error throwing with multi threads in prepare_train_data.py about sfmlearner-pytorch HOT 12 CLOSED

Hi, @ClementPinard,
Thanks for the update. Now when I run prepare_train_data.py, it got stuck shortly after the script started running (just stuck and gave no output files), do you know why?

from sfmlearner-pytorch.

Is is still working with only one thread ?

from sfmlearner-pytorch.

No, it also got stuck with only one thread after several iterations.

from sfmlearner-pytorch.

I'm sorry I can't reproduce this bug :( can you test with a smaller version of KITTI raw maybe ?

from sfmlearner-pytorch.

Yeah. I got the same problem. It just got stuck here.

Do you know how to solve it? @ClementPinard @sanweiliti

from sfmlearner-pytorch.

It's not actually stuck. The progress bar only display how the tasks are dispatched. the bar being at 100% only means that every task is either finished or dispatched. If you wait a little you will see that the program will end.

I am making a coming to use concurrent futures library instead of joblib to use the function as_completed and display the real progress of finished tasks.

from sfmlearner-pytorch.

I just uploaded a commit that now uses pebble instead of joblib. That way the colorbar will reflect finished tasks.

from sfmlearner-pytorch.

Thank you very much. I waited for about 2 hours. Then the program ended.
So I'm running train.py now. But there's an error now.

I checked the size of depth and intrinsics are all same as you comment in the code.

from sfmlearner-pytorch.

The batch inverse function is only available in pytorch 1.0.1 what's your version? I highly suspect you have an outdated version of pytorch and you should update it :)

from sfmlearner-pytorch.

My version of pytorch is 1.0.0. It seems the batch inverse function can't work. And unfortunately，I can't update the pytorch. I want to ask if there's any way to realize it.

from sfmlearner-pytorch.

Before 1.0.1, the inverse function works only for 2D matrix, so you can just decompose the batch matrix, inverse every matrix and then reassemble everythin with a stack

M = torch.randn(10,3,3)  # batch matrix
invM_list = [m.inverse() for m in M]
invM = torch.stack(invM_list, dim=0)

from sfmlearner-pytorch.

Yeah. Thank you very much for your help~

from sfmlearner-pytorch.