Fix Vaihingen datamodule #853
Conversation
github-actions
bot
added
datamodules
There was a problem hiding this comment.
I'm sure some of these comments apply to other existing data modules too. Someday I would like to have all public and private methods properly documented. We're using pydocstyle to try to enforce this, but it's broken, so it doesn't complain like it should when methods aren't fully documented.
|
I don't really understand why the minimum test is failing. |
|
The best I can tell is that torch is throwing a UserWarning about the |
|
Okay no idea. @adamjstewart any suggestions? Can we bump kornia's minimum version or ignore this warning? |
|
Since this is a warning (not an error) and the warning has been removed in newer versions of PyTorch, I went with silencing the warning over bumping the minimum version of PyTorch. Latest commit should hopefully get the tests to pass. |
adamjstewart
added
the
backwards-incompatible
|
@adamjstewart do you know why rtd is saying "/home/docs/checkouts/readthedocs.org/user_builds/torchgeo/checkouts/853/docs/tutorials/transforms.ipynb:: WARNING: Pygments lexer name 'ipython3' is not known"? Is that why the build failed? |
|
Should be fixed by #922, just need to rebase. |
calebrob6
force-pushed
the
datamodule/vaihingen
branch
from
2b39215 to
6c36f31
Compare
adamjstewart
force-pushed
the
datamodule/vaihingen
branch
from
ba6a82d to
9d98914
Compare
github-actions
bot
added
datasets
adamjstewart
dismissed
isaaccorley’s stale review
Requested changes have been implemented
* fix datamodule * requested changes to vaihingen * data loader * fix error an clarity * fix failing test * fix failing test crop augmentation * found a bug * remove same_batch param * Trying to get minimum tests to pass * Formatting * Formatting again * Update torchgeo/datamodules/vaihingen.py * Sort imports * Isort, yousort, we all sort * Same logic as deepglobe * More-specific types * Missing import Co-authored-by: Caleb Robinson <calebrob6@gmail.com> Co-authored-by: Adam J. Stewart <ajstewart426@gmail.com>
This PR closes #851. As suggested, based on the
OSCDDatamodule, random patch crops are taken during training. For validation and testing the batch size is fixed to 1 like inOSCDDatamodulebut instead of a fixed padding size, each sample is padded to the next larger multiple of 32 for the encoder-decoder architectures.