remove tf.stop_gradient

Author: GreenWizard2015

NN/RestorationModel/CRepeatedRestorator.py

@@ -28,7 +28,6 @@ def call(self, latents, pos, T, V, residual, training=None):
 
   def reverse(self, latents, pos, reverseArgs, training, value, residual, index):
     for i in range(self._N):
-      if tf.is_tensor(value): value = tf.stop_gradient(value)
       value = self._restorator.reverse(
         latents=self._withID(latents, i, training),
         pos=pos, reverseArgs=reverseArgs,
@@ -50,11 +49,11 @@ def train_step(self, x0, latents, positions, params, xT=None):
 
     for i in range(1, self._N):
       trainStep = self._restorator.train_step(
-        x0=tf.stop_gradient(x0),
-        xT=tf.stop_gradient(xT),
-        latents=self._withID(tf.stop_gradient(latents), i, training=True),
-        positions=tf.stop_gradient(positions),
-        params={k: tf.stop_gradient(v) if tf.is_tensor(v) else v for k, v in params.items()}
+        x0=x0,
+        xT=xT,
+        latents=self._withID(latents, i, training=True),
+        positions=positions,
+        params=params
       )
       loss += trainStep['loss']
       xT = tf.stop_gradient(trainStep['value'])

NN/RestorationModel/CSequentialRestorator.py

@@ -17,7 +17,6 @@ def call(self, latents, pos, T, V, residual, training=None):
 
   def reverse(self, latents, pos, reverseArgs, training, value, residual, index):
     for restorator in self._restorators:
-      if tf.is_tensor(value): value = tf.stop_gradient(value)
       value = restorator.reverse(
         latents=latents, pos=pos, reverseArgs=reverseArgs,
         residual=residual,
@@ -37,11 +36,11 @@ def train_step(self, x0, latents, positions, params, xT=None):
     # the rest of the restorators are trained sequentially
     for restorator in self._restorators:
       trainStep = restorator.train_step(
-        x0=tf.stop_gradient(x0),
-        xT=tf.stop_gradient(xT),
-        latents=tf.stop_gradient(latents),
-        positions=tf.stop_gradient(positions),
-        params={k: tf.stop_gradient(v) if isinstance(v, tf.Tensor) else v for k, v in params.items()}
+        x0=x0,
+        xT=xT,
+        latents=latents,
+        positions=positions,
+        params=params
       )
       loss += trainStep['loss']
       xT = tf.stop_gradient(trainStep['value'])