MultiBands, and tools improvements (mostly)

AUTHORS.md

@@ -1,3 +1,5 @@
 Daniel J. Hofmann <[email protected]> https://github.com/daniel-j-h
 
 Bhargav Kowshik <[email protected]> https://github.com/bkowshik
+
+Olivier Courtin <[email protected]> https://github.com/ocourtin

config/dataset-parking.toml

@@ -21,3 +21,11 @@
 # Note: use `./rs weights -h` to compute these for new datasets.
 [weights]
   values = [1.6248, 5.762827]
+
+
+# Channels configuration let your indicate wich dataset sub-directory and bands to take as input
+# You could so, add several channels blocks to compose your input Tensor. Orders are meaningful.
+[[channels]]
+type = "file"
+sub = "images"
+bands = [1,2,3]

config/model-unet.toml

@@ -5,9 +5,6 @@
 # Model specific common attributes.
 [common]
 
-  # Use CUDA for GPU acceleration.
-  cuda       = true
-
   # Batch size for training.
   batch_size = 2
 
@@ -32,3 +29,6 @@
 
   # Loss function name (e.g 'Lovasz', 'mIoU' or 'CrossEntropy')
   loss = 'Lovasz'
+
+  # Data augmentation, Flip or Rotate probability
+  data_augmentation = 0.75

deps/requirements-lock.txt

@@ -20,15 +20,15 @@ more-itertools==4.2.0
 numpy==1.14.4
 opencv-contrib-python==3.4.1.15
 osmium==2.14.1
-Pillow==5.1.0
+Pillow-simd==5.3.0
 pluggy==0.6.0
 py==1.5.3
 pyparsing==2.2.0
 pyproj==1.9.5.1
 pytest==3.6.1
 python-dateutil==2.7.3
 pytz==2018.4
-rasterio==1.0b1
+rasterio==1.0.9
 requests==2.18.4
 Rtree==0.8.3
 scipy==1.1.0
@@ -37,7 +37,7 @@ six==1.11.0
 snuggs==1.4.1
 supermercado==0.0.5
 toml==0.9.4
-torch==0.4.0
+torch==0.4.1
 torchvision==0.2.1
 tqdm==4.23.4
 urllib3==1.22

deps/requirements.txt

@@ -1,6 +1,6 @@
 torchvision
 numpy
-pillow
+pillow-simd
 scipy
 opencv-contrib-python
 tqdm
@@ -10,7 +10,7 @@ geojson
 mercantile
 osmium
 matplotlib
-rasterio>=1.0a12
+rasterio
 supermercado
 shapely
 rtree

docker/Dockerfile.cpu

@@ -6,7 +6,7 @@ FROM ubuntu:16.04
 # See: https://github.com/skvark/opencv-python/issues/90
 RUN apt-get update -qq && \
     apt-get install -qq -y -o quiet=1 \
-    python3 python3-dev python3-tk python3-pip build-essential libboost-python-dev libexpat1-dev zlib1g-dev libbz2-dev libspatialindex-dev libsm6
+    python3 python3-dev python3-tk python3-pip build-essential libboost-python-dev libexpat1-dev zlib1g-dev libbz2-dev libspatialindex-dev libsm6 libwebp-dev libjpeg-turbo8-dev
 
 WORKDIR /app
 ADD . /app

docker/Dockerfile.gpu

@@ -6,7 +6,7 @@ FROM nvidia/cuda:9.1-cudnn7-runtime-ubuntu16.04
 # See: https://github.com/skvark/opencv-python/issues/90
 RUN apt-get update -qq && \
     apt-get install -qq -y -o quiet=1 \
-    python3 python3-dev python3-tk python3-pip build-essential libboost-python-dev libexpat1-dev zlib1g-dev libbz2-dev libspatialindex-dev libsm6
+    python3 python3-dev python3-tk python3-pip build-essential libboost-python-dev libexpat1-dev zlib1g-dev libbz2-dev libspatialindex-dev libsm6 libwebp-dev libjpeg-turbo8-dev
 
 WORKDIR /app
 ADD . /app

robosat/colors.py

@@ -2,6 +2,7 @@
 """
 
 import colorsys
+import numpy as np
 
 from enum import Enum, unique
 
@@ -49,9 +50,10 @@ def make_palette(*colors):
       colors: variable number of color names.
     """
 
+    assert 0 < len(colors) <= 256
     rgbs = [Mapbox[color].value for color in colors]
-    flattened = sum(rgbs, ())
-    return list(flattened)
+
+    return list(sum(rgbs, ()))
 
 
 def color_string_to_rgb(color):
@@ -84,12 +86,24 @@ def continuous_palette_for_color(color, bins=256):
     r, g, b = [v / 255 for v in Mapbox[color].value]
     h, s, v = colorsys.rgb_to_hsv(r, g, b)
 
-    palette = []
+    assert 0 < bins <= 256
 
+    palette = []
     for i in range(bins):
-        ns = (1 / bins) * (i + 1)
-        palette.extend([int(v * 255) for v in colorsys.hsv_to_rgb(h, ns, v)])
-
-    assert len(palette) // 3 == bins
+        r, g, b = [int(v * 255) for v in colorsys.hsv_to_rgb(h, (1 / bins) * (i + 1), v)]
+        palette.extend(r, g, b)
 
     return palette
+
+
+def complementary_palette(palette):
+
+    comp_palette = []
+    colors = [palette[i : i + 3] for i in range(0, len(palette), 3)]
+
+    for color in colors:
+        r, g, b = [v for v in color]
+        h, s, v = colorsys.rgb_to_hsv(r, g, b)
+        comp_palette.extend(map(int, colorsys.hsv_to_rgb((h + 0.5) % 1, s, v)))
+
+    return comp_palette

robosat/datasets.py

@@ -5,9 +5,13 @@
 See: http://pytorch.org/docs/0.3.1/data.html
 """
 
+import os
+import sys
 import torch
 from PIL import Image
 import torch.utils.data
+import cv2
+import numpy as np
 
 from robosat.tiles import tiles_from_slippy_map, buffer_tile_image
 
@@ -17,21 +21,35 @@ class SlippyMapTiles(torch.utils.data.Dataset):
     """Dataset for images stored in slippy map format.
     """
 
-    def __init__(self, root, transform=None):
+    def __init__(self, root, mode, transform=None):
         super().__init__()
 
         self.tiles = []
         self.transform = transform
 
         self.tiles = [(tile, path) for tile, path in tiles_from_slippy_map(root)]
         self.tiles.sort(key=lambda tile: tile[0])
+        self.mode = mode
 
     def __len__(self):
         return len(self.tiles)
 
     def __getitem__(self, i):
         tile, path = self.tiles[i]
-        image = Image.open(path)
+
+        if self.mode == "image":
+            image = cv2.cvtColor(cv2.imread(path), cv2.COLOR_BGR2RGB)
+
+        elif self.mode == "multibands":
+            image = cv2.imread(path, cv2.IMREAD_ANYCOLOR)
+            if len(image.shape) == 3 and image.shape[2] >= 3:
+                # FIXME Look twice to find an in-place way to perform a multiband BGR2RGB
+                g = image[:, :, 0]
+                image[:, :, 0] = image[:, :, 2]
+                image[:, :, 2] = g
+
+        elif self.mode == "mask":
+            image = np.array(Image.open(path).convert("P"))
 
         if self.transform is not None:
             image = self.transform(image)
@@ -40,42 +58,49 @@ def __getitem__(self, i):
 
 
 # Multiple Slippy Map directories.
-# Think: one with images, one with masks, one with rasterized traces.
 class SlippyMapTilesConcatenation(torch.utils.data.Dataset):
     """Dataset to concate multiple input images stored in slippy map format.
     """
 
-    def __init__(self, inputs, target, joint_transform=None):
+    def __init__(self, path, channels, target, joint_transform=None):
         super().__init__()
 
-        # No transformations in the `SlippyMapTiles` instead joint transformations in getitem
-        self.joint_transform = joint_transform
+        assert len(channels), "Channels configuration empty"
+        self.channels = channels
+        self.inputs = dict()
+
+        for channel in channels:
+            for band in channel["bands"]:
+                self.inputs[channel["sub"]] = SlippyMapTiles(os.path.join(path, channel["sub"]), mode="multibands")
 
-        self.inputs = [SlippyMapTiles(inp) for inp in inputs]
-        self.target = SlippyMapTiles(target)
+        self.target = SlippyMapTiles(target, mode="mask")
 
-        assert len(set([len(dataset) for dataset in self.inputs])) == 1, "same number of tiles in all images"
-        assert len(self.target) == len(self.inputs[0]), "same number of tiles in images and label"
+        # No transformations in the `SlippyMapTiles` instead joint transformations in getitem
+        self.joint_transform = joint_transform
 
     def __len__(self):
         return len(self.target)
 
     def __getitem__(self, i):
-        # at this point all transformations are applied and we expect to work with raw tensors
-        inputs = [dataset[i] for dataset in self.inputs]
 
-        images = [image for image, _ in inputs]
-        tiles = [tile for _, tile in inputs]
+        mask, tile = self.target[i]
 
-        mask, mask_tile = self.target[i]
+        for channel in self.channels:
+            try:
+                data, band_tile = self.inputs[channel["sub"]][i]
+                assert band_tile == tile
 
-        assert len(set(tiles)) == 1, "all images are for the same tile"
-        assert tiles[0] == mask_tile, "image tile is the same as label tile"
+                for band in channel["bands"]:
+                    data_band = data[:, :, int(band) - 1] if len(data.shape) == 3 else data_band
+                    data_band = data_band.reshape(mask.shape[0], mask.shape[1], 1)
+                    tensor = np.concatenate((tensor, data_band), axis=2) if "tensor" in locals() else data_band
+            except:
+                sys.exit("Unable to concatenate input Tensor")
 
         if self.joint_transform is not None:
-            images, mask = self.joint_transform(images, mask)
+            tensor, mask = self.joint_transform(tensor, mask)
 
-        return torch.cat(images, dim=0), mask, tiles
+        return tensor, mask, tile
 
 
 # Todo: once we have the SlippyMapDataset this dataset should wrap
@@ -113,7 +138,7 @@ def __len__(self):
 
     def __getitem__(self, i):
         tile, path = self.tiles[i]
-        image = buffer_tile_image(tile, self.tiles, overlap=self.overlap, tile_size=self.size)
+        image = np.array(buffer_tile_image(tile, self.tiles, overlap=self.overlap, tile_size=self.size))
 
         if self.transform is not None:
             image = self.transform(image)

robosat/log.py

@@ -10,9 +10,9 @@ class Log:
     """Create a log instance on a log file
     """
 
-    def __init__(self, path, out=sys.stdout):
+    def __init__(self, path, out=sys.stdout, mode="a"):
         self.out = out
-        self.fp = open(path, "a")
+        self.fp = open(path, mode)
         assert self.fp, "Unable to open log file"
 
     """Log a new message to the opened log file, and optionnaly on stdout or stderr too

robosat/metrics.py

@@ -24,16 +24,19 @@ def __init__(self, labels):
         self.fp = 0
         self.tp = 0
 
-    def add(self, actual, predicted):
+    def add(self, label, predicted, is_prob=True):
         """Adds an observation to the tracker.
 
         Args:
-          actual: the ground truth labels.
-          predicted: the predicted labels.
+          label: the ground truth labels.
+          predicted: the predicted prob or mask.
+          is_prob: as predicted could be either a prob or a mask.
         """
 
-        masks = torch.argmax(predicted, 0)
-        confusion = masks.view(-1).float() / actual.view(-1).float()
+        if is_prob:
+            predicted = torch.argmax(predicted, 0)
+
+        confusion = predicted.view(-1).float() / label.view(-1).float()
 
         self.tn += torch.sum(torch.isnan(confusion)).item()
         self.fn += torch.sum(confusion == float("inf")).item()
@@ -46,7 +49,13 @@ def get_miou(self):
         Returns:
           The mean Intersection over Union score for all observations seen so far.
         """
-        return np.nanmean([self.tn / (self.tn + self.fn + self.fp), self.tp / (self.tp + self.fn + self.fp)])
+
+        try:
+            miou = np.nanmean([self.tn / (self.tn + self.fn + self.fp), self.tp / (self.tp + self.fn + self.fp)])
+        except ZeroDivisionError:
+            miou = float("NaN")
+
+        return miou
 
     def get_fg_iou(self):
         """Retrieves the foreground Intersection over Union score.
@@ -58,7 +67,7 @@ def get_fg_iou(self):
         try:
             iou = self.tp / (self.tp + self.fn + self.fp)
         except ZeroDivisionError:
-            iou = float("Inf")
+            iou = float("NaN")
 
         return iou
 
@@ -74,7 +83,7 @@ def get_mcc(self):
                 (self.tp + self.fp) * (self.tp + self.fn) * (self.tn + self.fp) * (self.tn + self.fn)
             )
         except ZeroDivisionError:
-            mcc = float("Inf")
+            mcc = float("NaN")
 
         return mcc