", frame_interval)
styled_images = []
losses = []
start = time.time()
content_layers = ["block5_conv2"]
style_layers = [
"block1_conv1",
"block2_conv1",
"block3_conv1",
"block4_conv1",
"block5_conv1",
]