menyifang / ADGAN

The Implementation of paper "Controllable Person Image Synthesis with Attribute-Decomposed GAN" CVPR 2020 (Oral); Pose and Appearance Attributes Transfer;
476 stars 90 forks source link

the number of learning iteration or epoch #32

Open ryhhtn opened 3 years ago

ryhhtn commented 3 years ago

In the description of paper, the number of trials is 120K. No other items, epoch or mini batch size, are mentioned

On the other hand, the pretrained model says 1000 epoch.

What criteria can be used to reproduce the same environment as the pre-trained model published in the paper or GitHub Code?

https://github.com/menyifang/ADGAN/blob/4dd70649ad136829b92dd6a1a823af7594a0220f/data/keypoint.py#L106

In addition, the paper states that 101,966 pairs were used as train data, but as shown above, the source code is limited to 4,000. Was this done to limit the number of iterations when learning at 1,000 epoch?