fix #1207

open_spiel/python/algorithms/rcfr_test.py

@@ -37,7 +37,7 @@ def _new_model():
   return rcfr.DeepRcfrModel(
       _GAME,
       num_hidden_layers=1,
-      num_hidden_units=13,
+      num_hidden_units=26,
       num_hidden_factors=1,
       use_skip_connections=True)
 
@@ -476,12 +476,16 @@ def test_rcfr_functions(self):
       data = data.batch(12)
       data = data.repeat(num_epochs)
 
-      optimizer = tf.keras.optimizers.Adam(lr=0.005, amsgrad=True)
+      optimizer = tf.keras.optimizers.Adam(learning_rate=0.005, amsgrad=True)
 
+      model = models[regret_player]
       for x, y in data:
-        optimizer.minimize(
-            lambda: tf.losses.huber_loss(y, models[regret_player](x)),  # pylint: disable=cell-var-from-loop
-            models[regret_player].trainable_variables)
+        with tf.GradientTape() as tape:
+          loss = tf.losses.huber_loss(y, model(x))
+        optimizer.apply_gradients(
+          zip(
+            tape.gradient(loss, model.trainable_variables),
+            model.trainable_variables))
 
       regret_player = reach_weights_player
 
@@ -504,12 +508,15 @@ def _train(model, data):
       data = data.batch(12)
       data = data.repeat(num_epochs)
 
-      optimizer = tf.keras.optimizers.Adam(lr=0.005, amsgrad=True)
+      optimizer = tf.keras.optimizers.Adam(learning_rate=0.005, amsgrad=True)
 
       for x, y in data:
-        optimizer.minimize(
-            lambda: tf.losses.huber_loss(y, model(x)),  # pylint: disable=cell-var-from-loop
-            model.trainable_variables)
+        with tf.GradientTape() as tape:
+          loss = tf.losses.huber_loss(y, model(x))
+        optimizer.apply_gradients(
+          zip(
+            tape.gradient(loss, model.trainable_variables),
+            model.trainable_variables))
 
     average_policy = patient.average_policy()
     self.assertGreater(pyspiel.nash_conv(_GAME, average_policy), 0.91)
@@ -565,12 +572,15 @@ def _train(model, data):
       data = data.batch(12)
       data = data.repeat(num_epochs)
 
-      optimizer = tf.keras.optimizers.Adam(lr=0.005, amsgrad=True)
+      optimizer = tf.keras.optimizers.Adam(learning_rate=0.005, amsgrad=True)
 
       for x, y in data:
-        optimizer.minimize(
-            lambda: tf.losses.huber_loss(y, model(x)),  # pylint: disable=cell-var-from-loop
-            model.trainable_variables)
+        with tf.GradientTape() as tape:
+          loss = tf.losses.huber_loss(y, model(x))
+        optimizer.apply_gradients(
+          zip(
+            tape.gradient(loss, model.trainable_variables),
+            model.trainable_variables))
 
     average_policy = patient.average_policy()
     self.assertGreater(pyspiel.nash_conv(_GAME, average_policy), 0.91)

open_spiel/python/examples/rcfr_example.py

@@ -87,14 +87,17 @@ def _train_fn(model, data):
     data = data.batch(FLAGS.batch_size)
     data = data.repeat(FLAGS.num_epochs)
 
-    optimizer = tf.keras.optimizers.Adam(lr=FLAGS.step_size, amsgrad=True)
+    optimizer = tf.keras.optimizers.Adam(learning_rate=FLAGS.step_size, amsgrad=True)
 
     @tf.function
     def _train():
       for x, y in data:
-        optimizer.minimize(
-            lambda: tf.losses.huber_loss(y, model(x), delta=0.01),  # pylint: disable=cell-var-from-loop
-            model.trainable_variables)
+        with tf.GradientTape() as tape:
+          loss = tf.losses.huber_loss(y, model(x), delta=0.01)
+        optimizer.apply_gradients(
+          zip(
+            tape.gradient(loss, model.trainable_variables),
+            model.trainable_variables))
 
     _train()