bensonren / bdimnna Goto Github PK

Hi,

I found that the ballistics is implemented as (line127 in generate_ballistics.py)

 output[:, 1] = x2 - m / k / k * ((g*m + v2 * k) * exponential_part + g*t*k)

However, the formula in the original paper (Benchmarking Invertible Architectures on Inverse Problems) suggests

 output[:, 1] = x2 - m / k / k * (g*m + (v2 * k) * exponential_part + g*t*k)

Is this a bug?

Dear authors, I have run your codes to reproduce the results represented in your paper, but I cannot get the same results as yours. I just download the codes and did not change any parameters. I found the loss is a little different from yours. Is anything needed to modify when training or after training? I just run the evaluate.py after training the models. And can you tell me how to calculate the loss of the model on the meta-material dataset because it does not create a file of Ypred? The figure of the loss histogram is empty. Thanks a lot!