os.mkdir(output_dir)
# this directory will contain the frames to build the video
frame_dir = os.path.join(opt.checkpoints_dir, opt.name, 'frames')
if os.path.isdir(frame_dir):
shutil.rmtree(frame_dir)
os.mkdir(frame_dir)
FRAMES_PER_EPOCH = 48
frame_index = 1
for data in dataset:
t = data['left_frame']
video_utils.save_tensor(
t,
frame_dir + "/frame-%s.jpg" % str(frame_index).zfill(5),
text="original video",
)
frame_index += 1
current_frame = t
frame_count = (opt.pstop - opt.pstart + 1) * FRAMES_PER_EPOCH
pbar = tqdm(total=frame_count)
duration_s = frame_count / opt.fps
video_id = "progress_epoch-%s-%s_%s_%.1f-s_%.1f-fps%s" % (
str(opt.pstart), str(opt.pstop), opt.name, duration_s, opt.fps,
"_with-%d-zoom" % opt.zoom_lvl if opt.zoom_lvl != 0 else "")
for epoch_index in range(opt.pstart, opt.pstop + 1):
############## Forward Pass - frame t -> frame t+1 ######################
if opt.scheduled_sampling and (
latest_generated_frame
is not None) and np.random.randn(1) < opt.ss_recursion_prob:
left_frame = latest_generated_frame.detach()
recursion += 1
else:
left_frame = Variable(data['left_frame'])
recursion = 0
right_frame = Variable(data['right_frame'])
if opt.debug:
video_utils.save_tensor(
left_frame,
debug_dir + "/step-%d-left-r%d.jpg" % (total_steps, recursion))
video_utils.save_tensor(
right_frame, debug_dir + "/step-%d-right.jpg" % total_steps)
losses, latest_generated_frame = model(left_frame,
None,
right_frame,
None,
infer=opt.scheduled_sampling)
# sum per device losses
losses = [
torch.mean(x) if not isinstance(x, int) else x for x in losses
]
loss_dict = dict(zip(model.module.loss_names, losses))
shutil.rmtree(frame_dir)
os.mkdir(frame_dir)
frame_index = 0
if opt.start_from == "noise":
# careful, default value is 1024x512
t = torch.rand(1, 3, opt.fineSize, opt.loadSize)
elif opt.start_from == "video":
# use initial frames from the dataset
for data in dataset:
t = data['left_frame'] #kokokaeru! Specify first raw parts
video_utils.save_tensor(
t,
frame_dir + "/frame-%s.png" % str(frame_index).zfill(5),
text="",
)
frame_index += 1
else:
# use specified image
filepath = opt.start_from
if os.path.isfile(filepath):
t = video_utils.im2tensor(Image.open(filepath))
for i in range(50):
video_utils.save_tensor(
t,
frame_dir + "/frame-%s.png" % str(frame_index).zfill(5),
)
frame_index += 1
for i, data in enumerate(tqdm(positive_ds)):
cur_frame = Image.open(data['left_path'])
next_frame = Image.open(data['right_path'])
cur_frame = video_utils.im2tensor(cur_frame)
next_frame = video_utils.im2tensor(next_frame)
if opt.gpu:
cur_frame = cur_frame.to('cuda')
next_frame = next_frame.to('cuda')
generated_next_frame = video_utils.next_frame_prediction(
model, cur_frame)
video_utils.save_tensor(
generated_next_frame,
frame_dir + "/frame-%s.png" % str(frame_index).zfill(5),
)
frame_index += 1
print('Finished generating images')
duration_s = frame_index / opt.fps
video_id = "epoch-%s_%s_%.1f-s_%.1f-fps" % (str(
opt.which_epoch), opt.name, duration_s, opt.fps)
print(f'created video id {video_id}')
video_path = output_dir + "/" + video_id + ".mp4"
while os.path.isfile(video_path):
video_path = video_path[:-4] + "-.mp4"
print(f'modified video path {video_path}')
video_utils.video_from_frame_directory(frame_dir,
video_path,