keras_extensions.py

@@ -0,0 +1,105 @@
+import numpy as np
+from keras import backend as K
+
+
+def to_categorical(y, nb_classes=None):
+    """ Convert class vector (integers from 0 to nb_classes)
+    to binary class matrix, for use with categorical_crossentropy
+    :param y: label image with values between 0 and 255 (ex. shape: (h, w))
+    :param nb_classes: number of classes
+    :return: "binary" label image with as many channels as number of labels (ex. shape: (h, w, #lab))
+    """
+    y = np.asarray(y, dtype='int32')
+    if not nb_classes:
+        nb_classes = np.max(y) + 1
+
+    yy = y.flatten()
+    Y = np.zeros((len(yy), nb_classes))
+    for i in range(len(yy)):
+        Y[i, yy[i]] = 1.
+    Y = np.reshape(Y, list(y.shape) + [nb_classes])
+    return Y
+
+
+def softmax_3d(class_dim=-1):
+    """ 3D extension of softmax, class is last dim"""
+    def activation(x):
+        ndim = K.ndim(x)
+        if ndim == 2:
+            return K.softmax(x)
+        elif ndim > 2:
+            e = K.exp(x - K.max(x, axis=class_dim, keepdims=True))
+            s = K.sum(e, axis=class_dim, keepdims=True)
+            return e / s
+        else:
+            raise Exception('Cannot apply softmax to a tensor that is not 2D or '
+                            '3D. Here, ndim=' + str(ndim))
+    return activation
+
+
+def categorical_crossentropy_3d_w(alpha, class_dim=-1):
+    """ Weighted 3D extension CCE, class is last dim"""
+    def loss(y_true, y_pred):
+        y_pred /= K.sum(y_pred, axis=class_dim, keepdims=True)
+        y_pred = K.clip(y_pred, K.epsilon(), 1.0 - K.epsilon())
+        cce = y_true * K.log(y_pred)
+        cce = K.sum(cce, axis=class_dim)
+        cce *= 1 + alpha * K.clip(K.cast(K.argmax(y_true, axis=class_dim), K.floatx()), 0, 1)
+        cce = -K.sum(K.sum(cce, axis=-1), axis=-1)
+        return cce
+    return loss
+
+
+def categorical_crossentropy_3d(class_dim=-1):
+    """2D categorical crossentropy loss
+    """
+    def loss(y_true, y_pred):
+        # scale preds so that the class probas of each sample sum to 1
+        y_pred /= K.sum(y_pred, axis=class_dim, keepdims=True)
+        # avoid numerical instability with _EPSILON clipping
+        y_pred = K.clip(y_pred, K.epsilon(), 1.0 - K.epsilon())
+        cce = -K.sum(K.sum(K.sum(y_true * K.log(y_pred), axis=class_dim), axis=-1), axis=-1)
+        return cce
+    return loss
+
+
+def softmax_2d(class_dim=-1):
+    """2D softmax activation
+    """
+    def activation(x):
+        ndim = K.ndim(x)
+        if ndim == 2:
+            return K.softmax(x)
+        elif ndim > 2:
+            e = K.exp(x - K.max(x, axis=class_dim, keepdims=True))
+            s = K.sum(e, axis=class_dim, keepdims=True)
+            return e / s
+        else:
+            raise Exception('Cannot apply softmax to a tensor that is not 2D or '
+                            '3D. Here, ndim=' + str(ndim))
+    return activation
+
+
+def categorical_crossentropy_2d(class_dim=-1):
+    """2D categorical crossentropy loss
+    """
+    def loss(y_true, y_pred):
+        # scale preds so that the class probas of each sample sum to 1
+        y_pred /= K.sum(y_pred, axis=class_dim, keepdims=True)
+        # avoid numerical instability with _EPSILON clipping
+        y_pred = K.clip(y_pred, K.epsilon(), 1.0 - K.epsilon())
+        cce = -K.sum(K.sum(y_true * K.log(y_pred), axis=class_dim), axis=-1)
+        return cce
+    return loss
+
+
+def categorical_crossentropy_2d_w(alpha, class_dim=-1):
+    def loss(y_true, y_pred):
+        y_pred /= K.sum(y_pred, axis=class_dim, keepdims=True)
+        y_pred = K.clip(y_pred, K.epsilon(), 1.0 - K.epsilon())
+        cce = y_true * K.log(y_pred)
+        cce = K.sum(cce, axis=class_dim)
+        cce *= 1 + alpha * K.clip(K.argmax(y_true, axis=class_dim), 0, 1)
+        cce = -K.sum(cce, axis=-1)
+        return cce
+    return loss