def _check_environment_trains(env, config): # Create controller and begin training. with tempfile.TemporaryDirectory() as dir: run_id = "id" save_freq = 99999 seed = 1337 trainer_config = yaml.safe_load(config) env_manager = SimpleEnvManager(env, FloatPropertiesChannel()) trainer_factory = TrainerFactory( trainer_config=trainer_config, summaries_dir=dir, run_id=run_id, model_path=dir, keep_checkpoints=1, train_model=True, load_model=False, seed=seed, meta_curriculum=None, multi_gpu=False, ) tc = TrainerController( trainer_factory=trainer_factory, summaries_dir=dir, model_path=dir, run_id=run_id, meta_curriculum=None, train=True, training_seed=seed, sampler_manager=SamplerManager(None), resampling_interval=None, save_freq=save_freq, ) # Begin training tc.start_learning(env_manager) print(tc._get_measure_vals()) for brain_name, mean_reward in tc._get_measure_vals().items(): assert not math.isnan(mean_reward) assert mean_reward > 0.99
def run_training(run_seed: int, options: RunOptions) -> None: """ Launches training session. :param options: parsed command line arguments :param run_seed: Random seed used for training. :param run_options: Command line arguments for training. """ with hierarchical_timer("run_training.setup"): model_path = f"./models/{options.run_id}" maybe_init_path = ( f"./models/{options.initialize_from}" if options.initialize_from else None ) summaries_dir = "./summaries" port = options.base_port # Configure CSV, Tensorboard Writers and StatsReporter # We assume reward and episode length are needed in the CSV. csv_writer = CSVWriter( summaries_dir, required_fields=[ "Environment/Cumulative Reward", "Environment/Episode Length", ], ) handle_existing_directories( model_path, summaries_dir, options.resume, options.force, maybe_init_path ) tb_writer = TensorboardWriter(summaries_dir, clear_past_data=not options.resume) gauge_write = GaugeWriter() console_writer = ConsoleWriter() StatsReporter.add_writer(tb_writer) StatsReporter.add_writer(csv_writer) StatsReporter.add_writer(gauge_write) StatsReporter.add_writer(console_writer) if options.env_path is None: port = UnityEnvironment.DEFAULT_EDITOR_PORT env_factory = create_environment_factory( options.env_path, options.no_graphics, run_seed, port, options.env_args ) engine_config = EngineConfig( width=options.width, height=options.height, quality_level=options.quality_level, time_scale=options.time_scale, target_frame_rate=options.target_frame_rate, capture_frame_rate=options.capture_frame_rate, ) env_manager = SubprocessEnvManager(env_factory, engine_config, options.num_envs) maybe_meta_curriculum = try_create_meta_curriculum( options.curriculum_config, env_manager, options.lesson ) sampler_manager, resampling_interval = create_sampler_manager( options.sampler_config, run_seed ) trainer_factory = TrainerFactory( options.trainer_config, summaries_dir, options.run_id, model_path, options.keep_checkpoints, not options.inference, options.resume, run_seed, maybe_init_path, maybe_meta_curriculum, options.multi_gpu, ) # Create controller and begin training. tc = TrainerController( trainer_factory, model_path, summaries_dir, options.run_id, options.save_freq, maybe_meta_curriculum, not options.inference, run_seed, sampler_manager, resampling_interval, ) # Begin training try: tc.start_learning(env_manager) finally: env_manager.close() write_timing_tree(summaries_dir, options.run_id)
def run_training(run_seed: int, options: RunOptions) -> None: """ Launches training session. :param options: parsed command line arguments :param run_seed: Random seed used for training. :param run_options: Command line arguments for training. """ # Recognize and use docker volume if one is passed as an argument if not options.docker_target_name: model_path = f"./models/{options.run_id}" summaries_dir = "./summaries" else: model_path = f"/{options.docker_target_name}/models/{options.run_id}" summaries_dir = f"/{options.docker_target_name}/summaries" port = options.base_port # Configure CSV, Tensorboard Writers and StatsReporter # We assume reward and episode length are needed in the CSV. csv_writer = CSVWriter( summaries_dir, required_fields=[ "Environment/Cumulative Reward", "Environment/Episode Length" ], ) tb_writer = TensorboardWriter(summaries_dir) StatsReporter.add_writer(tb_writer) StatsReporter.add_writer(csv_writer) if options.env_path is None: port = UnityEnvironment.DEFAULT_EDITOR_PORT env_factory = create_environment_factory( options.env_path, options.docker_target_name, options.no_graphics, run_seed, port, options.env_args, ) engine_config = EngineConfig( options.width, options.height, options.quality_level, options.time_scale, options.target_frame_rate, ) env_manager = SubprocessEnvManager(env_factory, engine_config, options.num_envs) maybe_meta_curriculum = try_create_meta_curriculum( options.curriculum_config, env_manager, options.lesson) sampler_manager, resampling_interval = create_sampler_manager( options.sampler_config, run_seed) trainer_factory = TrainerFactory( options.trainer_config, summaries_dir, options.run_id, model_path, options.keep_checkpoints, options.train_model, options.load_model, run_seed, maybe_meta_curriculum, options.multi_gpu, ) # Create controller and begin training. tc = TrainerController( trainer_factory, model_path, summaries_dir, options.run_id, options.save_freq, maybe_meta_curriculum, options.train_model, run_seed, sampler_manager, resampling_interval, ) # Begin training try: tc.start_learning(env_manager) finally: env_manager.close()
def run_training(run_seed: int, options: RunOptions) -> None: """ Launches training session. :param options: parsed command line arguments :param run_seed: Random seed used for training. :param run_options: Command line arguments for training. """ options.checkpoint_settings.run_id = "test8" with hierarchical_timer("run_training.setup"): checkpoint_settings = options.checkpoint_settings env_settings = options.env_settings engine_settings = options.engine_settings base_path = "results" write_path = os.path.join(base_path, checkpoint_settings.run_id) maybe_init_path = (os.path.join(base_path, checkpoint_settings.initialize_from) if checkpoint_settings.initialize_from else None) run_logs_dir = os.path.join(write_path, "run_logs") port: Optional[int] = env_settings.base_port # Check if directory exists handle_existing_directories( write_path, checkpoint_settings.resume, checkpoint_settings.force, maybe_init_path, ) # Make run logs directory os.makedirs(run_logs_dir, exist_ok=True) # Load any needed states if checkpoint_settings.resume: GlobalTrainingStatus.load_state( os.path.join(run_logs_dir, "training_status.json")) # Configure CSV, Tensorboard Writers and StatsReporter # We assume reward and episode length are needed in the CSV. csv_writer = CSVWriter( write_path, required_fields=[ "Environment/Cumulative Reward", "Environment/Episode Length", ], ) tb_writer = TensorboardWriter( write_path, clear_past_data=not checkpoint_settings.resume) gauge_write = GaugeWriter() console_writer = ConsoleWriter() StatsReporter.add_writer(tb_writer) StatsReporter.add_writer(csv_writer) StatsReporter.add_writer(gauge_write) StatsReporter.add_writer(console_writer) engine_config = EngineConfig( width=engine_settings.width, height=engine_settings.height, quality_level=engine_settings.quality_level, time_scale=engine_settings.time_scale, target_frame_rate=engine_settings.target_frame_rate, capture_frame_rate=engine_settings.capture_frame_rate, ) if env_settings.env_path is None: port = None # Begin training env_settings.env_path = "C:/Users/Sebastian/Desktop/RLUnity/Training/mFindTarget_new/RLProject.exe" env_factory = create_environment_factory( env_settings.env_path, engine_settings.no_graphics, run_seed, port, env_settings.env_args, os.path.abspath( run_logs_dir), # Unity environment requires absolute path ) env_manager = SubprocessEnvManager(env_factory, engine_config, env_settings.num_envs) maybe_meta_curriculum = try_create_meta_curriculum( options.curriculum, env_manager, restore=checkpoint_settings.resume) sampler_manager, resampling_interval = create_sampler_manager( options.parameter_randomization, run_seed) max_steps = options.behaviors['Brain'].max_steps options.behaviors['Brain'].max_steps = 10 trainer_factory = TrainerFactory(options, write_path, not checkpoint_settings.inference, checkpoint_settings.resume, run_seed, maybe_init_path, maybe_meta_curriculum, False, total_steps=0) trainer_factory.trainer_config[ 'Brain'].hyperparameters.learning_rate_schedule = ScheduleType.CONSTANT # Create controller and begin training. tc = TrainerController( trainer_factory, write_path, checkpoint_settings.run_id, maybe_meta_curriculum, not checkpoint_settings.inference, run_seed, sampler_manager, resampling_interval, ) try: # Get inital weights tc.init_weights(env_manager) inital_weights = deepcopy(tc.weights) finally: env_manager.close() write_run_options(write_path, options) write_timing_tree(run_logs_dir) write_training_status(run_logs_dir) options.behaviors['Brain'].max_steps = max_steps step = 0 counter = 0 max_meta_updates = 200 while counter < max_meta_updates: sample = np.random.random_sample() if (sample > 1): print("Performing Meta-learning on Carry Object stage") env_settings.env_path = "C:/Users/Sebastian/Desktop/RLUnity/Training/mCarryObject_new/RLProject.exe" else: print("Performing Meta-learning on Find Target stage") env_settings.env_path = "C:/Users/Sebastian/Desktop/RLUnity/Training/mFindTarget_new/RLProject.exe" env_factory = create_environment_factory( env_settings.env_path, engine_settings.no_graphics, run_seed, port, env_settings.env_args, os.path.abspath( run_logs_dir), # Unity environment requires absolute path ) env_manager = SubprocessEnvManager(env_factory, engine_config, env_settings.num_envs) maybe_meta_curriculum = try_create_meta_curriculum( options.curriculum, env_manager, restore=checkpoint_settings.resume) sampler_manager, resampling_interval = create_sampler_manager( options.parameter_randomization, run_seed) trainer_factory = TrainerFactory(options, write_path, not checkpoint_settings.inference, checkpoint_settings.resume, run_seed, maybe_init_path, maybe_meta_curriculum, False, total_steps=step) trainer_factory.trainer_config[ 'Brain'].hyperparameters.learning_rate_schedule = ScheduleType.CONSTANT trainer_factory.trainer_config[ 'Brain'].hyperparameters.learning_rate = 0.0005 * ( 1 - counter / max_meta_updates) trainer_factory.trainer_config[ 'Brain'].hyperparameters.beta = 0.005 * ( 1 - counter / max_meta_updates) trainer_factory.trainer_config[ 'Brain'].hyperparameters.epsilon = 0.2 * ( 1 - counter / max_meta_updates) print("Current lr: {}\nCurrent beta: {}\nCurrent epsilon: {}".format( trainer_factory.trainer_config['Brain'].hyperparameters. learning_rate, trainer_factory.trainer_config['Brain'].hyperparameters.beta, trainer_factory.trainer_config['Brain'].hyperparameters.epsilon)) # Create controller and begin training. tc = TrainerController( trainer_factory, write_path, checkpoint_settings.run_id, maybe_meta_curriculum, not checkpoint_settings.inference, run_seed, sampler_manager, resampling_interval, ) try: # Get inital weights print("Start learning at step: " + str(step) + " meta_step: " + str(counter)) print("Inital weights: " + str(inital_weights[8])) weights_after_train = tc.start_learning(env_manager, inital_weights) print(tc.trainers['Brain'].optimizer) # weights_after_train = tc.weights # print("Trained weights: " + str(weights_after_train[8])) step += options.behaviors['Brain'].max_steps print("meta step:" + str(step)) # print(weights_after_train) # equal = [] # for i, weight in enumerate(tc.weights): # equal.append(np.array_equal(inital_weights[i], weights_after_train[i])) # print(all(equal)) finally: print(len(weights_after_train), len(inital_weights)) for i, weight in enumerate(weights_after_train): inital_weights[i] = weights_after_train[i] env_manager.close() write_run_options(write_path, options) write_timing_tree(run_logs_dir) write_training_status(run_logs_dir) counter += 1
def run_training(run_seed: int, options: RunOptions) -> None: """ Launches training session. :param options: parsed command line arguments :param run_seed: Random seed used for training. :param run_options: Command line arguments for training. """ with hierarchical_timer("run_training.setup"): checkpoint_settings = options.checkpoint_settings env_settings = options.env_settings engine_settings = options.engine_settings base_path = "results" write_path = os.path.join(base_path, checkpoint_settings.run_id) maybe_init_path = ( os.path.join(base_path, checkpoint_settings.initialize_from) if checkpoint_settings.initialize_from is not None else None ) run_logs_dir = os.path.join(write_path, "run_logs") port: Optional[int] = env_settings.base_port # Check if directory exists handle_existing_directories( write_path, checkpoint_settings.resume, checkpoint_settings.force, maybe_init_path, ) # Make run logs directory os.makedirs(run_logs_dir, exist_ok=True) # Load any needed states if checkpoint_settings.resume: GlobalTrainingStatus.load_state( os.path.join(run_logs_dir, "training_status.json") ) # Configure Tensorboard Writers and StatsReporter tb_writer = TensorboardWriter( write_path, clear_past_data=not checkpoint_settings.resume ) gauge_write = GaugeWriter() console_writer = ConsoleWriter() StatsReporter.add_writer(tb_writer) StatsReporter.add_writer(gauge_write) StatsReporter.add_writer(console_writer) if env_settings.env_path is None: port = None env_factory = create_environment_factory( env_settings.env_path, engine_settings.no_graphics, run_seed, port, env_settings.env_args, os.path.abspath(run_logs_dir), # Unity environment requires absolute path ) engine_config = EngineConfig( width=engine_settings.width, height=engine_settings.height, quality_level=engine_settings.quality_level, time_scale=engine_settings.time_scale, target_frame_rate=engine_settings.target_frame_rate, capture_frame_rate=engine_settings.capture_frame_rate, ) env_manager = SubprocessEnvManager( env_factory, engine_config, env_settings.num_envs ) env_parameter_manager = EnvironmentParameterManager( options.environment_parameters, run_seed, restore=checkpoint_settings.resume ) trainer_factory = TrainerFactory( trainer_config=options.behaviors, output_path=write_path, train_model=not checkpoint_settings.inference, load_model=checkpoint_settings.resume, seed=run_seed, param_manager=env_parameter_manager, init_path=maybe_init_path, multi_gpu=False, force_torch="torch" in DetectDefault.non_default_args, ) # Create controller and begin training. tc = TrainerController( trainer_factory, write_path, checkpoint_settings.run_id, env_parameter_manager, not checkpoint_settings.inference, run_seed, ) # Begin training try: tc.start_learning(env_manager) finally: env_manager.close() write_run_options(write_path, options) write_timing_tree(run_logs_dir) write_training_status(run_logs_dir)
def run_training(run_seed: int, options: RunOptions) -> None: """ Launches training session. :param options: parsed command line arguments :param run_seed: Random seed used for training. :param run_options: Command line arguments for training. """ with hierarchical_timer("run_training.setup"): checkpoint_settings = options.checkpoint_settings env_settings = options.env_settings engine_settings = options.engine_settings base_path = "results" write_path = os.path.join(base_path, checkpoint_settings.run_id) maybe_init_path = ( os.path.join(base_path, checkpoint_settings.initialize_from) if checkpoint_settings.initialize_from else None ) run_logs_dir = os.path.join(write_path, "run_logs") port: Optional[int] = env_settings.base_port # Check if directory exists handle_existing_directories( write_path, checkpoint_settings.resume, checkpoint_settings.force, maybe_init_path, ) # Make run logs directory os.makedirs(run_logs_dir, exist_ok=True) # Load any needed states if checkpoint_settings.resume: GlobalTrainingStatus.load_state( os.path.join(run_logs_dir, "training_status.json") ) # Configure CSV, Tensorboard Writers and StatsReporter # We assume reward and episode length are needed in the CSV. csv_writer = CSVWriter( write_path, required_fields=[ "Environment/Cumulative Reward", "Environment/Episode Length", ], ) tb_writer = TensorboardWriter( write_path, clear_past_data=not checkpoint_settings.resume ) gauge_write = GaugeWriter() console_writer = ConsoleWriter() StatsReporter.add_writer(tb_writer) StatsReporter.add_writer(csv_writer) StatsReporter.add_writer(gauge_write) StatsReporter.add_writer(console_writer) if env_settings.env_path is None: port = None env_factory = create_environment_factory( env_settings.env_path, engine_settings.no_graphics, run_seed, port, env_settings.env_args, os.path.abspath(run_logs_dir), # Unity environment requires absolute path ) engine_config = EngineConfig( width=engine_settings.width, height=engine_settings.height, quality_level=engine_settings.quality_level, time_scale=engine_settings.time_scale, target_frame_rate=engine_settings.target_frame_rate, capture_frame_rate=engine_settings.capture_frame_rate, ) env_manager = SubprocessEnvManager( env_factory, engine_config, env_settings.num_envs ) maybe_meta_curriculum = try_create_meta_curriculum( options.curriculum, env_manager, restore=checkpoint_settings.resume ) maybe_add_samplers(options.parameter_randomization, env_manager, run_seed) trainer_factory = TrainerFactory( options.behaviors, write_path, not checkpoint_settings.inference, checkpoint_settings.resume, run_seed, maybe_init_path, maybe_meta_curriculum, False, ) # Create controller and begin training. tc = TrainerController( trainer_factory, write_path, checkpoint_settings.run_id, maybe_meta_curriculum, not checkpoint_settings.inference, run_seed, ) # Begin training try: tc.start_learning(env_manager) finally: env_manager.close() write_run_options(write_path, options) write_timing_tree(run_logs_dir) write_training_status(run_logs_dir)
def run_training( sub_id: int, run_seed: int, options: CommandLineOptions, process_queue: Queue ) -> None: """ Launches training session. :param process_queue: Queue used to send signal back to main. :param sub_id: Unique id for training session. :param options: parsed command line arguments :param run_seed: Random seed used for training. :param run_options: Command line arguments for training. """ # Docker Parameters trainer_config_path = options.trainer_config_path curriculum_folder = options.curriculum_folder # Recognize and use docker volume if one is passed as an argument if not options.docker_target_name: model_path = "./models/{run_id}-{sub_id}".format( run_id=options.run_id, sub_id=sub_id ) summaries_dir = "./summaries" else: trainer_config_path = "/{docker_target_name}/{trainer_config_path}".format( docker_target_name=options.docker_target_name, trainer_config_path=trainer_config_path, ) if curriculum_folder is not None: curriculum_folder = "/{docker_target_name}/{curriculum_folder}".format( docker_target_name=options.docker_target_name, curriculum_folder=curriculum_folder, ) model_path = "/{docker_target_name}/models/{run_id}-{sub_id}".format( docker_target_name=options.docker_target_name, run_id=options.run_id, sub_id=sub_id, ) summaries_dir = "/{docker_target_name}/summaries".format( docker_target_name=options.docker_target_name ) trainer_config = load_config(trainer_config_path) port = options.base_port + (sub_id * options.num_envs) if options.env_path is None: port = 5004 # This is the in Editor Training Port env_factory = create_environment_factory( options.env_path, options.docker_target_name, options.no_graphics, run_seed, port, options.env_args, ) engine_config = EngineConfig( options.width, options.height, options.quality_level, options.time_scale, options.target_frame_rate, ) env_manager = SubprocessEnvManager(env_factory, engine_config, options.num_envs) maybe_meta_curriculum = try_create_meta_curriculum( curriculum_folder, env_manager, options.lesson ) sampler_manager, resampling_interval = create_sampler_manager( options.sampler_file_path, run_seed ) trainer_factory = TrainerFactory( trainer_config, summaries_dir, options.run_id, model_path, options.keep_checkpoints, options.train_model, options.load_model, run_seed, maybe_meta_curriculum, options.multi_gpu, ) # Create controller and begin training. tc = TrainerController( trainer_factory, model_path, summaries_dir, options.run_id + "-" + str(sub_id), options.save_freq, maybe_meta_curriculum, options.train_model, run_seed, sampler_manager, resampling_interval, ) # Signal that environment has been launched. process_queue.put(True) # Begin training try: tc.start_learning(env_manager) finally: env_manager.close()
def run_training_aai(run_seed: int, options: RunOptionsAAI) -> None: """ Launches training session. :param run_seed: Random seed used for training. :param options: training parameters """ with hierarchical_timer("run_training.setup"): # Recognize and use docker volume if one is passed as an argument # if not options.docker_target_name: model_path = f"./models/{options.run_id}" summaries_dir = "./summaries" # else: # model_path = f"/{options.docker_target_name}/models/{options.run_id}" # summaries_dir = f"/{options.docker_target_name}/summaries" port = options.base_port # Configure CSV, Tensorboard Writers and StatsReporter # We assume reward and episode length are needed in the CSV. csv_writer = CSVWriter( summaries_dir, required_fields=[ "Environment/Cumulative Reward", "Environment/Episode Length", ], ) tb_writer = TensorboardWriter(summaries_dir) gauge_write = GaugeWriter() StatsReporter.add_writer(tb_writer) StatsReporter.add_writer(csv_writer) StatsReporter.add_writer(gauge_write) if options.env_path is None: port = AnimalAIEnvironment.DEFAULT_EDITOR_PORT env_factory = create_environment_factory_aai( options.env_path, # options.docker_target_name, run_seed, port, options.n_arenas_per_env, options.arena_config, options.resolution, ) if options.train_model: engine_config = EngineConfig( options.width, options.height, AnimalAIEnvironment.QUALITY_LEVEL.train, AnimalAIEnvironment.TIMESCALE.train, AnimalAIEnvironment.TARGET_FRAME_RATE.train, ) else: engine_config = EngineConfig( AnimalAIEnvironment.WINDOW_WIDTH.play, AnimalAIEnvironment.WINDOW_HEIGHT.play, AnimalAIEnvironment.QUALITY_LEVEL.play, AnimalAIEnvironment.TIMESCALE.play, AnimalAIEnvironment.TARGET_FRAME_RATE.play, ) env_manager = SubprocessEnvManagerAAI(env_factory, engine_config, options.num_envs) maybe_meta_curriculum = try_create_meta_curriculum( options.curriculum_config, env_manager, options.lesson) trainer_factory = TrainerFactory( options.trainer_config, summaries_dir, options.run_id, model_path, options.keep_checkpoints, options.train_model, options.load_model, run_seed, maybe_meta_curriculum, # options.multi_gpu, ) # Create controller and begin training. tc = TrainerControllerAAI( trainer_factory, model_path, summaries_dir, options.run_id, options.save_freq, maybe_meta_curriculum, options.train_model, run_seed, ) # Begin training try: tc.start_learning(env_manager) finally: env_manager.close() write_timing_tree(summaries_dir, options.run_id)
def run_training(run_seed: int, options: RunOptions) -> None: """ Launches training session. :param options: parsed command line arguments :param run_seed: Random seed used for training. :param run_options: Command line arguments for training. """ model_path = f"./models/{options.run_id}" summaries_dir = "./summaries" port = options.base_port # Configure CSV, Tensorboard Writers and StatsReporter # We assume reward and episode length are needed in the CSV. csv_writer = CSVWriter( summaries_dir, required_fields=[ "Environment/Cumulative Reward", "Environment/Episode Length" ], ) tb_writer = TensorboardWriter(summaries_dir) StatsReporter.add_writer(tb_writer) StatsReporter.add_writer(csv_writer) if options.env_path is None: port = 5004 # This is the in Editor Training Port env_factory = create_environment_factory(options.env_path, options.no_graphics, run_seed, port, options.env_args, options.env_id, options.n_steps) env_manager = SubprocessEnvManager(env_factory=env_factory, n_env=options.num_envs) maybe_meta_curriculum = try_create_meta_curriculum( options.curriculum_config, env_manager, options.lesson) sampler_manager, resampling_interval = create_sampler_manager( options.sampler_config, run_seed) trainer_factory = TrainerFactory( options.trainer_config, summaries_dir, options.run_id, model_path, options.keep_checkpoints, options.train_model, options.load_model, run_seed, maybe_meta_curriculum, options.multi_gpu, ) # Create controller and begin training. tc = TrainerController(trainer_factory=trainer_factory, model_path=model_path, summaries_dir=summaries_dir, run_id=options.run_id, save_freq=options.save_freq, meta_curriculum=maybe_meta_curriculum, train=options.train_model, training_seed=run_seed, sampler_manager=sampler_manager, resampling_interval=resampling_interval, n_steps=options.n_steps) # Begin training try: tc.start_learning(env_manager) finally: env_manager.close()