diff --git a/rllib/models/tf/fcnet.py b/rllib/models/tf/fcnet.py
index b016b5a97..36f0f3819 100644
--- a/rllib/models/tf/fcnet.py
+++ b/rllib/models/tf/fcnet.py
@@ -33,7 +33,7 @@ class FullyConnectedNetwork(TFModelV2):
 
         # We are using obs_flat, so take the flattened shape as input.
         inputs = tf.keras.layers.Input(
-            shape=(np.product(obs_space.shape), ), name="observations")
+            shape=(int(np.product(obs_space.shape)), ), name="observations")
         # Last hidden layer output (before logits outputs).
         last_layer = inputs
         # The action distribution outputs.
@@ -75,7 +75,7 @@ class FullyConnectedNetwork(TFModelV2):
             # Adjust num_outputs to be the number of nodes in the last layer.
             else:
                 self.num_outputs = (
-                    [np.product(obs_space.shape)] + hiddens[-1:])[-1]
+                    [int(np.product(obs_space.shape))] + hiddens[-1:])[-1]
 
         # Concat the log std vars to the end of the state-dependent means.
         if free_log_std and logits_out is not None:
diff --git a/rllib/models/torch/fcnet.py b/rllib/models/torch/fcnet.py
index 109b282b5..747fc221a 100644
--- a/rllib/models/torch/fcnet.py
+++ b/rllib/models/torch/fcnet.py
@@ -78,7 +78,7 @@ class FullyConnectedNetwork(TorchModelV2, nn.Module):
                     activation_fn=None)
             else:
                 self.num_outputs = (
-                    [np.product(obs_space.shape)] + hiddens[-1:])[-1]
+                    [int(np.product(obs_space.shape))] + hiddens[-1:])[-1]
 
         # Layer to add the log std vars to the state-dependent means.
         if self.free_log_std and self._logits: