class Agent: """An agent. It is able: - to choose an action given an observation, - to analyze the feedback (i.e. reward and done state) of its action.""" def __init__(self, obs_space, action_space, model_dir, device=None, argmax=False, num_envs=1, use_rim=False): obs_space, self.preprocess_obss = utils.get_obss_preprocessor( obs_space) self.acmodel = ACModel(obs_space, action_space, use_rim=use_rim) self.device = device self.argmax = argmax self.num_envs = num_envs if self.acmodel.recurrent: self.memories = torch.zeros(self.num_envs, self.acmodel.memory_size).to(device) self.acmodel.load_state_dict(utils.get_model_state(model_dir)) self.acmodel.to(self.device) self.acmodel.eval() if hasattr(self.preprocess_obss, "vocab"): self.preprocess_obss.vocab.load_vocab(utils.get_vocab(model_dir)) def get_actions(self, obss): preprocessed_obss = self.preprocess_obss(obss, device=self.device) with torch.no_grad(): if self.acmodel.recurrent: dist, _, self.memories = self.acmodel(preprocessed_obss, self.memories) else: dist, _ = self.acmodel(preprocessed_obss) if self.argmax: actions = dist.probs.max(1, keepdim=True)[1] else: actions = dist.sample() return actions.cpu().numpy() def get_action(self, obs): return self.get_actions([obs])[0] def analyze_feedbacks(self, rewards, dones): if self.acmodel.recurrent: masks = 1 - torch.tensor(dones, dtype=torch.float).to( self.device).unsqueeze(1) self.memories *= masks def analyze_feedback(self, reward, done): return self.analyze_feedbacks([reward], [done])
class Agent: """An agent. It is able: - to choose an action given an observation, - to analyze the feedback (i.e. reward and done state) of its action.""" def __init__(self, obs_space, action_space, model_dir, device=None, argmax=False, num_envs=1): obs_space, self.preprocess_obss = utils.get_obss_preprocessor(obs_space) self.acmodel = ACModel(obs_space, action_space) self.device = device self.argmax = argmax self.num_envs = num_envs self.acmodel.load_state_dict(utils.get_model_state(model_dir)) self.acmodel.to(self.device) self.acmodel.eval() def get_actions(self, obss): preprocessed_obss = self.preprocess_obss(obss, device=self.device) with torch.no_grad(): dist, _ = self.acmodel(preprocessed_obss) if self.argmax: actions = dist.probs.max(1, keepdim=True)[1] else: actions = dist.sample() return actions.cpu().numpy() def get_action(self, obs): return self.get_actions([obs])[0] def analyze_feedbacks(self, rewards, dones): pass def analyze_feedback(self, reward, done): return self.analyze_feedbacks([reward], [done])
txt_logger.info("Training status loaded\n") except OSError: status = {"num_frames": 0, "update": 0} # Load observations preprocessor obs_space, preprocess_obs_goals = utils.get_obs_goals_preprocessor(envs[0].observation_space) if "vocab" in status: preprocess_obs_goals.vocab.load_vocab(status["vocab"]) txt_logger.info("observations preprocessor loaded") # Load model acmodel = ACModel(obs_space, envs[0].action_space, args.mem, args.text) if "model_state" in status: acmodel.load_state_dict(status["model_state"]) txt_logger.info("Model loaded\n") acmodel.to(device) txt_logger.info("{}\n".format(acmodel)) # Load algo if args.algo == "a2c": algo = a2c.A2CAlgo(envs, acmodel, device, args.frames_per_proc, args.discount, args.lr, args.gae_lambda, args.entropy_coef, args.value_loss_coef, args.max_grad_norm, args.recurrence, args.optim_alpha, args.optim_eps, preprocess_obs_goals) elif args.algo == "ppo": algo = ppo.PPOAlgo(envs, acmodel, device, args.frames_per_proc, args.discount, args.lr, args.gae_lambda, args.entropy_coef, args.value_loss_coef, args.max_grad_norm, args.recurrence, args.optim_eps, args.clip_eps, args.epochs, args.batch_size, preprocess_obs_goals) else:
def main(): # Parse arguments parser = argparse.ArgumentParser() ## General parameters parser.add_argument( "--algo", required=True, help="algorithm to use: a2c | ppo | ppo_intrinsic (REQUIRED)") parser.add_argument("--env", required=True, help="name of the environment to train on (REQUIRED)") parser.add_argument( "--model", default=None, help="name of the model (default: {ENV}_{ALGO}_{TIME})") parser.add_argument("--seed", type=int, default=1, help="random seed (default: 1)") parser.add_argument("--log-interval", type=int, default=1, help="number of updates between two logs (default: 1)") parser.add_argument( "--save-interval", type=int, default=10, help= "number of updates between two saves (default: 10, 0 means no saving)") parser.add_argument("--procs", type=int, default=16, help="number of processes (default: 16)") parser.add_argument("--frames", type=int, default=10**7, help="number of frames of training (default: 1e7)") ## Parameters for main algorithm parser.add_argument("--epochs", type=int, default=4, help="number of epochs for PPO (default: 4)") parser.add_argument("--batch-size", type=int, default=256, help="batch size for PPO (default: 256)") parser.add_argument( "--frames-per-proc", type=int, default=None, help= "number of frames per process before update (default: 5 for A2C and 128 for PPO)" ) parser.add_argument("--discount", type=float, default=0.99, help="discount factor (default: 0.99)") parser.add_argument("--lr", type=float, default=0.001, help="learning rate (default: 0.001)") parser.add_argument( "--gae-lambda", type=float, default=0.95, help="lambda coefficient in GAE formula (default: 0.95, 1 means no gae)" ) parser.add_argument("--entropy-coef", type=float, default=0.01, help="entropy term coefficient (default: 0.01)") parser.add_argument("--value-loss-coef", type=float, default=0.5, help="value loss term coefficient (default: 0.5)") parser.add_argument("--max-grad-norm", type=float, default=0.5, help="maximum norm of gradient (default: 0.5)") parser.add_argument( "--optim-eps", type=float, default=1e-8, help="Adam and RMSprop optimizer epsilon (default: 1e-8)") parser.add_argument("--optim-alpha", type=float, default=0.99, help="RMSprop optimizer alpha (default: 0.99)") parser.add_argument("--clip-eps", type=float, default=0.2, help="clipping epsilon for PPO (default: 0.2)") parser.add_argument( "--recurrence", type=int, default=1, help= "number of time-steps gradient is backpropagated (default: 1). If > 1, a LSTM is added to the model to have memory." ) parser.add_argument("--text", action="store_true", default=False, help="add a GRU to the model to handle text input") parser.add_argument("--visualize", default=False, help="show real time CNN layer weight changes") args = parser.parse_args() args.mem = args.recurrence > 1 # Set run dir date = datetime.datetime.now().strftime("%y-%m-%d-%H-%M-%S") default_model_name = f"{args.env}_{args.algo}_seed{args.seed}_{date}" model_name = args.model or default_model_name model_dir = utils.get_model_dir(model_name) # Load loggers and Tensorboard writer txt_logger = utils.get_txt_logger(model_dir) csv_file, csv_logger = utils.get_csv_logger(model_dir) tb_writer = tensorboardX.SummaryWriter(model_dir) # Log command and all script arguments txt_logger.info("{}\n".format(" ".join(sys.argv))) txt_logger.info("{}\n".format(args)) # Set seed for all randomness sources utils.seed(args.seed) # Set device device = torch.device("cuda" if torch.cuda.is_available() else "cpu") txt_logger.info(f"Device: {device}\n") # Load environments envs = [] for i in range(args.procs): envs.append(utils.make_env(args.env, args.seed + 10000 * i)) txt_logger.info("Environments loaded\n") # Load training status try: status = utils.get_status(model_dir) except OSError: status = {"num_frames": 0, "update": 0} txt_logger.info("Training status loaded\n") # Load observations preprocessor obs_space, preprocess_obss = utils.get_obss_preprocessor( envs[0].observation_space) if "vocab" in status: preprocess_obss.vocab.load_vocab(status["vocab"]) txt_logger.info("Observations preprocessor loaded") # Load model acmodel = ACModel(obs_space, envs[0].action_space, args.mem, args.text) if "model_state" in status: acmodel.load_state_dict(status["model_state"]) acmodel.to(device) txt_logger.info("Model loaded\n") txt_logger.info("{}\n".format(acmodel)) # Load algo if args.algo == "a2c": algo = torch_ac.A2CAlgo(envs, acmodel, device, args.frames_per_proc, args.discount, args.lr, args.gae_lambda, args.entropy_coef, args.value_loss_coef, args.max_grad_norm, args.recurrence, args.optim_alpha, args.optim_eps, preprocess_obss) elif args.algo == "ppo": algo = torch_ac.PPOAlgo(envs, acmodel, device, args.frames_per_proc, args.discount, args.lr, args.gae_lambda, args.entropy_coef, args.value_loss_coef, args.max_grad_norm, args.recurrence, args.optim_eps, args.clip_eps, args.epochs, args.batch_size, preprocess_obss) elif args.algo == "ppo_intrinsic": algo = torch_ac.PPOAlgoIntrinsic( envs, acmodel, device, args.frames_per_proc, args.discount, args.lr, args.gae_lambda, args.entropy_coef, args.value_loss_coef, args.max_grad_norm, args.recurrence, args.optim_eps, args.clip_eps, args.epochs, args.batch_size, preprocess_obss) elif args.algo == "a2c_intrinsic": algo = torch_ac.A2CAlgoIntrinsic( envs, acmodel, device, args.frames_per_proc, args.discount, args.lr, args.gae_lambda, args.entropy_coef, args.value_loss_coef, args.max_grad_norm, args.recurrence, args.optim_alpha, args.optim_eps, preprocess_obss) else: raise ValueError("Incorrect algorithm name: {}".format(args.algo)) if "optimizer_state" in status: algo.optimizer.load_state_dict(status["optimizer_state"]) txt_logger.info("Optimizer loaded\n") # Train model num_frames = status["num_frames"] update = status["update"] start_time = time.time() print_visual = args.visualize if print_visual: fig, axs = plt.subplots(1, 3) fig.suptitle('Convolution Layer Weights Normalized Difference') while num_frames < args.frames: # Store copies of s_t model params old_parameters = {} for name, param in acmodel.named_parameters(): old_parameters[name] = param.detach().numpy().copy() # Update model parameters update_start_time = time.time() exps, logs1 = algo.collect_experiences() logs2 = algo.update_parameters(exps) logs = {**logs1, **logs2} update_end_time = time.time() # Store copies of s_t+1 model params new_parameters = {} for name, param in acmodel.named_parameters(): new_parameters[name] = param.detach().numpy().copy() # Compute L2 Norm of model state differences # Print model weight change visualization for index in range(len(old_parameters.keys())): if index == 0 or index == 2 or index == 4: key = list(old_parameters.keys())[index] old_weights = old_parameters[key] new_weights = new_parameters[key] norm_diff = numpy.linalg.norm(new_weights - old_weights) diff_matrix = abs(new_weights - old_weights) diff_matrix[:, :, 0, 0] = normalize(diff_matrix[:, :, 0, 0], norm='max', axis=0) if print_visual: axs[int(index / 2)].imshow(diff_matrix[:, :, 0, 0], cmap='Greens', interpolation='nearest') # This allows the plots to update as the model trains if print_visual: plt.ion() plt.show() plt.pause(0.001) num_frames += logs["num_frames"] update += 1 # Print logs if update % args.log_interval == 0: fps = logs["num_frames"] / (update_end_time - update_start_time) duration = int(time.time() - start_time) return_per_episode = utils.synthesize(logs["return_per_episode"]) rreturn_per_episode = utils.synthesize( logs["reshaped_return_per_episode"]) num_frames_per_episode = utils.synthesize( logs["num_frames_per_episode"]) header = ["update", "frames", "FPS", "duration"] data = [update, num_frames, fps, duration] header += ["rreturn_" + key for key in rreturn_per_episode.keys()] data += rreturn_per_episode.values() header += [ "num_frames_" + key for key in num_frames_per_episode.keys() ] data += num_frames_per_episode.values() header += [ "entropy", "value", "policy_loss", "value_loss", "grad_norm" ] data += [ logs["entropy"], logs["value"], logs["policy_loss"], logs["value_loss"], logs["grad_norm"] ] txt_logger.info( "U {} | F {:06} | FPS {:04.0f} | D {} | rR:μσmM {:.2f} {:.2f} {:.2f} {:.2f} | F:μσmM {:.1f} {:.1f} {} {} | H {:.3f} | V {:.3f} | pL {:.3f} | vL {:.3f} | ∇ {:.3f}" .format(*data)) header += ["return_" + key for key in return_per_episode.keys()] data += return_per_episode.values() if status["num_frames"] == 0: csv_logger.writerow(header) csv_logger.writerow(data) csv_file.flush() for field, value in zip(header, data): tb_writer.add_scalar(field, value, num_frames) # Save status if args.save_interval > 0 and update % args.save_interval == 0: status = { "num_frames": num_frames, "update": update, "model_state": acmodel.state_dict(), "optimizer_state": algo.optimizer.state_dict() } if hasattr(preprocess_obss, "vocab"): status["vocab"] = preprocess_obss.vocab.vocab utils.save_status(status, model_dir) txt_logger.info("Status saved")
class Agent: def __init__(self, env, model_dir, model_type='PPO2', logger=None, argmax=False, use_memory=False, use_text=False, num_cpu=1, frames_per_proc=None, discount=0.99, lr=0.001, gae_lambda=0.95, entropy_coef=0.01, value_loss_coef=0.5, max_grad_norm=0.5, recurrence=1, optim_eps=1e-8, optim_alpha=None, clip_eps=0.2, epochs=4, batch_size=256): """ Initialize the Agent object. This primarily includes storing of the configuration parameters, but there is some other logic for correctly initializing the agent. :param env: the environment for training :param model_dir: the save directory (appended with the goal_id in initialization) :param model_type: the type of model {'PPO2', 'A2C'} :param logger: existing text logger :param argmax: if we use determinsitic or probabilistic action selection :param use_memory: if we are using an LSTM :param use_text: if we are using NLP to parse the goal :param num_cpu: the number of parallel instances for training :param frames_per_proc: max time_steps per process (versus constant) :param discount: the discount factor (gamma) :param lr: the learning rate :param gae_lambda: the generalized advantage estimator lambda parameter (training smoothing parameter) :param entropy_coef: relative weight for entropy loss :param value_loss_coef: relative weight for value function loss :param max_grad_norm: max scaling factor for the gradient :param recurrence: number of recurrent steps :param optim_eps: minimum value to prevent numerical instability :param optim_alpha: RMSprop decay parameter (A2C only) :param clip_eps: clipping parameter for the advantage and value function (PPO2 only) :param epochs: number of epochs in the parameter update (PPO2 only) :param batch_size: number of samples for the parameter update (PPO2 only) """ if hasattr( env, 'goal' ) and env.goal: # if the environment has a goal, set the model_dir to the goal folder self.model_dir = model_dir + env.goal.goalId + '/' else: # otherwise just use the model_dir as is self.model_dir = model_dir # store all of the input parameters self.model_type = model_type self.num_cpu = num_cpu self.frames_per_proc = frames_per_proc self.discount = discount self.lr = lr self.gae_lambda = gae_lambda self.entropy_coef = entropy_coef self.value_loss_coef = value_loss_coef self.max_grad_norm = max_grad_norm self.recurrence = recurrence self.optim_eps = optim_eps self.optim_alpha = optim_alpha self.clip_eps = clip_eps self.epochs = epochs self.batch_size = batch_size # use the existing logger and create two new ones self.txt_logger = logger self.csv_file, self.csv_logger = utils.get_csv_logger(self.model_dir) self.tb_writer = tensorboardX.SummaryWriter(self.model_dir) self.set_env( env ) # set the environment to with some additional checks and init of training_envs self.algo = None # we don't initialize the algorithm until we call init_training_algo() device = torch.device("cuda" if torch.cuda.is_available() else "cpu") self.txt_logger.info(f"Device: {device}\n") try: # if we have a saved model, load it self.status = utils.get_status(self.model_dir) except OSError: # otherwise initialize the status print('error loading saved model. initializing empty model...') self.status = {"num_frames": 0, "update": 0} if self.txt_logger: self.txt_logger.info("Training status loaded\n") if "vocab" in self.status: preprocess_obss.vocab.load_vocab(self.status["vocab"]) if self.txt_logger: self.txt_logger.info("Observations preprocessor loaded") # get the obs_space and the observation pre-processor # (for manipulating gym observations into a torch-friendly format) obs_space, self.preprocess_obss = utils.get_obss_preprocessor( self.env.observation_space) self.acmodel = ACModel(obs_space, self.env.action_space, use_memory=use_memory, use_text=use_text) self.device = device # store the device {'cpu', 'cuda:N'} self.argmax = argmax # if we are using greedy action selection # or are we using probabilistic action selection if self.acmodel.recurrent: # initialize the memories self.memories = torch.zeros(num_cpu, self.acmodel.memory_size, device=self.device) if "model_state" in self.status: # if we have a saved model ('model_state') in the status # load that into the initialized model self.acmodel.load_state_dict(self.status["model_state"]) self.acmodel.to( device) # make sure the model is located on the correct device self.txt_logger.info("Model loaded\n") self.txt_logger.info("{}\n".format(self.acmodel)) # some redundant code. uncomment if there are issues and delete after enough testing #if 'model_state' in self.status: # self.acmodel.load_state_dict(self.status['model_state']) #self.acmodel.to(self.device) self.acmodel.eval() if hasattr(self.preprocess_obss, "vocab"): self.preprocess_obss.vocab.load_vocab(utils.get_vocab(model_dir)) def init_training_algo(self, num_envs=None): """ Initialize the training algorithm. This primarily calls the object creation functions for the A2C or PPO2 and the optimizer, but this also spawns a number of parallel environments, based on the self.num_cpu or num_envs input (if provided). Note, the spawning of parallel environments is VERY slow due to deepcopying the termination sets. I tried some work arounds, but nothing worked properly, so we are stuck with it for now. :param num_envs: an override for the default number of environments to spawn (in self.num_cpu) """ if not num_envs: num_envs = self.num_cpu if self.model_type == "A2C": # check to make sure that the A2C parameters are set assert self.optim_alpha self.training_envs = [deepcopy(self.env) for i in range(num_envs) ] # spawn parallel environments if self.acmodel.recurrent: self.memories = torch.zeros(num_envs, self.acmodel.memory_size, device=self.device) self.algo = torch_ac.A2CAlgo( self.training_envs, self.acmodel, self.device, self.frames_per_proc, self.discount, self.lr, self.gae_lambda, self.entropy_coef, self.value_loss_coef, self.max_grad_norm, self.recurrence, self.optim_alpha, self.optim_eps, self.preprocess_obss) elif self.model_type == "PPO2": # check to see if the PPO2 parameters are set assert self.clip_eps and self.epochs and self.batch_size self.training_envs = [deepcopy(self.env) for i in range(num_envs) ] # spawn parallel environments if self.acmodel.recurrent: self.memories = torch.zeros(num_envs, self.acmodel.memory_size, device=self.device) self.algo = torch_ac.PPOAlgo( self.training_envs, self.acmodel, self.device, self.frames_per_proc, self.discount, self.lr, self.gae_lambda, self.entropy_coef, self.value_loss_coef, self.max_grad_norm, self.recurrence, self.optim_eps, self.clip_eps, self.epochs, self.batch_size, self.preprocess_obss) else: raise ValueError("Incorrect algorithm name: {}".format(algo_type)) # load the optimizer state, if it exists if "optimizer_state" in self.status: self.algo.optimizer.load_state_dict(self.status["optimizer_state"]) self.txt_logger.info("Optimizer loaded\n") def learn(self, total_timesteps, log_interval=1, save_interval=10, save_env_info=False, save_loc=None): """ The primary training loop. :param total_timesteps: the total number of timesteps :param log_interval: the period between logging/printing updates :param save_interval: the number of updates between model saving :param save_env_info: if we save the environment info (termination set) VERY SLOW :return: True, if training is successful """ self.init_training_algo( ) # initialize the training algo/environment list/optimizer if save_loc: print( 'ignoring save_loc override. if this is not intended, fix me') # initialize parameters self.num_frames = self.status["num_frames"] self.update = self.status["update"] start_time = time.time() # loop until we reach the desired number of timesteps while self.num_frames < total_timesteps: # Update model parameters update_start_time = time.time( ) # store the time (for fps calculations) exps, logs1 = self.algo.collect_experiences( ) # collect a number of data points for training logs2 = self.algo.update_parameters( exps) # update the parameters based on the experiences logs = {**logs1, **logs2} # merge the logs for printing update_end_time = time.time() self.num_frames += logs["num_frames"] self.update += 1 # all of this messy stuff is just storing and printing the log info if self.update % log_interval == 0: fps = logs["num_frames"] / (update_end_time - update_start_time) duration = int(time.time() - start_time) return_per_episode = utils.synthesize( logs["return_per_episode"]) rreturn_per_episode = utils.synthesize( logs["reshaped_return_per_episode"]) num_frames_per_episode = utils.synthesize( logs["num_frames_per_episode"]) header = ["update", "frames", "FPS", "duration"] data = [self.update, self.num_frames, fps, duration] header += [ "rreturn_" + key for key in rreturn_per_episode.keys() ] data += rreturn_per_episode.values() header += [ "num_frames_" + key for key in num_frames_per_episode.keys() ] data += num_frames_per_episode.values() header += [ "entropy", "value", "policy_loss", "value_loss", "grad_norm" ] data += [ logs["entropy"], logs["value"], logs["policy_loss"], logs["value_loss"], logs["grad_norm"] ] self.txt_logger.info( "U {} | F {:06} | FPS {:04.0f} | D {} | rR:usmM {:.2f} {:.2f} {:.2f} {:.2f} | F:usmM {:.1f} {:.1f} {} {} | H {:.3f} | V {:.3f} | pL {:.3f} | vL {:.3f} | D {:.3f}" .format(*data)) header += [ "return_" + key for key in return_per_episode.keys() ] data += return_per_episode.values() if self.status["num_frames"] == 0: self.csv_logger.writerow(header) self.csv_logger.writerow(data) self.csv_file.flush() for field, value in zip(header, data): self.tb_writer.add_scalar(field, value, self.num_frames) # Save status if save_interval > 0 and self.update % save_interval == 0: self._save_training_info() if save_env_info: for e in self.training_envs: if hasattr(e, 'save_env_info'): e.save_env_info() self._clear_training_envs() return True def _save_training_info(self): """ Function to save the training info. """ # update the status dictionary self.status = { "num_frames": self.num_frames, "update": self.update, "model_state": self.acmodel.state_dict(), "optimizer_state": self.algo.optimizer.state_dict() } if hasattr(self.preprocess_obss, "vocab"): # if we are using NLP save, NLP info self.status["vocab"] = self.preprocess_obss.vocab.vocab utils.save_status(self.status, self.model_dir) # save the status info to model_dir self.txt_logger.info("Status saved") def _clear_training_envs(self): """ Clear the training environments to free up memory. """ # the termination set gets lost, so we need to store it again if hasattr(self.env, 'termination_set'): self.env.termination_set = [ s for e in self.training_envs for s in e.termination_set ] # clear the env and the training envs self.algo.env = None self.training_envs = None def save(self, f): """ Legacy function for saving the model. TODO: place the saving logic for the model here :param f: """ print('self.save() - currently not implemented') def set_env(self, env): """ Set the environment and clear the training environments :param env: environment for training/acting """ # check to make sure the environment is the correct type assert isinstance(env, gym.Env) self.env = env self.training_envs = None def predict(self, obs, state=None, deterministic=False): """ Wrapper for training code compatibility. Calls get_action() to predict the action to take based on the current observation. :param obs: observation for predicting the action :param state: state of the LSTM (unused) :param deterministic: whether to use deterministic or probabilistic actions (unused) :return: action and LSTM state """ # assert (state==None) and (deterministic==False) # still need to reimplement return self.get_action( obs ), None # return action, states - states is unused at the moment def get_actions(self, obss): """ Get a list of actions for a list of observations. :param obss: list of observations for predicting actions :return: list of actions for the associated observations """ preprocessed_obss = self.preprocess_obss(obss, device=self.device) with torch.no_grad( ): # don't calculate the gradients, since we are doing a forward pass if self.acmodel.recurrent: # if we are using a recurrent model dist, _, self.memories = self.acmodel(preprocessed_obss, self.memories) else: # otherwise dist, _ = self.acmodel(preprocessed_obss) # preprocess the observations to put them in a torch-friendly format # the acmodel returns a probability distribution if self.argmax: # if we are detemrinistic, take the action with the highest probability actions = dist.probs.max(1, keepdim=True)[1] else: # otherwise sample the distribution to select the action actions = dist.sample() return actions.cpu().numpy() # reaturn a numpy array, not a tensor def get_action(self, obs): """ Wrapper for get_actions() to produce just a single action (rather than a list of actions) for acting. :param obs: single observation :return: single action """ return self.get_actions([obs])[0] def analyze_feedbacks(self, rewards, dones): """ rl-starter-files code. Not sure what this does. :param rewards: :param dones: """ if self.acmodel.recurrent: masks = 1 - torch.tensor( dones, dtype=torch.float, device=self.device).unsqueeze(1) self.memories *= masks def analyze_feedback(self, reward, done): """ rl-starter-files code. Not sure what this does (other than wrap analyze_feedbacks(). :param reward: :param done: :return: """ return self.analyze_feedbacks([reward], [done])
class Agent: """An agent. It is able: - to choose an action given an observation, - to analyze the feedback (i.e. reward and done state) of its action.""" def __init__(self, env, obs_space, action_space, model_dir, device=None, argmax=False, num_envs=1, use_memory=False, use_text=False): obs_space, self.preprocess_obs_goals = utils.get_obs_goals_preprocessor( obs_space) self.acmodel = ACModel(obs_space, action_space, use_memory=use_memory, use_text=use_text) self.device = device self.argmax = argmax self.num_envs = num_envs status = utils.get_status(model_dir) self.goals = list(status['agent_goals'].values()) # for goal in self.goals: # goal = env.unwrapped.get_obs_render( goal, tile_size=32) # plt.imshow(goal) # plt.show() if self.acmodel.recurrent: self.memories = torch.zeros(self.num_envs, self.acmodel.memory_size, device=self.device) self.acmodel.load_state_dict(status["model_state"]) self.acmodel.to(self.device) self.acmodel.eval() if hasattr(self.preprocess_obs_goals, "vocab"): self.preprocess_obs_goals.vocab.load_vocab(status["vocab"]) def concat_obs_goal(self, obs): if 'image' in obs: obs_goals = [{ "image": np.concatenate((obs["image"], self.goals[i]), axis=2), "mission": obs['mission'] } for i in range(len(self.goals))] else: obs_goals = [ np.concatenate((obs, self.goals[i]), axis=2) for i in range(len(self.goals)) ] return obs_goals def get_actions(self, obss): actions = np.zeros(len(obss), dtype=int) for i in range(len(obss)): memory = self.memories[i] obs_goals = self.concat_obs_goal(obss[i]) preprocessed_obs_goals = self.preprocess_obs_goals( obs_goals, device=self.device) with torch.no_grad(): if self.acmodel.recurrent: memory = torch.stack([memory] * len(self.goals), 0) dists, values, memory = self.acmodel( preprocessed_obs_goals, memory) else: dists, values = self.acmodel(preprocessed_obs_goals) g = values.data.max(0)[1] print(values.data, g) if self.argmax: actions[i] = dists.probs.max(1, keepdim=True)[1][g].cpu().numpy() else: actions[i] = dists.sample()[g].cpu().numpy() if self.acmodel.recurrent: self.memories[i] = memory[g] return actions def reset(self): if self.acmodel.recurrent: self.memories = torch.zeros(self.num_envs, self.acmodel.memory_size, device=self.device) def get_action(self, obs): return self.get_actions([obs])[0] def analyze_feedbacks(self, rewards, dones): if self.acmodel.recurrent: masks = 1 - torch.tensor( dones, dtype=torch.float, device=self.device).unsqueeze(1) self.memories *= masks def analyze_feedback(self, reward, done): return self.analyze_feedbacks([reward], [done])
algo.env.change_difficulty(difficulty) best_val = 0 if status["num_frames"] == 0: csv_writer.writerow(header) csv_writer.writerow(data) csv_file.flush() if args.tb: for field, value in zip(header, data): tb_writer.add_scalar(field, value, num_frames) status = {"num_frames": num_frames, "update": update} # Save vocabulary and model if args.save_interval > 0 and update % args.save_interval == 0: preprocess_obss.vocab.save() if torch.cuda.is_available(): base_model.cpu() model_to_save = ACModel(obs_space, envs[0].action_space, args.mem, args.text) model_to_save.load_state_dict(best_model) utils.save_model(model_to_save, model_dir) logger.info("Model successfully saved") if torch.cuda.is_available(): base_model.cuda() utils.save_status(status, model_dir)
class Agent: """An agent. It is able: - to choose an action given an observation, - to analyze the feedback (i.e. reward and done state) of its action.""" def __init__(self, env, obs_space, action_space, model_dir, ignoreLTL, progression_mode, gnn, recurrence=1, dumb_ac=False, device=None, argmax=False, num_envs=1): try: print(model_dir) status = utils.get_status(model_dir) except OSError: status = {"num_frames": 0, "update": 0} using_gnn = (gnn != "GRU" and gnn != "LSTM") obs_space, self.preprocess_obss = utils.get_obss_preprocessor( env, using_gnn, progression_mode) if "vocab" in status and self.preprocess_obss.vocab is not None: self.preprocess_obss.vocab.load_vocab(status["vocab"]) if recurrence > 1: self.acmodel = RecurrentACModel(env, obs_space, action_space, ignoreLTL, gnn, dumb_ac, True) self.memories = torch.zeros(num_envs, self.acmodel.memory_size, device=device) else: self.acmodel = ACModel(env, obs_space, action_space, ignoreLTL, gnn, dumb_ac, True) self.device = device self.argmax = argmax self.num_envs = num_envs self.acmodel.load_state_dict(utils.get_model_state(model_dir)) self.acmodel.to(self.device) self.acmodel.eval() def get_actions(self, obss): preprocessed_obss = self.preprocess_obss(obss, device=self.device) with torch.no_grad(): if self.acmodel.recurrent: dist, _, self.memories = self.acmodel(preprocessed_obss, self.memories) else: dist, _ = self.acmodel(preprocessed_obss) if self.argmax: actions = dist.probs.max(1, keepdim=True)[1] else: actions = dist.sample() return actions.cpu().numpy() def get_action(self, obs): return self.get_actions([obs])[0] def analyze_feedbacks(self, rewards, dones): if self.acmodel.recurrent: masks = 1 - torch.tensor(dones, dtype=torch.float).unsqueeze(1) self.memories *= masks def analyze_feedback(self, reward, done): return self.analyze_feedbacks([reward], [done])
def tuner(icm_lr, reward_weighting, normalise_rewards, args): import argparse import datetime import torch import torch_ac import tensorboardX import sys import numpy as np from model import ACModel from .a2c import A2CAlgo # from .ppo import PPOAlgo frames_to_visualise = 200 # Parse arguments args.mem = args.recurrence > 1 def make_exploration_heatmap(args, plot_title): import numpy as np import matplotlib.pyplot as plt visitation_counts = np.load( f"{args.model}_visitation_counts.npy", allow_pickle=True ) plot_title = str(np.count_nonzero(visitation_counts)) + args.model plt.imshow(np.log(visitation_counts)) plt.colorbar() plt.title(plot_title) plt.savefig(f"{plot_title}_visitation_counts.png") # Set run dir date = datetime.datetime.now().strftime("%y-%m-%d-%H-%M-%S") default_model_name = f"{args.env}_{args.algo}_seed{args.seed}_{date}" model_name = args.model or default_model_name model_dir = utils.get_model_dir(model_name) # Load loggers and Tensorboard writer txt_logger = utils.get_txt_logger(model_dir) csv_file, csv_logger = utils.get_csv_logger(model_dir) tb_writer = tensorboardX.SummaryWriter(model_dir) # Log command and all script arguments txt_logger.info("{}\n".format(" ".join(sys.argv))) txt_logger.info("{}\n".format(args)) # Set seed for all randomness sources utils.seed(args.seed) # Set device device = "cpu" # torch.device("cuda" if torch.cuda.is_available() else "cpu") txt_logger.info(f"Device: {device}\n") # Load environments envs = [] for i in range(16): an_env = utils.make_env( args.env, int(args.frames_before_reset), int(args.environment_seed) ) envs.append(an_env) txt_logger.info("Environments loaded\n") # Load training status try: status = utils.get_status(model_dir) except OSError: status = {"num_frames": 0, "update": 0} txt_logger.info("Training status loaded\n") # Load observations preprocessor obs_space, preprocess_obss = utils.get_obss_preprocessor(envs[0].observation_space) if "vocab" in status: preprocess_obss.vocab.load_vocab(status["vocab"]) txt_logger.info("Observations preprocessor loaded") # Load model acmodel = ACModel(obs_space, envs[0].action_space, args.mem, args.text) if "model_state" in status: acmodel.load_state_dict(status["model_state"]) acmodel.to(device) txt_logger.info("Model loaded\n") txt_logger.info("{}\n".format(acmodel)) # Load algo # adapted from impact driven RL from .models import AutoencoderWithUncertainty autoencoder = AutoencoderWithUncertainty(observation_shape=(7, 7, 3)).to(device) autoencoder_opt = torch.optim.Adam( autoencoder.parameters(), lr=icm_lr, weight_decay=0 ) if args.algo == "a2c": algo = A2CAlgo( envs, acmodel, autoencoder, autoencoder_opt, args.uncertainty, args.noisy_tv, args.curiosity, args.randomise_env, args.uncertainty_budget, args.environment_seed, reward_weighting, normalise_rewards, args.frames_before_reset, device, args.frames_per_proc, args.discount, args.lr, args.gae_lambda, args.entropy_coef, args.value_loss_coef, args.max_grad_norm, args.recurrence, args.optim_alpha, args.optim_eps, preprocess_obss, None, args.random_action, ) elif args.algo == "ppo": algo = PPOAlgo( envs, acmodel, autoencoder, autoencoder_opt, args.uncertainty, args.noisy_tv, args.curiosity, args.randomise_env, args.uncertainty_budget, args.environment_seed, reward_weighting, normalise_rewards, device, args.frames_per_proc, args.discount, args.lr, args.gae_lambda, args.entropy_coef, args.value_loss_coef, args.max_grad_norm, args.recurrence, args.optim_eps, args.clip_eps, args.epochs, args.batch_size, preprocess_obss, ) else: raise ValueError("Incorrect algorithm name: {}".format(args.algo)) if "optimizer_state" in status: algo.optimizer.load_state_dict(status["optimizer_state"]) txt_logger.info("Optimizer loaded\n") # Train model num_frames = status["num_frames"] update = status["update"] start_time = time.time() while num_frames < args.frames: # Update model parameters update_start_time = time.time() exps, logs1 = algo.collect_experiences() logs2 = algo.update_parameters(exps) logs = {**logs1, **logs2} update_end_time = time.time() num_frames += logs["num_frames"] update += 1 log_to_wandb(logs, start_time, update_start_time, update_end_time) # Print logs if update % args.log_interval == 0: fps = logs["num_frames"] / (update_end_time - update_start_time) duration = int(time.time() - start_time) return_per_episode = utils.synthesize(logs["return_per_episode"]) rreturn_per_episode = utils.synthesize(logs["reshaped_return_per_episode"]) num_frames_per_episode = utils.synthesize(logs["num_frames_per_episode"]) header = ["update", "frames", "FPS", "duration"] data = [update, num_frames, fps, duration] header += ["rreturn_" + key for key in rreturn_per_episode.keys()] data += rreturn_per_episode.values() header += ["num_frames_" + key for key in num_frames_per_episode.keys()] data += num_frames_per_episode.values() header += [ "intrinsic_rewards", "uncertainties", "novel_states_visited", "entropy", "value", "policy_loss", "value_loss", "grad_norm", ] data += [ logs["intrinsic_rewards"].mean().item(), logs["uncertainties"].mean().item(), logs["novel_states_visited"].mean().item(), logs["entropy"], logs["value"], logs["policy_loss"], logs["value_loss"], logs["grad_norm"], ] txt_logger.info( "U {} | F {:06} | FPS {:04.0f} | D {} | rR:μσmM {:.2f} {:.2f} {:.2f} {:.2f} | F:μσmM {:.1f} {:.1f} {} {} | H {:.3f} | V {:.3f} | pL {:.3f}".format( *data ) ) # Save status if args.save_interval > 0 and update % args.save_interval == 0: status = { "num_frames": num_frames, "update": update, "model_state": acmodel.state_dict(), "optimizer_state": algo.optimizer.state_dict(), } if hasattr(preprocess_obss, "vocab"): status["vocab"] = preprocess_obss.vocab.vocab utils.save_status(status, model_dir) return