Train setting questions

[curie3170]

1. RCR setting

In the paper, you set the scale factor as 2.5, so I set scale value as (0.4,1.0) as below. But the ratio is just mentioned as random, so how should I set RCR params?
Compose([RandomCropandResize((args.image_height, args.image_width), scale=(0.4, 1.0), ratio = (0.5,2.0)), DownscaleFlow(), ToTensor()])

2. Optimizer setting
   Without any information, I assumed Adam optimizer as shown below. Have you used L2 regularization? If so, what value did you use for weight_decay?
   optimizer = optim.Adam(filter(lambda p: p.requires_grad, parameters()), lr=args.lr, weight_decay=args.weight_decay)

3. iteration vs. epoch
   The paper said that 100,000 iterations with batch_size=100 were used for the first stage to train pose network. If batch_size=100 and 400,000 images were trained, 4,000 iter is 1 epoch, so is it correct to train 25 epochs for the first stage?

[aojj]

because I can't find corresponding train data. Can I get your train code ?thank you!

[curie3170]

Hi, I found information from this repo: https://github.com/microsoft/COMPASS

[ghost]

Hi, I found information from this repo: https://github.com/microsoft/COMPASS

Thanks