Features

Features
 * FrameGetter supports len()
 * Per-Class Accuracy metric

mpctools/extensions/cvext.py

@@ -14,6 +14,7 @@
 
 Author: Michael P. J. Camilleri
 """
+import glob
 
 from numba import jit, uint8, uint16, double
 from queue import Queue, Empty, Full
@@ -1202,6 +1203,19 @@ def __getitem__(self, item):
             assert os.path.exists(_pth), f"Image {item} does not exist at {_pth}."
             return cv2.imread(_pth)
 
+    def __len__(self):
+        """
+        Retrieves the number of Frames
+
+        This is 'approximate', by counting the number of files matching the extension.
+
+        :return: Length
+        """
+        if self.Fmt.lower() == "video":
+            return int(cv2.VideoCapture(self.Path).get(cv2.CAP_PROP_FRAME_COUNT))
+        else:
+            return len(glob.glob(os.path.join(self.Path, f'*{os.path.splitext(self.Fmt)[1]}')))
+
 
 class SwCLAHE:
     """

mpctools/extensions/skext.py

@@ -14,10 +14,32 @@
 """
 from scipy.cluster.hierarchy import linkage, dendrogram, fcluster
 from scipy.spatial.distance import squareform
-from mpctools.extensions import npext
+from mpctools.extensions import npext, utils
+from sklearn import metrics as skmetrics
 import numpy as np
 
 
+def class_accuracy(y_true, y_pred, labels=None, normalize=True):
+    """
+    Computes per-class, one-v-rest accuracy
+
+    :param y_true: True labels (N)
+    :param y_pred: Predicted labels (N)
+    :param labels: If not None, specifies labels to consider: otherwise any label that appears in
+                   y_true or y_pred is considered
+    :param normalize: If True, normalise relative to all samples: else report number of samples.
+    :return: Accuracy-score per-class
+    """
+    # Define Labels
+    labels = utils.default(labels, np.union1d(np.unique(y_pred), np.unique(y_true)))
+
+    # compute per-class accuracy
+    accuracy = np.empty(len(labels))
+    for i, lbl in enumerate(labels):
+        accuracy[i] = skmetrics.accuracy_score(y_true == lbl, y_pred == lbl, normalize=normalize)
+    return accuracy
+
+
 def hierarchical_log_loss(y_true, y_prob, mapping, eps=1e-15):
     """
     Compute the Log-Loss, when y_true contains over-arching labels which are not predictable in