[RLlib] Add 2 Transformer learning test cases on StatelessCartPole (PPO and IMPALA). (#8624)

rllib/examples/attention_net.py

@@ -50,7 +50,7 @@ if __name__ == "__main__":
         "model": {
             "custom_model": GTrXLNet,
             "max_seq_len": 50,
-            "custom_options": {
+            "custom_model_config": {
                 "num_transformer_units": 1,
                 "attn_dim": 64,
                 "num_heads": 2,

rllib/examples/custom_loss.py

@@ -53,7 +53,7 @@ if __name__ == "__main__":
         "num_workers": 0,
         "model": {
             "custom_model": "custom_loss",
-            "custom_options": {
+            "custom_model_config": {
                 "input_files": args.input_files,
             },
         },

rllib/examples/mobilenet_v2_with_lstm.py

@@ -36,7 +36,7 @@ if __name__ == "__main__":
         "model": {
             "custom_model": "my_model",
             # Extra config passed to the custom model's c'tor as kwargs.
-            "custom_options": {
+            "custom_model_config": {
                 "cnn_shape": cnn_shape_torch if args.torch else cnn_shape,
             },
             "max_seq_len": 20,

rllib/examples/models/custom_loss_model.py

@@ -38,7 +38,8 @@ class CustomLossModel(TFModelV2):
     @override(ModelV2)
     def custom_loss(self, policy_loss, loss_inputs):
         # Create a new input reader per worker.
-        reader = JsonReader(self.model_config["custom_options"]["input_files"])
+        reader = JsonReader(
+            self.model_config["custom_model_config"]["input_files"])
         input_ops = reader.tf_input_ops()
 
         # Define a secondary loss by building a graph copy with weight sharing.
@@ -80,7 +81,7 @@ class DeprecatedCustomLossModelV1(Model):
 
     def custom_loss(self, policy_loss, loss_inputs):
         # create a new input reader per worker
-        reader = JsonReader(self.options["custom_options"]["input_files"])
+        reader = JsonReader(self.options["custom_model_config"]["input_files"])
         input_ops = reader.tf_input_ops()
 
         # define a secondary loss by building a graph copy with weight sharing