Fix for segmentation tests

Author: ashnair1

torchgeo/datamodules/spacenet.py

@@ -53,12 +53,12 @@ def __init__(
             K.RandomVerticalFlip(p=0.5),
             K.RandomSharpness(p=0.5),
             K.ColorJitter(p=0.5, brightness=0.1, contrast=0.1, saturation=0.1, hue=0.1),
-            data_keys=['image', 'mask'],
+            data_keys=None,
         )
         self.aug = K.AugmentationSequential(
             K.Normalize(mean=self.mean, std=self.std),
             K.PadTo((448, 448)),
-            data_keys=['image', 'mask'],
+            data_keys=None,
         )
 
     def setup(self, stage: str) -> None:

torchgeo/trainers/segmentation.py

@@ -9,6 +9,7 @@
 import matplotlib.pyplot as plt
 import segmentation_models_pytorch as smp
 import torch.nn as nn
+from einops import rearrange
 from matplotlib.figure import Figure
 from torch import Tensor
 from torchmetrics import MetricCollection
@@ -225,6 +226,9 @@ def training_step(
         Returns:
             The loss tensor.
         """
+        if 'mask' in batch and batch['mask'].shape[1] == 1:
+            batch['mask'] = rearrange(batch['mask'], 'b () h w -> b h w')
+
         x = batch['image']
         y = batch['mask']
         batch_size = x.shape[0]
@@ -245,6 +249,8 @@ def validation_step(
             batch_idx: Integer displaying index of this batch.
             dataloader_idx: Index of the current dataloader.
         """
+        if 'mask' in batch and batch['mask'].shape[1] == 1:
+            batch['mask'] = rearrange(batch['mask'], 'b () h w -> b h w')
         x = batch['image']
         y = batch['mask']
         batch_size = x.shape[0]
@@ -289,6 +295,8 @@ def test_step(self, batch: Any, batch_idx: int, dataloader_idx: int = 0) -> None
             batch_idx: Integer displaying index of this batch.
             dataloader_idx: Index of the current dataloader.
         """
+        if 'mask' in batch and batch['mask'].shape[1] == 1:
+            batch['mask'] = rearrange(batch['mask'], 'b () h w -> b h w')
         x = batch['image']
         y = batch['mask']
         batch_size = x.shape[0]