Add PixelwiseRegressionTask #1241
Conversation
|
Not sure if this belongs in |
|
I vote for |
isaaccorley
force-pushed
the
trainers/dense-regression
branch
from
dd2f9a0
to
6a2c519
Compare
| @@ -35,8 +36,10 @@ class RegressionTask(LightningModule): # type: ignore[misc] | |||
| print(timm.list_models()) | |||
| """ | |||
|
|
|||
| def config_task(self) -> None: | |||
| """Configures the task based on kwargs parameters.""" | |||
| target_key: str = "label" | |||
There was a problem hiding this comment.
This allows to not have to duplicate all the train/val/test steps just to change label to mask. Let me know if you have any other suggestions.
tests/trainers/test_regression.py
Outdated
| ("inria", InriaAerialImageLabelingDataModule, 1, "mse"), | ||
| ("inria", InriaAerialImageLabelingDataModule, 2, "mae"), |
There was a problem hiding this comment.
Testing regression on Inria binary [0, 1] masks for now since we don't have a readily available pixelwise regression datamodule.
| y_hat = self(x) | ||
|
|
||
| loss = F.mse_loss(y_hat, y) | ||
| if y_hat.ndim != y.ndim: |
There was a problem hiding this comment.
If num_outputs=1 the target variable ground truth is missing the necessary channel dim e.g.
- (b,) instead of (b, 1)
- (b, h, w) instead of (b, 1, h, w)
while the output of the models will be:
- (b, 1)
- (b, 1, h, w)
| self.log("train_loss", loss) # logging to TensorBoard | ||
| self.train_metrics(y_hat, y) | ||
| self.train_metrics(y_hat, y.to(torch.float)) |
There was a problem hiding this comment.
Cast to float only for loss and metrics in case the plotting expects a different dtype
isaaccorley
force-pushed
the
trainers/dense-regression
branch
from
e05a3b5
to
d71a252
Compare
…rchgeo into trainers/dense-regression
|
Looks like pixel-wise regression won the poll. Can you rebase and use the new hydra-style configs? |
|
Dimension stuff looks super confusing. Wonder if there's a way to simplify that. We should consider abstracting the segmentation model stuff into a shared utility. Will update things to the new style in a separate PR. |
|
The dimension lines are essentially the same as the except it only does it if necessary. It's a result of not adding a channel dimension to the output from the dataset targets. |
This PR adds a
PixelwiseRegressionTaskwhich can be used for regression on 2D imagery, e.g. height estimation or other continuous per-pixel variables. It's basically a mixture of the currentRegressionTaskandSemanticSegmentationTaskin that it performs regression but uses smp models e.g. U-Net with a single output channel and L1 or L2 loss for training.Unless I'm mistaken, we don't have any datasets to actually test this on.
Closes #849