I know the parameters chosen for training a RL model depend heavily on the model itself as well as the problem. Nevertheless, I am trying to train a bunch of these agents on different environments, and would like to do a preliminar training in order to check which Environment-agent combination seems more fitting or promising for my problem to then develop this one more precisely. The thing is that I do not count with advanced resources (I have a dedicated CPU, but no GPU) and the executions take very long, so I must select these parameters very wisely in order to run a training process significant enough but also fast enough to have the results at least in two weeks or three. I'd like to train 3 agents on 4 different environments, and with my current code and parameters, a leaning episode takes 2 hours, while an evaluation episode (and I'm tying to run 5 of them per each learning one) takes about 8 hours, so I think my code must have some kind of flaw that causes it to last so much on the evaluation part...

As you will see in the code, I constantly store model versions. This is because I am using a selenium webdriver script in my agent's step method, and for some reason every now and then (like every hour or so) Chrome DevTools disconnect, to which the only solution I have found has been to store constant copies of the model and the rewards in order to be able to continue from the last checkpoint every time I rerun the training (to be able to not start from the beginning over and over again)

Therefore, I'd like to know:

  • Which would be a good balance of episodes-iterations for learning and evaluation? I am currently doing this:

    • The environment returns "done" when a certain condition is fulfilled or after a maximum of 1024 iterations
    • Training episodes: 10
    • Learning iterations per training episode: 1024
    • Evaluation episodes per training step: 5
    • Evaluation iterations per evaluation episode: 1024
    • Iteration steps per to store evaluation checkpoints: 103
  • Is there anything I can change in my current code to make all these executions much more efficient? Because I think I must be doing something wrong for it to last for so long (specially on evaluation)

Any other suggestion to improve the training process or the code?

def learn(self, total_timesteps, load_if_previous = False, training_step = 0, force_more_trainig = False, checkpoint_interval=100, callback=None, log_interval=1, tb_log_name='AgentV1',
            reset_num_timesteps=True, progress_bar=True):
        print(f"\t\t\t\t\t\t LEARNING {datetime.datetime.now()}")
        current_step = 0
        final_model_name = f"{self.model_path}_{training_step}_final"

        # Check if the model already exists
        if os.path.exists(f"{final_model_name}.zip") and load_if_previous and not force_more_trainig:
            # Skip learning step o
            print(f"\tModel for training step {training_step} already exists ({final_model_name}.zip). Loading the model.")
            while current_step < total_timesteps:
                remaining_steps = total_timesteps - current_step
                steps_to_train = min(checkpoint_interval, remaining_steps)
                print(f"\t ------ Training from {current_step} to {current_step + steps_to_train}  ------ ")
                # Check if the model already exists
                iternmediate_model_file = f"{self.model_path}_{training_step}_{current_step}.zip"
                print(f"\t\t Finding {iternmediate_model_file}...")
                if os.path.exists(iternmediate_model_file) and load_if_previous:
                    print(f"\t\t Model for training step and batch {training_step} already exists ({iternmediate_model_file}.zip). Loading the model")
                    print(f"\t\t Learning batch #{current_step} of {total_timesteps}")
                    # Set the model's environment before training
                    self.model.set_env(self.gym_env )
                    self.model.learn(total_timesteps=steps_to_train, callback=callback, log_interval=log_interval,
                                    tb_log_name=tb_log_name, reset_num_timesteps=reset_num_timesteps, progress_bar=progress_bar)

                    # Save the model after every checkpoint interval
                    if current_step % checkpoint_interval == 0:
                current_step += steps_to_train

            # Remove non-final intermediate files
            for intermediate_file in glob.glob(f"{self.model_path}_{training_step}_*.zip"):
                if intermediate_file != f"{final_model_name}.zip" and intermediate_file != f"{final_model_name}":
                    print(f"\t\t - Deleted intermediate file: {intermediate_file}")

            # Save the final model
        print(f"\t--------------------- ENDOF Learning {datetime.datetime.now()} ---------------------\n")
        return self.load(f"{final_model_name}.zip")

    def evaluate(self, checkpoint_interval = 100, deterministic=False, training_step = 0):
        The model will be evaluated for self.n_eval_episodes times until "is done"
        n_batch_iterations = math.ceil(self.gym_env.max_steps / checkpoint_interval)

        print(f"\t\t\t\t\t\t EVALUATION {datetime.datetime.now()}")
        print(f"\t\t - Evaluation episodes: {self.n_eval_episodes}")
        print(f"\t\t - Iterations for the step: {self.gym_env.max_steps}")
        print(f"\t\t - Checkpoints for the step: {checkpoint_interval}")

        # ------------------------------------- Checkpoint Loadng ------------------------------------- 
        # LOAD last execution's episode, batch and agent information
        checkpoint_files = glob.glob(f"{self.model_path}_evaluation_Teps_{training_step}_episode_*_batch_*.zip")
        if len(checkpoint_files) > 0:
            last_checkpoint_file = checkpoint_files[-1]
            episode_str, batch_str = last_checkpoint_file.split(f"_evaluation_Teps_{training_step}_episode_")[1].split("_batch_")
            last_episode = int(episode_str)
            # Obtain the batch from the string (if errors, batch = 0)
            error = True
            attempts = 5
            loops = 0
            last_batch = 0
            while error or loops < attempts:  
                loops += 1
                    batch_str = batch_str[:-4]
                    last_batch = int(batch_str)
                    error = False
                except ValueError as e:

            print(f"\t\t  Found checkpoint file: {last_checkpoint_file}")
            print(f"\t\t   - Last episode: {last_episode}")
            print(f"\t\t   - Last batch: {last_batch}")
            # Load the file's information into "self"
            print("\t\t - No checkpoint file found. Starting evaluation from scratch.")
            last_episode = 0
            last_batch = 0
        # Create a temporary environment to evaluate the policy
        eval_env = self.gym_env
        reward_file_path = f"{self.monitor_gym_log_folder}/Rewards_{self.model_name}_Evaluation_Teps_{self.gym_env.current_episode}.csv"

        # Initialize lists to store rewards
        episode_rewards = []

        # LOAD episode rewards from previous executions
        episode_rewards_file = f"{self.model_path}_evaluation_rewards_for_training_step_{training_step}.npy"
        if os.path.exists(episode_rewards_file):
            print(f"\t\t  Rewards loaded from: {episode_rewards_file}")
            episode_rewards = np.load(episode_rewards_file)
        # ------------------------------------- ENDOF Loading ------------------------------------- 
        batch = last_batch
        iteration_count = 0
        step = 0
        for episode in range(last_episode, self.n_eval_episodes):
            print(f"\n\t\t Executing evaluation EPISODE #{episode + 1}/{self.n_eval_episodes + 1}")
            obs = eval_env.reset()[0]

            done = False
            episode_reward = episode_rewards[episode] if episode < len(episode_rewards) else 0.0

            print(f"\n\t\t Executing evaluation BATCH #{batch}/{n_batch_iterations}")
            while not done:
                step += 1
                # Get the action from the model
                action, _ = self.model.predict(obs, deterministic=deterministic)

                # Perform the action in the environment
                obs, reward, done, truncated, info = eval_env.step(action, episode = episode + 1, step = step, reward_file_path = reward_file_path)
                # obs, reward, done, info = eval_env.step(action, episode = episode, step = iteration_count, reward_file_path = reward_file_path)

                # Accumulate the reward
                episode_reward = reward
                # Increment the iteration count in order to save (or not) a new checkpoint
                iteration_count += 1
                if iteration_count > checkpoint_interval:
                    # ------------------------------------- Checkpoint Storage ------------------------------------- 
                    # After each batch, store the information about batch, episode and the agent
                    print(f"\t\t STORING BATCH: info Batch checkpoint stored at {self.model_path}_evaluation_Teps_{training_step}_episode_{episode}_batch_{batch}.zip")

                    # Save the rewards after that batch
                    print(f"\t\t STORING REWARDS")
                    np.save(episode_rewards_file, episode_rewards)
                    # ------------------------------------- ENDOF Checkpoint Storage ------------------------------------- 
                    batch += 1
                    iteration_count = 0
                    print(f"\n\t\t Executing evaluation BATCH #{batch}/{n_batch_iterations}")

            done = False
            step = 0

            if episode < len(episode_rewards):
                episode_rewards[episode] = episode_reward
                if isinstance(episode_rewards, np.ndarray):
                    episode_rewards = np.append(episode_rewards, episode_reward)

            episode_reward = 0
            # ------------------------------------------------------------ Checkpoint Deletion ------------------------------------------------------------ 
            # After each episode, delete the intermidiate file where episode is lower than current episode (the same for rewards)
            # Delete intermediate files for episodes and batches lower than the current episode
            for episode_file in glob.glob(f"{self.model_path}_evaluation_Teps_{training_step}_episode_*.zip"):
                episode_num = int(episode_file.split(f"_evaluation_Teps_{training_step}_episode_")[1].split("_batch")[0])
                if episode_num < episode:
                    print(f"\t\t - Deleted intermediate episode file: {episode_file}")

            for batch_file in glob.glob(f"{self.model_path}_evaluation_Teps_{training_step}_episode_{episode}_batch_*.zip"):
                batch_num = int(batch_file.split("_batch_")[1][:-4])
                if batch_num < batch:
                    print(f"\t\t - Deleted intermediate batch file: {batch_file}")
            # ------------------------------------------------------------ ENDOF Checkpoint Deletion ------------------------------------------------------
            if episode >= self.n_eval_episodes:
        # Calculate mean and standard deviation of rewards
        mean_reward = np.mean(episode_rewards)
        std_reward = np.std(episode_rewards)

        print(f"\t\t Mean reward = {mean_reward:.2f} +/- {std_reward}")
        print(f"\t--------------------- ENDOF Evaluation {datetime.datetime.now()} ---------------------\n")
        return mean_reward

    def choose_action(self, obs, deterministic = False):
        return self.model.predict(obs, deterministic=deterministic)
    def train_agent(self, n_steps = 2048, total_timesteps = 50000, callback=None, log_interval=1, tb_log_name='AgentV1', reset_num_timesteps=True, progress_bar=False, checkpoint = True, force_more_trainig = False):
        # Set the model's environment before training
        self.model.set_env(self.gym_env )

        base_path =  f"{self.model_path}_{total_timesteps}_{n_steps}_"
        extension = ".zip"
        plot_files = []
        print(f"           Training agent {self.model_name}")

        print(f" Training info:", flush=True)
        print(f"    - Total training steps: {n_steps}", flush=True)
        print(f"    - Iterations per step: {total_timesteps}", flush=True)
        print(f"    - Model stored as: {base_path}_*{extension}")
        mean_reward = 0

        if self.trained and not force_more_trainig:
            print("A trained model has already been found, training will be skipped.")
            print(f"If you want to keep training the trained model {self.model_name}, rerun the training function by adding force_more_trainig = True")
            # train the agent
            # Loading the last training info to continue from the same point in case the execution ends abruptly
            if checkpoint:
                best_mean_reward, best_model_path = -float('inf'), None
                model_paths =  glob.glob(f"{base_path}_*{extension}")
                # Check if there is a saved model
                if model_paths:
                    best_model_path = max(model_paths, key=lambda path: int(path.split('_')[-1].split('.')[0]))
                    print("Loading saved model...")
                    last_step = int(best_model_path.split("_")[-1].split('.')[0])
                    print(f"Continuing training from step #{last_step}")
                    last_step = 0

            best_mean_reward, best_model_path = -float('inf'), None
            for i in range(1, n_steps+1):
                print(f"\n_______________________________________________________________Training EPISODE #{i}/{n_steps}_______________________________________________________________", flush=True)

                print(f"\tIterations for the step: {total_timesteps}", flush=True)
                self.gym_env.set_test_case(data_wrapper.insert("Test_cases", {
                    "Execution_uid": self.gym_env.execution['uid'],
                    "Initial_time": utils.getDateTime(),
                    "Error_count": 0,   
                    "Unique_error_count": 0,
                initial_time = utils.getDateTime()
                self.learn(total_timesteps=total_timesteps, load_if_previous = True, force_more_trainig = force_more_trainig, training_step = i)
                # evaluate the agent and log the results

                # If a model for the next step already exists, skip evaluation of the current one
                if len(glob.glob(f"{self.model_path}_{i + 1}_*.zip")) > 0:
                    print(f"\t\t Evaluation for the current Episode has already been completed")
                    mean_reward = self.evaluate(checkpoint_interval = math.ceil(self.gym_env.max_steps / 10) , deterministic=False, training_step = i)
                    if mean_reward > best_mean_reward:
                        best_mean_reward = mean_reward
                        best_model_path = f"{base_path}_{i}{extension}"
                        print(f"\tNew best model saved to {best_model_path}")
                        print(f"\tNew best model saved to {best_model_path}", flush=True)   

                data_wrapper.update("Test_cases", {"uid": self.gym_env.test_case['uid'] }, { "$set":  self.gym_env.test_case })
            data_wrapper.update("Test_cases", {"uid": self.gym_env.test_case['uid'] }, { "$set": {"End_time":  utils.getDateTime(), 'Execution_uid': self.gym_env.execution['uid'], 'Initial_time': initial_time }})

                # Generate plots after each episode
                if self.monitor:
                        print("Generating plot to send...")
                        # Call plot_results and redirect the output to a StringIO object
                        results_plotter.plot_results([ self.monitor_gym_log_folder ], total_timesteps + 1, results_plotter.X_TIMESTEPS, f"{self.model_name} training for episode {i}")

                        # Rename the monitor file in self.monitor_gym_log_folder/monitor.csv to self.monitor_gym_log_folder/monitor_{i}.csv by making a copy
                            monitor_file_path = os.path.join(self.monitor_gym_log_folder, "monitor.csv")
                            renamed_monitor_file_path = os.path.join(self.monitor_gym_log_folder, f"monitor_{i}.csv")
                            shutil.copy(monitor_file_path, renamed_monitor_file_path)
                        except Exception as e:
                            print("An error occurred when trying to rename the monitor file")

                        # Capture the plot
                        fig = plt.gcf()

                        # Save the figure to a file
                        plot_file = f"{self.monitor_gym_log_folder}/training_results_{self.model_name}_episode_{i}_{ str(int(time.time())) }.png"

                        # Save the plot to a file
                        print(f"Plot file saved at {plot_file}")

                        plt.close(fig)  # Close the plot to free up memory
                    except Exception as e:
                        print("Error when generating the plot with the built in function. ERROR:")
                        print("Plot information: ") 
                        print(f"    - Source Monitor file name: {self.monitor_gym_log_folder}/monitor.csv")
                        print(f"    - Renamed Monitor file name: {self.monitor_gym_log_folder}/monitor_{i}.csv")
                        print(f"    - Target file name: {self.monitor_gym_log_folder}/training_results_{self.model_name}_episode_{i}_{ str(int(time.time())) }.png")
                        plot_file = ''
                print(f"\n_________________________________________ Training step #{i}/{n_steps} ended at: {datetime.datetime.now()} _________________________________________\n\n\n", flush=True)
            # SAFE THE FINAL MODEL
        # Modify the execution to save the model name and mean reward
        self.gym_env.execution["Mean_reward"] = mean_reward
        self.gym_env.execution["End_time"] = utils.getDateTime()
        data_wrapper.update("Executions", {"uid": self.gym_env.execution['uid'] }, { "$set": self.gym_env.execution })

        # Send email and plot
        utils.sendTrainingEndEmail( self.gym_env.execution['uid'], tb_log_name, mean_reward, plot_files, self.gym_env.test_case)

  • 1
    $\begingroup$ There are too many questions together. Please revise it, or split into multiple threads. $\endgroup$
    – lpounng
    Jul 27 at 8:38
  • $\begingroup$ Please edit the question to limit it to a specific problem with enough detail to identify an adequate answer. $\endgroup$
    – Community Bot
    Jul 29 at 13:39


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.