Models should not compute loss when encrypted backprop enabled.

google · Oct 30, 2024 · 44ece50 · 44ece50
1 parent da280db
commit 44ece50
Show file tree

Hide file tree

Showing 3 changed files with 15 additions and 17 deletions.
diff --git a/tf_shell_ml/dpsgd_sequential_model.py b/tf_shell_ml/dpsgd_sequential_model.py
@@ -247,18 +247,18 @@ def rebalance(x, s):
         for metric in self.metrics:
             if metric.name == "loss":
                 if self.disable_encryption:
-                    metric.update_state(0)  # Loss is not known when encrypted.
-                else:
                     loss = self.loss_fn(y, y_pred)
                     metric.update_state(loss)
+                else:
+                    # Loss is unknown when encrypted.
+                    metric.update_state(0.0)
             else:
                 if self.disable_encryption:
                     metric.update_state(y, y_pred)
                 else:
+                    # Other metrics are uknown when encrypted.
                     zeros = tf.broadcast_to(0, tf.shape(y_pred))
-                    metric.update_state(
-                        zeros, zeros
-                    )  # No other metrics when encrypted.
+                    metric.update_state(zeros, zeros)
 
         metric_results = {m.name: m.result() for m in self.metrics}
 

diff --git a/tf_shell_ml/loss.py b/tf_shell_ml/loss.py
@@ -16,7 +16,7 @@
 from tensorflow.nn import softmax
 from tensorflow.nn import sigmoid
 from tensorflow.math import log
-import tf_shell
+import tensorflow as tf
 
 
 class CategoricalCrossentropy:
@@ -27,10 +27,9 @@ def __init__(self, from_logits=False, lazy_normalization=True):
     def __call__(self, y_true, y_pred):
         if self.from_logits:
             y_pred = softmax(y_pred)
-        batch_size = y_true.shape.as_list()[0]
-        batch_size_inv = 1 / batch_size
+        batch_size = tf.shape(y_true)[0]
         out = -y_true * log(y_pred)
-        cce = tf_shell.reduce_sum(out, axis=0) * batch_size_inv
+        cce = tf.reduce_sum(out, axis=0) / tf.cast(batch_size, out.dtype)
         return cce
 
     def grad(self, y_true, y_pred):
@@ -61,10 +60,9 @@ def __call__(self, y_true, y_pred):
         if self.from_logits:
             y_pred = sigmoid(y_pred)
 
-        batch_size = y_true.shape.as_list()[0]
-        batch_size_inv = 1 / batch_size
+        batch_size = tf.shape(y_true)[0]
         out = -(y_true * log(y_pred) + (1 - y_true) * log(1 - y_pred))
-        bce = tf_shell.reduce_sum(out, axis=0) * batch_size_inv
+        bce = tf.reduce_sum(out, axis=0) / tf.cast(batch_size, out.dtype)
         return bce
 
     def grad(self, y_true, y_pred):

diff --git a/tf_shell_ml/postscale_sequential_model.py b/tf_shell_ml/postscale_sequential_model.py
@@ -249,18 +249,18 @@ def rebalance(x, s):
         for metric in self.metrics:
             if metric.name == "loss":
                 if self.disable_encryption:
-                    metric.update_state(0)  # Loss is not known when encrypted.
-                else:
                     loss = self.loss_fn(y, y_pred)
                     metric.update_state(loss)
+                else:
+                    # Loss is unknown when encrypted.
+                    metric.update_state(0.0)
             else:
                 if self.disable_encryption:
                     metric.update_state(y, y_pred)
                 else:
+                    # Other metrics are uknown when encrypted.
                     zeros = tf.broadcast_to(0, tf.shape(y_pred))
-                    metric.update_state(
-                        zeros, zeros
-                    )  # No other metrics when encrypted.
+                    metric.update_state(zeros, zeros)
 
         metric_results = {m.name: m.result() for m in self.metrics}