Merge pull request #74 from BloodAxe/develop

Release 0.5.1

Author: BloodAxe

.github/workflows/ci.yml

@@ -25,11 +25,11 @@ jobs:
         matrix.operating-system == 'ubuntu-latest' ||
         matrix.operating-system == 'windows-latest'
       run: >
-        pip install torch==1.8.1+cpu torchvision==0.9.1+cpu
+        pip install torch==1.10.1+cpu torchvision==0.11.2+cpu
         -f https://download.pytorch.org/whl/torch_stable.html
     - name: Install PyTorch on MacOS
       if: matrix.operating-system == 'macos-latest'
-      run: pip install torch==1.8.1 torchvision==0.9.1
+      run: pip install torch==1.10.1 torchvision==0.11.2
     - name: Install dependencies
       run: pip install .[${{ matrix.pytorch-toolbelt-version }}]
     - name: Install linters
@@ -59,6 +59,6 @@ jobs:
     - name: Update pip
       run: python -m pip install --upgrade pip
     - name: Install Black
-      run: pip install black==20.8b1
+      run: pip install black==22.3.0
     - name: Run Black
       run: black --config=black.toml --check .

.gitignore

@@ -20,3 +20,4 @@ var/
 .pytest_cache/
 /tests/tta_eval.csv
 /tests/tmp.onnx
+/tests/test_plot_confusion_matrix.png

notebooks/tiled_inference.ipynb

@@ -1,3 +1,3 @@
 from __future__ import absolute_import
 
-__version__ = "0.5.0"
+__version__ = "0.5.1"

pytorch_toolbelt/__init__.py

@@ -139,7 +139,7 @@ def __getitem__(self, index):
 
             if self.need_supervision_masks:
                 for i in range(1, 6):
-                    stride = 2 ** i
+                    stride = 2**i
                     mask = block_reduce(mask, (2, 2), partial(_block_reduce_dominant_label))
                     sample[name_for_stride(TARGET_MASK_KEY, stride)] = self.make_target(mask)
 

pytorch_toolbelt/datasets/segmentation.py

@@ -7,26 +7,27 @@
 from ..utils.support import pytorch_toolbelt_deprecated
 
 __all__ = [
+    "geometric_mean",
+    "harmonic_mean",
+    "logodd_mean",
+    "pad_image_tensor",
+    "torch_fliplr",
+    "torch_flipud",
     "torch_none",
+    "torch_rot180",
+    "torch_rot270",
     "torch_rot90",
-    "torch_rot90_cw",
     "torch_rot90_ccw",
-    "torch_transpose_rot90_cw",
-    "torch_transpose_rot90_ccw",
     "torch_rot90_ccw_transpose",
+    "torch_rot90_cw",
     "torch_rot90_cw_transpose",
-    "torch_rot180",
-    "torch_rot270",
-    "torch_fliplr",
-    "torch_flipud",
     "torch_transpose",
     "torch_transpose2",
     "torch_transpose_",
-    "pad_image_tensor",
+    "torch_transpose_rot90_ccw",
+    "torch_transpose_rot90_cw",
     "unpad_image_tensor",
     "unpad_xyxy_bboxes",
-    "geometric_mean",
-    "harmonic_mean",
 ]
 
 
@@ -240,3 +241,23 @@ def harmonic_mean(x: Tensor, dim: int, eps: float = 1e-6) -> Tensor:
     x = torch.mean(x, dim=dim)
     x = torch.reciprocal(x.clamp_min(eps))
     return x
+
+
+def logodd_mean(x: Tensor, dim: int, eps: float = 1e-6) -> Tensor:
+    """
+    Compute log-odd mean along given dimension.
+    logodd = log(p / (1 - p))
+
+    This implementation assume values are in range [0, 1] (Probabilities)
+    Args:
+        x: Input tensor of arbitrary shape
+        dim: Dimension to reduce
+
+    Returns:
+        Tensor
+    """
+    x = x.clamp(min=eps, max=1.0 - eps)
+    x = torch.log(x / (1 - x))
+    x = torch.mean(x, dim=dim)
+    x = torch.exp(x) / (1 + torch.exp(x))
+    return x

pytorch_toolbelt/inference/functional.py

@@ -60,7 +60,7 @@ class ImageSlicer:
     Helper class to slice image into tiles and merge them back
     """
 
-    def __init__(self, image_shape, tile_size, tile_step=0, image_margin=0, weight="mean"):
+    def __init__(self, image_shape: Tuple[int, int], tile_size, tile_step=0, image_margin=0, weight="mean"):
         """
 
         :param image_shape: Shape of the source image (H, W)
@@ -122,12 +122,6 @@ def __init__(self, image_shape, tile_size, tile_step=0, image_margin=0, weight="
             else:
                 margin_left = margin_right = margin_top = margin_bottom = image_margin
 
-            if (self.image_width + margin_left + margin_right) % self.tile_size[1] != 0:
-                raise ValueError()
-
-            if (self.image_height + margin_top + margin_bottom) % self.tile_size[0] != 0:
-                raise ValueError()
-
             self.margin_left = margin_left
             self.margin_right = margin_right
             self.margin_top = margin_top
@@ -337,6 +331,10 @@ def integrate_batch(self, batch: torch.Tensor, crop_coords):
         if batch.device != self.image.device:
             batch = batch.to(device=self.image.device)
 
+        # Ensure that input batch dtype match the target dtyle of the accumulator
+        if batch.dtype != self.image.dtype:
+            batch = batch.type_as(self.image)
+
         for tile, (x, y, tile_width, tile_height) in zip(batch, crop_coords):
             self.image[:, y : y + tile_height, x : x + tile_width] += tile * self.weight
             self.norm_mask[:, y : y + tile_height, x : x + tile_width] += self.weight

pytorch_toolbelt/inference/tiles.py

@@ -79,6 +79,8 @@ def _deaugment_averaging(x: Tensor, reduction: MaybeStrOrCallable) -> Tensor:
         x = F.geometric_mean(x, dim=0)
     elif reduction in {"hmean", "harmonic_mean"}:
         x = F.harmonic_mean(x, dim=0)
+    elif reduction == "logodd":
+        x = F.logodd_mean(x, dim=0)
     elif callable(reduction):
         x = reduction(x, dim=0)
     elif reduction in {None, "None", "none"}:

pytorch_toolbelt/inference/tta.py

@@ -264,7 +264,15 @@ def instantiate_activation_block(activation_name: str, **kwargs) -> nn.Module:
 
     act_params = {}
 
-    if "inplace" in kwargs and activation_name in {ACT_RELU, ACT_RELU6, ACT_LEAKY_RELU, ACT_SELU, ACT_CELU, ACT_ELU}:
+    if "inplace" in kwargs and activation_name in {
+        ACT_RELU,
+        ACT_RELU6,
+        ACT_LEAKY_RELU,
+        ACT_SELU,
+        ACT_SILU,
+        ACT_CELU,
+        ACT_ELU,
+    }:
         act_params["inplace"] = kwargs["inplace"]
 
     if "slope" in kwargs and activation_name in {ACT_LEAKY_RELU}:

pytorch_toolbelt/modules/activations.py

@@ -100,7 +100,7 @@ def __init__(self, feature_maps: List[int], decoder_features: int, mask_channels
         super().__init__()
 
         if not isinstance(decoder_features, list):
-            decoder_features = [decoder_features * (2 ** i) for i in range(len(feature_maps))]
+            decoder_features = [decoder_features * (2**i) for i in range(len(feature_maps))]
 
         blocks = []
         for block_index, in_enc_features in enumerate(feature_maps[:-1]):