Open iraj465 opened 1 year ago
Which GPU are you using? In my experience the second pass failing is usually due to either running out of VRAM, or the initial mesh pass output being of very low quality. What was the loss you were seeing on the first pass?
The img-loss gets somewhat constant after 100-200 iterations
From the log above, it looks like the reshape operation is failing
reflvec = ru.xfm_vectors(reflvec.view(reflvec.shape[0], reflvec.shape[1] * reflvec.shape[2], reflvec.shape[3]), mtx).view(*reflvec.shape)
RuntimeError: shape '[2, 720, 1280, 3]' is invalid for input of size 2764800
which is strange, as I assume reflvec has shape '[2, 720, 1280, 3]
not '[1, 720, 1280, 3]
when this triggered.
That said, that part of the code should not execute unless you have
"camera_space_light" : true,
set in the config. This should only be set if you have a setup with moving object and static camera and light. If this is not your capture setup, results will be really bad.
In most other examples, we have a static object, static light and only the camera is moving, and for that setup, you should have
"camera_space_light" : false,
or remove that line from the config.
I have static camera and light but the object is rotating in a turntable. We would need camera_space_light: true
in this case right? Considering camera_space_light: true
being set, how can i resolve this?
Just after saving the dmtet_pass, the training is failing. I have printed the required shapes to be more clear, also batch_size of 2 is taken for 720x1280 res images.