fix(simclr): use two tapes for representations and embeddings

sslforslr/models/simclr/SimCLR.py

@@ -74,7 +74,7 @@ def train_step(self, data):
         X_1, X_2, _ = data
         # X shape: (B, H, W, C) = (B, 40, 200, 1)
 
-        with tf.GradientTape() as tape:
+        with tf.GradientTape() as tape1, tf.GradientTape() as tape2:
             Z_1 = self.encoder(X_1, training=True)
             Z_2 = self.encoder(X_2, training=True)
             representations_loss, representations_accuracy = self.representations_loss(
@@ -91,15 +91,17 @@ def train_step(self, data):
                 )
 
         # Apply representations loss
-        params = self.encoder.trainable_weights
-        grads = tape.gradient(representations_loss, params)
-        self.optimizer.apply_gradients(zip(grads, params))
+        if self.representations_loss_vic or self.representations_loss_nce:
+            params = self.encoder.trainable_weights
+            grads = tape1.gradient(representations_loss, params)
+            self.optimizer.apply_gradients(zip(grads, params))
 
         # Aplly embeddings loss
-        params = self.encoder.trainable_weights
-        params += self.mlp.trainable_weights
-        grads = tape.gradient(embeddings_loss, params)
-        self.optimizer.apply_gradients(zip(grads, params))
+        if self.embeddings_loss_vic or self.embeddings_loss_nce:
+            params = self.encoder.trainable_weights
+            params += self.mlp.trainable_weights
+            grads = tape2.gradient(embeddings_loss, params)
+            self.optimizer.apply_gradients(zip(grads, params))
 
         return {
             'representations_loss': representations_loss,