About I3D model for VFID score.

[kyuyeonpooh]

Hi, thanks for sharing great codes!

I have a question about the mode (training / testing) you have used for I3D networks for evaluating VFID score.

Free-Form-Video-Inpainting/src/evaluate.py

Lines 35 to 48 in 19bfc6f

	def init_i3d_model():
	global i3d_model
	if i3d_model is not None:
	return
	logger.info("Loading I3D model for FID score ..")
	i3d_model_weight = '../libs/model_weights/i3d_rgb_imagenet.pt'
	if not os.path.exists(i3d_model_weight):
	make_dirs(os.path.dirname(i3d_model_weight))
	urllib.request.urlretrieve('https://github.com/piergiaj/pytorch-i3d/'
	'raw/master/models/rgb_imagenet.pt', i3d_model_weight)
	i3d_model = InceptionI3d(400, in_channels=3, final_endpoint='Logits')
	i3d_model.load_state_dict(torch.load(i3d_model_weight))
	i3d_model.to(torch.device('cuda:0'))

As shown above in I3D model initialization part for I3D feature extracting, it seems that you haven't converted the model into evaluation mode by i3d_model.eval().

When I evaluate your model without converting I3D into eval mode, I was able to reproduce the similar results in the paper.
However, when I turned the model into evaluation mode (because using .eval() is common for using the model as feature extractor), I found the VFID score goes up to over 1.0.

Is there any reason why the I3D network is used as training mode?
Thank you for reading! Have a nice day.