592c161032bfd697ef6f2f334c5cd007062bec4b,rllib/models/tf/recurrent_net.py,LSTMWrapper,forward,#LSTMWrapper#Any#Any#Any#,176

Before Change

        // Concat. prev-action/reward if required.
        if self.model_config["lstm_use_prev_action_reward"]:
            wrapped_out = tf.concat(
                [
                    wrapped_out,
                    tf.reshape(
                        tf.cast(input_dict[SampleBatch.PREV_ACTIONS],
                                tf.float32), [-1, self.action_dim]),
                    tf.reshape(
                        tf.cast(input_dict[SampleBatch.PREV_REWARDS],
                                tf.float32), [-1, 1]),
                ],
                axis=1)

        // Then through our LSTM.
        input_dict["obs_flat"] = wrapped_out
        return super().forward(input_dict, state, seq_lens)

After Change

        prev_a_r = []
        if self.model_config["lstm_use_prev_action"]:
            prev_a = input_dict[SampleBatch.PREV_ACTIONS]
            if isinstance(self.action_space, (Discrete, MultiDiscrete)):
                prev_a = one_hot(prev_a, self.action_space)
            prev_a_r.append(
                tf.reshape(tf.cast(prev_a, tf.float32), [-1, self.action_dim]))
        if self.model_config["lstm_use_prev_reward"]:
            prev_a_r.append(

Instances