DEV: Add Auxiliary Functionality For SlideGraph (#208)

vqdang · web-flow · commit 90e84e5954e2 · 2021-12-07T13:26:55.000Z
Migrate existing .py in #180 to another branch. Included functionalities are: - Generic graph plotting onto canvas - Plat Scaler for calibrating probabilities - Pin scikit-image to avoid errors due to changes in the updated version. - [x] Add functionalities - [x] Add tests Co-authored-by: @vqdang
diff --git a/requirements.conda.yml b/requirements.conda.yml
@@ -15,7 +15,7 @@ dependencies:
   - jupyterlab
   - matplotlib
   - numpy
-  - opencv>=4.0
+  - opencv==4.5.4.60
   - openjpeg>=2.3.0
   - openslide
   - pandas
@@ -27,7 +27,7 @@ dependencies:
   - pytorch
   - pyyaml>=5.1
   - requests
-  - scikit-image
+  - scikit-image==0.18.3
   - scikit-learn>=0.23.2
   - shapely
   - tifffile
diff --git a/requirements.dev.conda.yml b/requirements.dev.conda.yml
@@ -18,7 +18,7 @@ dependencies:
   - jupyterlab
   - matplotlib
   - numpy
-  - opencv>=4.0
+  - opencv==4.5.4.60
   - openjpeg>=2.3.0
   - openslide
   - pandas
@@ -34,7 +34,7 @@ dependencies:
   - pyyaml>=5.1
   - recommonmark
   - requests
-  - scikit-image
+  - scikit-image==0.18.3
   - scikit-learn>=0.23.2
   - shapely
   - tifffile
diff --git a/requirements.txt b/requirements.txt
@@ -6,13 +6,13 @@ imagecodecs
 jupyterlab
 matplotlib
 numpy
-opencv-python>=4.0
+opencv-python==4.5.4.60
 openslide-python==1.1.2
 pandas
 pillow
 pyyaml>=5.1
 requests
-scikit-image
+scikit-image==0.18.3
 scikit-learn>=0.23.2
 shapely
 sphinx==4.1.2
diff --git a/requirements.win64.conda.yml b/requirements.win64.conda.yml
@@ -15,7 +15,7 @@ dependencies:
   - jupyterlab
   - matplotlib
   - numpy
-  - opencv>=4.0
+  - opencv==4.5.4.60
   - openjpeg>=2.3.0
   - pandas
   - pillow
@@ -25,7 +25,7 @@ dependencies:
   - pytorch
   - pyyaml>=5.1
   - requests
-  - scikit-image
+  - scikit-image==0.18.3
   - scikit-learn>=0.23.2
   - shapely
   - tifffile
diff --git a/requirements_dev.txt b/requirements_dev.txt
@@ -10,7 +10,7 @@ imagecodecs
 jupyterlab
 matplotlib
 numpy
-opencv-python>=4.0
+opencv-python==4.5.4.60
 openslide-python==1.1.2
 pandas
 pillow
@@ -22,7 +22,7 @@ pytest==6.2.5
 pyyaml>=5.1
 recommonmark
 requests
-scikit-image
+scikit-image==0.18.3
 scikit-learn>=0.23.2
 shapely
 sphinx==4.1.2
diff --git a/setup.py b/setup.py
@@ -16,13 +16,13 @@
     "numpy",
     "pillow",
     "matplotlib",
-    "opencv-python>=4.0",
+    "opencv-python==4.5.4.60",
     "openslide-python==1.1.2",
     "pyyaml>=5.1",
     "pandas",
     "glymur",
     "scikit-learn>=0.23.2",
-    "scikit-image>=0.17",
+    "scikit-image==0.18.3",
     "shapely",
     "torchvision==0.10.1",
     "torch==1.9.1",
diff --git a/tests/test_scale.py b/tests/test_scale.py
@@ -0,0 +1,37 @@
+"""Tests for scaling methods."""
+
+import numpy as np
+import pytest
+
+from tiatoolbox.tools.scale import PlattScaling
+
+
+def test_platt_scaler():
+    """Test for Platt scaler."""
+    np.random.seed(5)
+    sample_size = 1000
+    logit = np.random.rand(sample_size)
+    # binary class
+    label = np.concatenate(
+        [np.full(int(0.9 * sample_size), -1), np.full(int(0.1 * sample_size), 1)]
+    )
+    scaler = PlattScaling(num_iters=1)
+    scaler._fixer_a = 0.0
+    scaler._fixer_b = 0.0
+    _ = scaler.fit_transform(logit * 0.01, label)
+
+    scaler = PlattScaling(num_iters=1)
+    scaler._fixer_a = 0.0
+    scaler._fixer_b = 1.0
+    _ = scaler.fit_transform(logit * 0.01, label)
+
+    scaler = PlattScaling(num_iters=10)
+    _ = scaler.fit_transform(logit * 100, label)
+
+    label = np.concatenate([np.full(int(sample_size), -1)])
+    scaler = PlattScaling(num_iters=1)
+    _ = scaler.fit_transform(logit * 0.01, label)
+
+    with pytest.raises(ValueError, match=r".*same shape.*"):
+        scaler.fit_transform(logit, label[:2])
+    print(scaler)
diff --git a/tests/test_visualization.py b/tests/test_visualization.py
@@ -11,6 +11,7 @@
 from tiatoolbox.utils.visualization import (
     overlay_prediction_contours,
     overlay_prediction_mask,
+    plot_graph,
 )
 from tiatoolbox.wsicore.wsireader import get_wsireader
 
@@ -130,3 +131,18 @@ def test_overlay_instance_prediction():
     # test crash
     with pytest.raises(ValueError, match=r"`.*inst_colours`.*tuple.*"):
         overlay_prediction_contours(canvas, inst_dict, inst_colours=inst_colours)
+
+
+def test_plot_graph():
+    """Test plotting graph."""
+    canvas = np.zeros([10, 10])
+    nodes = np.array([[1, 1], [2, 2], [2, 5]])
+    edges = np.array([[0, 1], [1, 2], [2, 0]])
+    node_colors = np.array([[0, 0, 0]] * 3)
+    edge_colors = np.array([[1, 1, 1]] * 3)
+    plot_graph(
+        canvas,
+        nodes,
+        edges,
+    )
+    plot_graph(canvas, nodes, edges, node_colors=node_colors, edge_colors=edge_colors)
diff --git a/tiatoolbox/models/__init__.py b/tiatoolbox/models/__init__.py
@@ -28,6 +28,7 @@
     WSIPatchDataset,
 )
 from tiatoolbox.models.engine.semantic_segmentor import (
+    FeatureExtractor,
     IOSegmentorConfig,
     SemanticSegmentor,
     WSIStreamDataset,
diff --git a/tiatoolbox/models/architecture/__init__.py b/tiatoolbox/models/architecture/__init__.py
@@ -28,6 +28,7 @@
 import torch
 
 from tiatoolbox import rcParam
+from tiatoolbox.models.architecture.vanilla import CNNExtractor, CNNModel
 from tiatoolbox.models.dataset.classification import predefined_preproc_func
 from tiatoolbox.utils.misc import download_data
 
diff --git a/tiatoolbox/tools/graph.py b/tiatoolbox/tools/graph.py
@@ -14,7 +14,7 @@
 # along with this program; if not, write to the Free Software Foundation,
 # Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
 #
-# The Original Code is Copyright (C) 2021, TIALab, University of Warwick
+# The Original Code is Copyright (C) 2021, TIA Centre, University of Warwick
 # All rights reserved.
 # ***** END GPL LICENSE BLOCK *****
 
diff --git a/tiatoolbox/tools/pyramid.py b/tiatoolbox/tools/pyramid.py
@@ -14,7 +14,7 @@
 # along with this program; if not, write to the Free Software Foundation,
 # Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
 #
-# The Original Code is Copyright (C) 2021, TIALab, University of Warwick
+# The Original Code is Copyright (C) 2021, TIA Centre, University of Warwick
 # All rights reserved.
 # ***** END GPL LICENSE BLOCK *****
 
diff --git a/tiatoolbox/tools/scale.py b/tiatoolbox/tools/scale.py
@@ -0,0 +1,192 @@
+# ***** BEGIN GPL LICENSE BLOCK *****
+#
+# This program is free software; you can redistribute it and/or
+# modify it under the terms of the GNU General Public License
+# as published by the Free Software Foundation; either version 2
+# of the License, or (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with this program; if not, write to the Free Software Foundation,
+# Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
+#
+# The Original Code is Copyright (C) 2021, TIA Centre, University of Warwick
+# All rights reserved.
+# ***** END GPL LICENSE BLOCK *****
+
+"""Scaler for transforming input.
+
+Included classes and methods are utilized to either pre-process input
+(such as standardization) or post-process predictions (such as re-calibrating
+logits to proper probabilities).
+
+"""
+
+import numpy as np
+
+
+# Fit model output to the label range
+class PlattScaling:
+    """Platt scaling.
+
+    Fitting a logistic regression model to a classifier scores such that
+    the model outputs are transformed into a probability distribution over classes.
+
+    Args:
+        num_iters (int): Number of iterations for training.
+
+    Examples:
+        >>> import numpy as np
+        >>> logit = np.random.rand(10)
+        >>> # binary class
+        >>> label = np.random.randint(0, 2, 10)
+        >>> scaler = PlattScaling()
+        >>> probabilities = scaler.fit_transform(label, logit)
+
+    """
+
+    def __init__(self, num_iters=100):
+        self.a = None
+        self.b = None
+        self.num_iters = num_iters + 1
+        self._fixer_a = 1.0
+        self._fixer_b = 1.0
+
+    def fit(self, logits, labels):
+        """Fit function like sklearn.
+
+        Fit the sigmoid to the classifier scores logits and labels
+        using the Platt Method.
+
+        Args:
+            logits (array-like): Classifier output scores.
+            labels (array like): Classifier labels, must be `+1` vs `-1` or `1` vs `0`.
+        Returns:
+            Model with fitted coefficients a and b for the sigmoid function.
+
+        """
+
+        def mylog(v):
+            """Log with epilon."""
+            return np.log(v + 1.0e-200)
+
+        out = np.array(logits)
+        labels = np.array(labels)
+
+        if len(logits) != len(labels):
+            raise ValueError(
+                (
+                    f"`logits` and `labels` must have same shape: "
+                    f"{len(logits)} vs {len(labels)}"
+                )
+            )
+
+        target = labels == 1
+        prior1 = float(np.sum(target))
+        prior0 = len(target) - prior1
+        a_ = 0
+        b_ = np.log((prior0 + 1) / (prior1 + 1))
+        self.a, self.b = a_, b_
+
+        hi_target = (prior1 + 1) / (prior1 + 2)
+        lo_target = 1 / (prior0 + 2)
+        labda = 1e-3
+        olderr = 1e300
+        pp = np.ones(out.shape) * (prior1 + 1) / (prior0 + prior1 + 2)
+        idx_t = np.zeros(target.shape)
+        for _ in range(1, self.num_iters):
+            a = 0
+            b = 0
+            c = 0
+            d = 0
+            e = 0
+            for i, _ in enumerate(out):
+                if target[i]:
+                    t = hi_target
+                    idx_t[i] = t
+                else:
+                    t = lo_target
+                    idx_t[i] = t
+                d1 = pp[i] - t
+                d2 = pp[i] * (1 - pp[i])
+                a += out[i] * out[i] * d2
+                b += d2
+                c += out[i] * d2
+                d += out[i] * d1
+                e += d1
+
+            flag = abs(d) < 1.0e-9 and abs(e) < 1.0e-9
+            if flag:
+                break
+
+            old_a_ = a_
+            old_b_ = b_
+            count = 0
+            while 1:
+                det = (a + labda) * (b + labda) - c * c
+                if self._fixer_a * det == 0:
+                    labda *= 10
+                    continue
+                a_ = old_a_ + ((b + labda) * d - c * e) / det
+                b_ = old_b_ + ((a + labda) * e - c * d) / det
+
+                self.a, self.b = a_, b_
+                err = 0
+                for i, _ in enumerate(out):
+                    p = self.transform(out[i])
+                    pp[i] = p
+                    t = idx_t[i]
+                    err -= t * mylog(p) + (1 - t) * mylog(1 - p)
+
+                if err < self._fixer_a * olderr * (1 + 1e-7):
+                    labda *= 0.1
+                    break
+                labda *= 10
+
+                if self._fixer_b * labda > 1e6:
+                    break
+                diff = err - olderr
+                scale = 0.5 * (err + olderr + 1)
+
+                flag = -1e-3 * scale < diff < 1e-7 * scale
+                if flag:
+                    count += 1
+                else:
+                    count = 0
+                olderr = err
+
+                if count == 3:
+                    break
+        self.a, self.b = a_, b_
+        return self
+
+    def transform(self, logits):
+        """Tranform input to probabilities basing on trained parameters.
+
+        Args:
+            labels (array like): Classifier labels, must be `+1` vs `-1` or `1` vs `0`.
+        Returns:
+            Array of probabilities.
+
+        """
+        return 1 / (1 + np.exp(logits * self.a + self.b))
+
+    def fit_transform(self, logits, labels):
+        """Fit and tranform input to probabilities.
+
+        Args:
+            logits (array-like): Classifier output scores.
+            labels (array like): Classifier labels, must be `+1` vs `-1` or `1` vs `0`.
+        Returns:
+            Array of probabilities.
+
+        """
+        return self.fit(logits, labels).transform(logits)
+
+    def __repr__(self):
+        a, b = self.a, self.b
+        return "Platt Scaling: " + f"a: {a}, b: {b}"
diff --git a/tiatoolbox/tools/stainaugment.py b/tiatoolbox/tools/stainaugment.py
diff --git a/tiatoolbox/utils/visualization.py b/tiatoolbox/utils/visualization.py

Original file line number	Diff line number	Diff line change
`@@ -28,6 +28,7 @@`
`28`	`28`	`WSIPatchDataset,`
`29`	`29`	`)`
`30`	`30`	`from tiatoolbox.models.engine.semantic_segmentor import (`
	`31`	`+ FeatureExtractor,`
`31`	`32`	`IOSegmentorConfig,`
`32`	`33`	`SemanticSegmentor,`
`33`	`34`	`WSIStreamDataset,`