About training details
Opened this issue · 0 comments
fuenwang commented
Hi, thanks for contributing this great work and dataset!
I would like to implement a pytorch version and some training details seems missing. Could you provide me the details like
- batch size
- the parameter of RMSProp (momentum, weight_decay, .....)
- Data augmentation
- Model initialization (xavier or uniform....)
- Is your conv/deconv layer including batchnorm layer?
- In your python code, you define " decay_heatmap(heatmap, sigma2=4):", in which sigma is 4, but the sigma in the paper is 2 instead. Which one is the final value you use for training?
Thanks very much!
Bests.