search

pfnet-research / tgan

The implementation of Temporal Generative Adversarial Nets with Singular Value Clipping
https://pfnet-research.github.io/tgan
MIT License
78 stars 34 forks source link

Inception Score on UCF101 #11

Closed vladyushchenko closed 5 years ago

vladyushchenko commented 5 years ago

Hi, nice work. I have a small question about the inception score evaluation procedure. As referenced in https://github.com/pfnet-research/tgan/issues/8 the normalization file for C3D network was not provided with the code. The proposed solution has a file, that was calculated for image resolution 112 112, while you are using 128 171.

Could you, please, explain how it was calculated? Is it taken from .binaryproto normalization file of Caffe C3D model, finetuned on UCF https://github.com/facebook/C3D/tree/master/C3D-v1.0/examples/c3d_finetuning or the initial model trained on Sports-1M https://github.com/facebook/C3D/tree/master/C3D-v1.0/examples/c3d_feature_extraction ? Do you perform any additional transformations?

Thank you in advance, Vlad

rezoo commented 5 years ago

Hello Vlad,

Sorry for replying late. We calculated the inception score with the pre-trained model trained on Sports-1M dataset and fine-tuned on UCF101. The model can be downloaded at this URL. To compute the inception score, we resized the video into 128x128px, and sent it into the above pre-trained model.

Please reply or ask me directly via email if you could have some additional questions.

Sincerely, Saito

vladyushchenko commented 5 years ago

Closing issue, normalization file is converted from C3D binaryproto. Data is provided as plain array, but can be reshaped and saved as numpy array, as discussed in C3D issues