Test moving the EPSILON & similar to free some RAM

Signed-off-by: Frederic Boisnard <frederic.boisnard@irt-saintexupery.com>

Author: fredericboisnard

xplique/attributions/grad_cam_pp.py

@@ -40,10 +40,6 @@ class GradCAMPP(GradCAM):
         If a string is provided it will look for the layer name.
     """
 
-    # Avoid zero division during procedure. (the value is not important, as if the denominator is
-    # zero, then the nominator will also be zero).
-    EPSILON = tf.constant(1e-4)
-
     @staticmethod
     @tf.function
     def _compute_weights(feature_maps_gradients: tf.Tensor,
@@ -71,7 +67,11 @@ def _compute_weights(feature_maps_gradients: tf.Tensor,
         nominator = feature_maps_gradients_square
         denominator = 2.0 * feature_maps_gradients_square + \
                       feature_maps_gradients_cube * feature_map_avg
-        denominator += tf.cast(denominator == 0, tf.float32) * GradCAMPP.EPSILON
+
+        # Avoid zero division during procedure. (the value is not important, as if the denominator is
+        # zero, then the nominator will also be zero).
+        EPSILON = tf.constant(1e-4)
+        denominator += tf.cast(denominator == 0, tf.float32) * EPSILON
 
         feature_map_alphas = nominator / denominator * tf.nn.relu(feature_maps_gradients)
         weights = tf.reduce_mean(feature_map_alphas, axis=(1, 2))

xplique/attributions/rise.py

@@ -42,10 +42,6 @@ class Rise(BlackBoxExplainer):
         Value used as when applying masks.
     """
 
-    # Avoid zero division during procedure. (the value is not important, as if the denominator is
-    # zero, then the nominator will also be zero).
-    EPSILON = tf.constant(1e-4)
-
     def __init__(self,
                  model: Callable,
                  batch_size: Optional[int] = 32,
@@ -113,7 +109,10 @@ def explain(self,
                 rise_nominator += tf.reduce_sum(predictions * masks_upsampled, 0)
                 rise_denominator += tf.reduce_sum(masks_upsampled, 0)
 
-            rise_map = rise_nominator / (rise_denominator + Rise.EPSILON)
+            # Avoid zero division during procedure. (the value is not important, as if the denominator is
+            # zero, then the nominator will also be zero).
+            EPSILON = tf.constant(1e-4)
+            rise_map = rise_nominator / (rise_denominator + EPSILON)
             rise_map = rise_map[tf.newaxis]
 
             rise_maps = rise_map if rise_maps is None else tf.concat([rise_maps, rise_map], axis=0)

xplique/commons/operators.py

@@ -7,7 +7,7 @@
 import tensorflow as tf
 
 from ..types import Callable, Optional
-from ..utils_functions.object_detection import _box_iou, _format_objects, _EPSILON
+from ..utils_functions.object_detection import _box_iou, _format_objects
 
 
 @tf.function
@@ -197,6 +197,7 @@ def batch_loop(args):
         class_refs = tf.repeat(tf.expand_dims(class_refs, axis=1), repeats=size, axis=1)
 
         # (nb_box_ref, nb_box_pred)
+        _EPSILON = tf.constant(1e-4)
         classification_score = tf.reduce_sum(class_refs * classification, axis=-1) \
                 / (tf.norm(classification, axis=-1) * tf.norm(class_refs, axis=-1)+ _EPSILON)
 

xplique/features_visualizations/preconditioning.py

@@ -16,13 +16,6 @@
 IMAGENET_SPECTRUM_URL = "https://storage.googleapis.com/serrelab/loupe/"\
                         "spectrums/imagenet_decorrelated.npy"
 
-imagenet_color_correlation = tf.cast(
-      [[0.56282854, 0.58447580, 0.58447580],
-       [0.19482528, 0.00000000,-0.19482528],
-       [0.04329450,-0.10823626, 0.06494176]], tf.float32
-)
-
-
 def recorrelate_colors(images: tf.Tensor) -> tf.Tensor:
     """
     Map uncorrelated colors to 'normal colors' by using empirical color
@@ -39,6 +32,11 @@ def recorrelate_colors(images: tf.Tensor) -> tf.Tensor:
     images
         Images recorrelated.
     """
+    imagenet_color_correlation = tf.cast(
+      [[0.56282854, 0.58447580, 0.58447580],
+       [0.19482528, 0.00000000,-0.19482528],
+       [0.04329450,-0.10823626, 0.06494176]], tf.float32
+    )
     images_flat = tf.reshape(images, [-1, 3])
     images_flat = tf.matmul(images_flat, imagenet_color_correlation)
     return tf.reshape(images_flat, tf.shape(images))

xplique/utils_functions/object_detection.py

@@ -4,9 +4,6 @@
 from typing import Tuple
 import tensorflow as tf
 
-_EPSILON = tf.constant(1e-4)
-
-
 def _box_iou(boxes_a: tf.Tensor, boxes_b: tf.Tensor) -> tf.Tensor:
     """
     Compute the intersection between two batched bounding boxes.
@@ -40,6 +37,7 @@ def _box_iou(boxes_a: tf.Tensor, boxes_b: tf.Tensor) -> tf.Tensor:
 
     union_area = a_area + b_area - intersection_area
 
+    _EPSILON = tf.constant(1e-4)
     iou_score = intersection_area / (union_area + _EPSILON)
 
     return iou_score