loss.backward()
tr_optimizer.step()
##
## TB log val
if (opt.IsTbLog):
tb_info = {
'loss': loss_val,
'recon_loss': recon_loss_val,
'entropy_loss': entropy_loss_val
}
for tag, value in tb_info.items():
tb_logger.scalar_summary(tag, value, global_ite_idx)
# TB log img
if ((global_ite_idx % tb_img_log_interval) == 0):
frames_vis = utils.vframes2imgs(unorm_trans(frames.data),
step=5,
batch_idx=0)
frames_vis = np.concatenate(frames_vis, axis=-1)
frames_vis = frames_vis[None, :, :] * np.ones(
3, dtype=int)[:, None, None]
frames_recon_vis = utils.vframes2imgs(unorm_trans(
recon_frames.data),
step=5,
batch_idx=0)
frames_recon_vis = np.concatenate(frames_recon_vis, axis=-1)
frames_recon_vis = frames_recon_vis[None, :, :] * np.ones(
3, dtype=int)[:, None, None]
tb_info = {'x': frames_vis, 'x_rec': frames_recon_vis}
for tag, imgs in tb_info.items():
tb_logger.image_summary(tag, imgs, global_ite_idx)
##
model.to(device)
model.eval()
img_crop_size = 0
recon_error_list = [None] * len(video_data_loader)
# recon_error_list = []
progress_bar = tqdm(video_data_loader)
for batch_idx, frames in enumerate(progress_bar):
progress_bar.update()
frames = frames.reshape([batch_size, num_frame, ch, height, width])
frames = frames.permute(0, 2, 1, 3, 4)
frames = frames.to(device)
if (ModelName == 'AE'):
recon_frames = model(frames)
###### calculate reconstruction error (MSE)
recon_np = utils.vframes2imgs(unorm_trans(recon_frames.data), step=1, batch_idx=0)
input_np = utils.vframes2imgs(unorm_trans(frames.data), step=1, batch_idx=0)
r = utils.crop_image(recon_np, img_crop_size) - utils.crop_image(input_np, img_crop_size)
# recon_error = np.mean(sum(r**2)**0.5)
recon_error = np.mean(r ** 2) # **0.5
elif (ModelName == 'MemAE'):
recon_res = model(frames)
recon_frames = recon_res['output']
recon_np = utils.vframes2imgs(unorm_trans(recon_frames.data), step=1, batch_idx=0)
input_np = utils.vframes2imgs(unorm_trans(frames.data), step=1, batch_idx=0)
r = utils.crop_image(recon_np, img_crop_size) - utils.crop_image(input_np, img_crop_size)
sp_error_map = sum(r ** 2)**0.5
recon_error = np.mean(sp_error_map.flatten())
else:
recon_error = -1
print('Wrong ModelName.')