Optimize spectral_normalization implementation.

PiperOrigin-RevId: 722757020

Author: tensorflower-gardener

tf_keras/layers/normalization/spectral_normalization.py

@@ -95,7 +95,7 @@ def build(self, input_shape):
 
     def call(self, inputs, training=False):
         if training:
-            self.normalize_weights()
+            self._update_weights()
 
         output = self.layer(inputs)
         return output
@@ -105,35 +105,42 @@ def compute_output_shape(self, input_shape):
             self.layer.compute_output_shape(input_shape).as_list()
         )
 
+    def _update_weights(self):
+        weights = self.kernel
+        vector_u = self.vector_u
+
+        kernel_weights, vector_u = tf.cond(
+            tf.reduce_all(tf.equal(weights, 0)),
+            lambda: (weights, vector_u),
+            lambda: self.normalize_weights(),
+        )
+        self.kernel.assign(kernel_weights)
+        self.vector_u.assign(vector_u)
+
     def normalize_weights(self):
         """Generate spectral normalized weights.
 
         This method will update the value of `self.kernel` with the
         spectral normalized value, so that the layer is ready for `call()`.
         """
-
-        weights = tf.reshape(self.kernel, [-1, self.kernel_shape[-1]])
+        # Initialize vector_v to hint the compiler it always exist.
         vector_u = self.vector_u
-
-        # check for zeroes weights
-        if not tf.reduce_all(tf.equal(weights, 0.0)):
-            for _ in range(self.power_iterations):
-                vector_v = tf.math.l2_normalize(
-                    tf.matmul(vector_u, weights, transpose_b=True)
-                )
-                vector_u = tf.math.l2_normalize(tf.matmul(vector_v, weights))
-            vector_u = tf.stop_gradient(vector_u)
-            vector_v = tf.stop_gradient(vector_v)
-            sigma = tf.matmul(
-                tf.matmul(vector_v, weights), vector_u, transpose_b=True
-            )
-            self.vector_u.assign(tf.cast(vector_u, self.vector_u.dtype))
-            self.kernel.assign(
-                tf.cast(
-                    tf.reshape(self.kernel / sigma, self.kernel_shape),
-                    self.kernel.dtype,
-                )
+        vector_v = self.vector_u
+        weights = tf.reshape(self.kernel, [-1, self.kernel_shape[-1]])
+        for _ in range(self.power_iterations):
+            vector_v = tf.math.l2_normalize(
+                tf.matmul(vector_u, weights, transpose_b=True)
             )
+            vector_u = tf.math.l2_normalize(tf.matmul(vector_v, weights))
+        vector_u = tf.stop_gradient(vector_u)
+        vector_v = tf.stop_gradient(vector_v)
+        sigma = tf.matmul(
+            tf.matmul(vector_v, weights),
+            vector_u,
+            transpose_b=True,
+        )
+        weights_normalized = tf.reshape(weights / sigma, self.kernel_shape)
+        return weights_normalized, vector_u
 
     def get_config(self):
         config = {"power_iterations": self.power_iterations}