-
Notifications
You must be signed in to change notification settings - Fork 35
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Data mismatch in F.conv2d() module #362
Comments
After some data tracing, I found that transformation function defined in torchgeo.datamodule.VHR10DataModule.init() transforms data, and move image data back to CPU.
Input data and output data of AugPipe().
|
Found that is is known issue of Kornia: “AugmentationSequential explicitly moves the output to the CPU if data_keys is given” (kornia/kornia#3066). Data is moved to CPU when AugmentationSequential() is given data_type with more than one data item in it. Why this bug has not appeared with terratorch ?
Two tentative solutions.
This works, but dependency requirement is not fully satisfied. Solution-2: Use latest Kornia from github.
This is tentative solution until new version of kornia (0.7.5 or 0.8) is released. Then we would pin down the version in requirements/required.txt |
@takaomoriyama According to my tests, it seems that the input tensor
Ad that step is passing. |
There is another ocurrence in ~ |
@Joao-L-S-Almeida Yes, |
@Joao-L-S-Almeida Can you please elaborate on the issue you are referring on line 110 of object_detection_task.py ?
|
The tensors |
@Joao-L-S-Almeida I see. I confirmed only x is on CPU. So I added |
Now that kornia 0.8.0 is released at pypi, the code should run without the workaround. |
@Joao-L-S-Almeida Removed the workaround in the commit, and confirmed the code ran with ordinary set up procedure described in README_Object_Detection.md. Thanks for your help. |
Describe the issue
When running a task for Object Detection (under development in branch), fit process stops due to data mismatch in F.conv2d() module.
To Reproduce (optional, but appreciated)
Steps to reproduce the behavior:
Screenshots or log output (optional)
I found that
input
data passed toF.conv2d()
is a tensor on CPU, whileweight
is a tensor on GPU.The text was updated successfully, but these errors were encountered: