def visual_test(epoch):
print("===> Running visual test...")
for i, tup in enumerate(visual_test_set):
result = interpolate(model, load_img(tup[0]), load_img(tup[2]))
result = pil_to_tensor(result)
tag = 'data/visual_test_{}'.format(i)
board_writer.add_image(tag, result, epoch)
def stitch_frames(src_path, dest_path, output_fps=None, drop_frames=False):
tick_t = timer()
frames = [join(src_path, x) for x in listdir(src_path)]
frames = [x for x in frames if is_image(x)]
frames.sort()
print('===> Loading frames...')
if drop_frames:
frames = [load_img(x) for i, x in enumerate(frames) if i % 2 == 0]
else:
frames = [load_img(x) for i, x in enumerate(frames)]
print('===> Writing results...')
write_video(dest_path, frames, output_fps)
tock_t = timer()
print("Done. Took ~{}s".format(round(tock_t - tick_t)))
def test_linear_interp(validation_set=None):
if validation_set is None:
validation_set = get_validation_set()
total_ssim = 0
total_psnr = 0
iters = len(validation_set.tuples)
crop = CenterCrop(config.CROP_SIZE)
for tup in validation_set.tuples:
x1, gt, x2, = [pil_to_tensor(crop(load_img(p))) for p in tup]
pred = torch.mean(torch.stack((x1, x2), dim=0), dim=0)
total_ssim += ssim(pred, gt).item()
total_psnr += psnr(pred, gt).item()
avg_ssim = total_ssim / iters
avg_psnr = total_psnr / iters
print(f'avg_ssim: {avg_ssim}, avg_psnr: {avg_psnr}')
def test_on_validation_set(model, validation_set=None):
if validation_set is None:
validation_set = get_validation_set()
total_ssim = 0
total_psnr = 0
iters = len(validation_set.tuples)
crop = CenterCrop(config.CROP_SIZE)
for i, tup in enumerate(validation_set.tuples):
x1, gt, x2, = [crop(load_img(p)) for p in tup]
pred = interpolate(model, x1, x2)
gt = pil_to_tensor(gt)
pred = pil_to_tensor(pred)
total_ssim += ssim(pred, gt).item()
total_psnr += psnr(pred, gt).item()
print(f'#{i+1} done')
avg_ssim = total_ssim / iters
avg_psnr = total_psnr / iters
print(f'avg_ssim: {avg_ssim}, avg_psnr: {avg_psnr}')
def find_icme_optical_flow_magnitude():
tuples = tuples_from_custom("./video/icme/")
pil_to_numpy = lambda x: np.array(x)[:, :, ::-1]
for tup_index in range(len(tuples)):
tup = tuples[tup_index]
imgs = [load_img(x) for x in tup]
imgs = [pil_to_numpy(x) for x in imgs]
for pair_idx in range(2, len(imgs)):
img1 = imgs[pair_idx - 2]
img2 = imgs[pair_idx]
flow = cv.optflow.calcOpticalFlowSF(img1,
img2,
layers=3,
averaging_block_size=2,
max_flow=4)
n = np.sum(1 - np.isnan(flow), axis=(0, 1))
flow[np.isnan(flow)] = 0
flow_magnitude = np.linalg.norm(flow.sum(axis=(0, 1)) / n)
print(
f"Flow magnitude {tup[pair_idx-2]} to {tup[pair_idx]} is {flow_magnitude}"
)
def __getitem__(self, index):
frames = self.tuples[index]
x1, target, x2 = (pil_to_tensor(self.crop(data_manager.load_img(x))) for x in frames)
input = torch.cat((x1, x2), dim=0)
return input, target
processed_magnitude += v[1]
if processed_magnitude > check_vectors_magnitude_ratio * sum_magnitudes:
break
diff = angle_difference(v[0], avg_direction_angle)
if diff > check_vectors_max_angle_difference:
num_check_fails += 1
fail_ratio = num_check_fails / float(checked_vecs)
return (fail_ratio < check_vectors_max_error_ratio, avg_direction)
print(len(tuples_from_custom("video/raw/frames")))
img1 = load_img("video/raw/frames/FjU_x1106pg/255000.jpg")
img2 = load_img("video/raw/frames/FjU_x1106pg/255004.jpg")
pil_to_numpy = lambda x: np.array(x)[:, :, ::-1]
img1 = pil_to_numpy(img1)
img2 = pil_to_numpy(img2)
#flow = cv.optflow.calcOpticalFlowSF(img1, img2, layers=3, averaging_block_size=2, max_flow=4)
img_h = img1.shape[0]
img_w = img1.shape[1]
patch_h = 150
patch_w = 150
flow_threshold = 6
