Semantic Segmentation

contrib/segmentation/README.md

@@ -0,0 +1,15 @@
+# Semantic Segmentation using PyTorch and Azure Machine Learning
+
+This subproject contains a production ready training pipeline for a semantic segmentation model using PyTorch and Azure Machine Learning.
+
+## Installation
+
+To install the Azure ML CLI v2, [follow these instructions](https://docs.microsoft.com/en-us/azure/machine-learning/how-to-configure-cli)
+
+To install the last set of known working python dependencies run
+
+```bash
+pip install requirements.txt
+```
+
+Note that this project utilizes [pip-tools](https://github.com/jazzband/pip-tools) to manage its dependencies. Direct dependencies that the project requires are specified in `requirements.in` and may be upgraded to greater versions than that of `requirements.txt` at your own risk.

contrib/segmentation/config/augmentation.py

@@ -0,0 +1,42 @@
+from typing import Tuple
+import albumentations as A
+
+
+def _preprocessing(patch_dim: Tuple[int, int] = (512, 512)):
+    transform = A.Compose(
+        [
+            # This allows meaningful yet stochastic cropped views
+            A.CropNonEmptyMaskIfExists(patch_dim[0], patch_dim[1], p=1),
+            A.RandomRotate90(p=0.5),
+            A.HorizontalFlip(p=0.5),
+            A.VerticalFlip(p=0.5),
+            A.Blur(p=0.25),
+            A.ColorJitter(p=0.25),
+            A.GaussNoise(p=0.25),
+            A.CoarseDropout(p=0.5, max_holes=64, max_height=8, max_width=8),
+            A.RandomBrightnessContrast(p=0.25),
+        ],
+    )
+    return transform
+
+
+def _augmentation(patch_dim: Tuple[int, int] = (512, 512)):
+    transform = A.Compose(
+        [
+            # This allows meaningful yet stochastic cropped views
+            A.CropNonEmptyMaskIfExists(patch_dim[0], patch_dim[1], p=1),
+            A.RandomRotate90(p=0.5),
+            A.HorizontalFlip(p=0.5),
+            A.VerticalFlip(p=0.5),
+            A.Blur(p=0.25),
+            A.ColorJitter(p=0.25),
+            A.GaussNoise(p=0.25),
+            A.CoarseDropout(p=0.5, max_holes=64, max_height=8, max_width=8),
+            A.RandomBrightnessContrast(p=0.25),
+        ],
+    )
+    return transform
+
+
+preprocessing = _preprocessing()
+augmentation = _augmentation()