gv1028 / videogan Goto Github PK

It seems there are quite a few differences between this implementation and the one from the paper. Namely the weight initializers for DeConv, Conv and BatchNorm, noise sample, and whether to use bias in the 3DdeConv layer. May I ask why these differences?

hello! Is there any pretrained models in your code? The video dataset is too large and I can't train a model from the very beginning. Thank you!

absl.flags._exceptions.IllegalFlagValueError: flag --video_dim=[32, 64, 64, 3]: Expect argument to be a string or int, found <type 'list'>

Python v2.7.12
TensorFlow v1.9.0 (gpu accelerated) [docker image tensorflow/tensorflow:latest-gpu]

CMD: python main.py

any suggested fixes or workarounds?

absl.flags._exceptions.IllegalFlagValueError: flag --video_dim=[32, 64, 64, 3]: Expect argument to be a string or int, found <type 'list'>

Hey,

On a totally unrelated note, I was just wondering if you used the annotation files provided in the dataset as the .mat are matlab v7.3 files and using them in Python as dict/arrays is seemingly impossible for me...

Hello,I read the original paper for VGAN, it says the authors normalize the images to [-1,1].
While I find in your tensorflow implement version, in founction:read_and_process_video,the normalization is:image batch * 2 - 1.
So I wonder if this is specifically for the dataset you use?Or this is a common choice for various datasets.
Thank you in advance!

Hello! Is it possible to download a pretrained model and generate a few samples?
Thanks