How to test the model? And why flow training is too slow?
haoyan14 opened this issue · comments
How to input parameters for testing after training the radiation model? And why is the training of flow model so slow? Training the flow model on 3090Ti only ran less than 10000 steps a day.
How for testing the model, you can add the --render_test flag to your command.
The flow model is slow due to the fact that the flow field becomes non-smooth over the course of training. So integrating the flow field is a bit unstable and requires many numerical computations. By adding suitable regularization losses (and increasing the integration tolerance of the flow integration) you can make flow training run faster.
Thank you for your answer. But I just have another question. What does the ".hdr" file in the training data represents for, and what does its each 4-channel pixel value o represents for? What does the ".npy" file in the training data represents for, and what does each 4-channel pixel value represents for? I hope you can help explain.
Thank you for your answer. But I just have another question. What does the ".hdr" file in the training data represents for, and what does its each 4-channel pixel value o represents for? What does the ".npy" file in the training data represents for, and what does each 4-channel pixel value represents for? I hope you can help explain.
Hi, the HDR file contains the depth of each pixel in the image -- each element of the ".npy" represents the optical flow from one image to the subsequent one.
Thanks！