import os import gc import copy import tqdm import wandb import random import argparse import numpy as np import utils from utils import ReplayBuffer, FreezeParameters, make_env, preprocess_obs, soft_update_params, save_image from models import ObservationEncoder, ObservationDecoder, TransitionModel, Actor, ValueModel, RewardModel, ProjectionHead, ContrastiveHead, CLUBSample from logger import Logger from video import VideoRecorder from dmc2gym.wrappers import set_global_var import torch import torch.nn as nn import torch.nn.functional as F import torchvision.transforms as T from torch.utils.tensorboard import SummaryWriter #from agent.baseline_agent import BaselineAgent #from agent.bisim_agent import BisimAgent #from agent.deepmdp_agent import DeepMDPAgent #from agents.navigation.carla_env import CarlaEnv def parse_args(): parser = argparse.ArgumentParser() # environment parser.add_argument('--domain_name', default='cheetah') parser.add_argument('--version', default=1, type=int) parser.add_argument('--task_name', default='run') parser.add_argument('--image_size', default=84, type=int) parser.add_argument('--channels', default=3, type=int) parser.add_argument('--action_repeat', default=2, type=int) parser.add_argument('--frame_stack', default=3, type=int) parser.add_argument('--collection_interval', default=100, type=int) parser.add_argument('--resource_files', type=str) parser.add_argument('--eval_resource_files', type=str) parser.add_argument('--img_source', default=None, type=str, choices=['color', 'noise', 'images', 'video', 'none']) parser.add_argument('--total_frames', default=1000, type=int) # 10000 parser.add_argument('--high_noise', action='store_true') # replay buffer parser.add_argument('--replay_buffer_capacity', default=50000, type=int) #50000 parser.add_argument('--episode_length', default=51, type=int) # train parser.add_argument('--agent', default='dpi', type=str, choices=['baseline', 'bisim', 'deepmdp', 'db', 'dpi', 'rpc']) parser.add_argument('--init_steps', default=10000, type=int) parser.add_argument('--num_train_steps', default=10000, type=int) parser.add_argument('--batch_size', default=30, type=int) #512 parser.add_argument('--state_size', default=256, type=int) parser.add_argument('--hidden_size', default=128, type=int) parser.add_argument('--history_size', default=128, type=int) parser.add_argument('--num-units', type=int, default=50, help='num hidden units for reward/value/discount models') parser.add_argument('--load_encoder', default=None, type=str) parser.add_argument('--imagine_horizon', default=15, type=str) parser.add_argument('--grad_clip_norm', type=float, default=100.0, help='Gradient clipping norm') # eval parser.add_argument('--eval_freq', default=10, type=int) # TODO: master had 10000 parser.add_argument('--num_eval_episodes', default=20, type=int) # value parser.add_argument('--value_lr', default=8e-5, type=float) parser.add_argument('--value_beta', default=0.9, type=float) parser.add_argument('--value_tau', default=0.005, type=float) parser.add_argument('--value_target_update_freq', default=100, type=int) parser.add_argument('--td_lambda', default=0.95, type=int) # actor parser.add_argument('--actor_lr', default=8e-5, type=float) parser.add_argument('--actor_beta', default=0.9, type=float) parser.add_argument('--actor_log_std_min', default=-10, type=float) parser.add_argument('--actor_log_std_max', default=2, type=float) parser.add_argument('--actor_update_freq', default=2, type=int) # world/encoder/decoder parser.add_argument('--encoder_type', default='pixel', type=str, choices=['pixel', 'pixelCarla096', 'pixelCarla098', 'identity']) parser.add_argument('--encoder_feature_dim', default=50, type=int) parser.add_argument('--world_model_lr', default=6e-4, type=float) parser.add_argument('--past_transition_lr', default=1e-3, type=float) parser.add_argument('--encoder_lr', default=1e-3, type=float) parser.add_argument('--encoder_tau', default=0.001, type=float) parser.add_argument('--encoder_stride', default=1, type=int) parser.add_argument('--decoder_type', default='pixel', type=str, choices=['pixel', 'identity', 'contrastive', 'reward', 'inverse', 'reconstruction']) parser.add_argument('--decoder_lr', default=1e-3, type=float) parser.add_argument('--decoder_update_freq', default=1, type=int) parser.add_argument('--decoder_weight_lambda', default=0.0, type=float) parser.add_argument('--num_layers', default=4, type=int) parser.add_argument('--num_filters', default=32, type=int) parser.add_argument('--aug', action='store_true') # sac parser.add_argument('--discount', default=0.99, type=float) parser.add_argument('--init_temperature', default=0.01, type=float) parser.add_argument('--alpha_lr', default=1e-3, type=float) parser.add_argument('--alpha_beta', default=0.9, type=float) # misc parser.add_argument('--seed', default=1, type=int) parser.add_argument('--logging_freq', default=100, type=int) parser.add_argument('--saving_interval', default=1000, type=int) parser.add_argument('--work_dir', default='.', type=str) parser.add_argument('--save_tb', default=False, action='store_true') parser.add_argument('--save_model', default=False, action='store_true') parser.add_argument('--save_buffer', default=False, action='store_true') parser.add_argument('--save_video', default=False, action='store_true') parser.add_argument('--transition_model_type', default='', type=str, choices=['', 'deterministic', 'probabilistic', 'ensemble']) parser.add_argument('--render', default=False, action='store_true') args = parser.parse_args() return args class DPI: def __init__(self, args): # wandb config #run = wandb.init(project="dpi") self.args = args # set environment noise set_global_var(self.args.high_noise) # environment setup self.env = make_env(self.args) #self.args.seed = np.random.randint(0, 1000) self.env.seed(self.args.seed) # noiseless environment setup self.args.version = 2 # env_id changes to v2 self.args.img_source = None # no image noise self.args.resource_files = None # stack several consecutive frames together if self.args.encoder_type.startswith('pixel'): self.env = utils.FrameStack(self.env, k=self.args.frame_stack) self.env = utils.ActionRepeat(self.env, self.args.action_repeat) self.env = utils.NormalizeActions(self.env) # create replay buffer self.data_buffer = ReplayBuffer(size=self.args.replay_buffer_capacity, obs_shape=(self.args.frame_stack*self.args.channels,self.args.image_size,self.args.image_size), action_size=self.env.action_space.shape[0], seq_len=self.args.episode_length, batch_size=args.batch_size, args=self.args) # create work directory utils.make_dir(self.args.work_dir) self.video_dir = utils.make_dir(os.path.join(self.args.work_dir, 'video')) self.model_dir = utils.make_dir(os.path.join(self.args.work_dir, 'model')) self.buffer_dir = utils.make_dir(os.path.join(self.args.work_dir, 'buffer')) # create models self.build_models(use_saved=False, saved_model_dir=self.model_dir) def build_models(self, use_saved, saved_model_dir=None): # World Models self.obs_encoder = ObservationEncoder( obs_shape=(self.args.frame_stack*self.args.channels,self.args.image_size,self.args.image_size), # (9,84,84) state_size=self.args.state_size # 128 ).to(device) self.obs_encoder_momentum = ObservationEncoder( obs_shape=(self.args.frame_stack*self.args.channels,self.args.image_size,self.args.image_size), # (9,84,84) state_size=self.args.state_size # 128 ).to(device) self.obs_decoder = ObservationDecoder( state_size=self.args.state_size, # 128 output_shape=(self.args.channels,self.args.image_size,self.args.image_size) # (3,84,84) ).to(device) self.transition_model = TransitionModel( state_size=self.args.state_size, # 128 hidden_size=self.args.hidden_size, # 256 action_size=self.env.action_space.shape[0], # 6 history_size=self.args.history_size, # 128 ).to(device) # Actor Model self.actor_model = Actor( state_size=self.args.state_size, # 128 hidden_size=self.args.hidden_size, # 256, action_size=self.env.action_space.shape[0], # 6 ).to(device) # Value Models self.value_model = ValueModel( state_size=self.args.state_size, # 128 hidden_size=self.args.hidden_size, # 256 ).to(device) self.target_value_model = ValueModel( state_size=self.args.state_size, # 128 hidden_size=self.args.hidden_size, # 256 ).to(device) self.reward_model = RewardModel( state_size=self.args.state_size, # 128 hidden_size=self.args.hidden_size, # 256 ).to(device) # Contrastive Models self.prjoection_head = ProjectionHead( state_size=self.args.state_size, # 128 action_size=self.env.action_space.shape[0], # 6 hidden_size=self.args.hidden_size, # 256 ).to(device) self.prjoection_head_momentum = ProjectionHead( state_size=self.args.state_size, # 128 action_size=self.env.action_space.shape[0], # 6 hidden_size=self.args.hidden_size, # 256 ).to(device) self.contrastive_head = ContrastiveHead( hidden_size=self.args.hidden_size, # 256 ).to(device) # model parameters self.world_model_parameters = list(self.obs_encoder.parameters()) + list(self.obs_decoder.parameters()) + \ list(self.value_model.parameters()) + list(self.transition_model.parameters()) + \ list(self.prjoection_head.parameters()) self.past_transition_parameters = self.transition_model.parameters() # optimizers self.world_model_opt = torch.optim.Adam(self.world_model_parameters, self.args.world_model_lr) self.value_opt = torch.optim.Adam(self.value_model.parameters(), self.args.value_lr) self.actor_opt = torch.optim.Adam(self.actor_model.parameters(), self.args.actor_lr) self.past_transition_opt = torch.optim.Adam(self.past_transition_parameters, self.args.past_transition_lr) # Create Modules self.world_model_modules = [self.obs_encoder, self.obs_decoder, self.reward_model, self.transition_model, self.prjoection_head] self.value_modules = [self.value_model] self.actor_modules = [self.actor_model] if use_saved: self._use_saved_models(saved_model_dir) def _use_saved_models(self, saved_model_dir): self.obs_encoder.load_state_dict(torch.load(os.path.join(saved_model_dir, 'obs_encoder.pt'))) self.obs_decoder.load_state_dict(torch.load(os.path.join(saved_model_dir, 'obs_decoder.pt'))) self.transition_model.load_state_dict(torch.load(os.path.join(saved_model_dir, 'transition_model.pt'))) def collect_sequences(self, episodes, random=True, actor_model=None, encoder_model=None): obs = self.env.reset() done = False all_rews = [] #video = VideoRecorder(self.video_dir if args.save_video else None, resource_files=args.resource_files) for episode_count in tqdm.tqdm(range(episodes), desc='Collecting episodes'): if args.save_video: self.env.video.init(enabled=True) epi_reward = 0 for i in range(self.args.episode_length): if random: action = self.env.action_space.sample() else: with torch.no_grad(): obs_torch = torch.unsqueeze(torch.tensor(obs).float(),0).to(device) state = self.obs_encoder(obs_torch)["distribution"].sample() action = self.actor_model(state).cpu().detach().numpy().squeeze() next_obs, rew, done, _ = self.env.step(action) self.data_buffer.add(obs, action, next_obs, rew, episode_count+1, done) if args.save_video: self.env.video.record(self.env) if done or i == self.args.episode_length-1: obs = self.env.reset() done=False else: obs = next_obs epi_reward += rew all_rews.append(epi_reward) if args.save_video: self.env.video.save('noisy/%d.mp4' % episode_count) print("Collected {} random episodes".format(episode_count+1)) return all_rews def train(self, step, total_steps): counter = 0 while step < total_steps: # collect experience if step !=0: encoder = self.obs_encoder actor = self.actor_model #all_rews = self.collect_sequences(self.args.batch_size, random=True) all_rews = self.collect_sequences(self.args.batch_size, random=False, actor_model=actor, encoder_model=encoder) else: all_rews = self.collect_sequences(self.args.batch_size, random=True) # Group by steps and sample random batch random_indices = self.data_buffer.sample_random_idx(self.args.batch_size * ((step//self.args.collection_interval)+1)) # random indices for batch #random_indices = np.arange(self.args.batch_size * ((step//self.args.collection_interval)),self.args.batch_size * ((step//self.args.collection_interval)+1)) last_observations = self.data_buffer.group_and_sample_random_batch(self.data_buffer,"observations", "cpu", random_indices=random_indices) current_observations = self.data_buffer.group_and_sample_random_batch(self.data_buffer,"next_observations", device="cpu", random_indices=random_indices) next_observations = self.data_buffer.group_and_sample_random_batch(self.data_buffer,"next_observations", device="cpu", offset=1, random_indices=random_indices) actions = self.data_buffer.group_and_sample_random_batch(self.data_buffer,"actions", device=device, is_obs=False, random_indices=random_indices) next_actions = self.data_buffer.group_and_sample_random_batch(self.data_buffer,"actions", device=device, is_obs=False, offset=1, random_indices=random_indices) rewards = self.data_buffer.group_and_sample_random_batch(self.data_buffer,"rewards", device=device, is_obs=False, offset=1, random_indices=random_indices) # Preprocessing last_observations = preprocess_obs(last_observations).to(device) current_observations = preprocess_obs(current_observations).to(device) next_observations = preprocess_obs(next_observations).to(device) # Initialize transition model states self.transition_model.init_states(self.args.batch_size, device) # (N,128) self.history = self.transition_model.prev_history # (N,128) # Train encoder if step == 0: step += 1 for _ in range(self.args.collection_interval // self.args.episode_length+1): counter += 1 for i in range(self.args.episode_length-1): if i > 0: # Encode observations and next_observations self.last_states_dict = self.get_features(last_observations[i]) self.current_states_dict = self.get_features(current_observations[i]) self.next_states_dict = self.get_features(next_observations[i], momentum=True) self.action = actions[i] # (N,6) self.next_action = next_actions[i] # (N,6) history = self.transition_model.prev_history # Encode negative observations idx = torch.randperm(current_observations[i].shape[0]) # random permutation on batch random_time_index = torch.randint(0, self.args.episode_length-2, (1,)).item() # random time index negative_current_observations = current_observations[random_time_index][idx] self.negative_current_states_dict = self.obs_encoder(negative_current_observations) # Predict current state from past state with transition model last_states_sample = self.last_states_dict["sample"] predicted_current_state_dict = self.transition_model.imagine_step(last_states_sample, self.action, self.history) self.history = predicted_current_state_dict["history"] # Calculate upper bound loss likeli_loss, ub_loss = self._upper_bound_minimization(self.last_states_dict, self.current_states_dict, self.negative_current_states_dict, predicted_current_state_dict ) # Calculate encoder loss encoder_loss = self._past_encoder_loss(self.current_states_dict, predicted_current_state_dict) # contrastive projection vec_anchor = predicted_current_state_dict["sample"] vec_positive = self.next_states_dict["sample"].detach() z_anchor = self.prjoection_head(vec_anchor, self.action) z_positive = self.prjoection_head_momentum(vec_positive, next_actions[i]).detach() # contrastive loss logits = self.contrastive_head(z_anchor, z_positive) labels = torch.arange(logits.shape[0]).long().to(device) lb_loss = F.cross_entropy(logits, labels) # behaviour learning with FreezeParameters(self.world_model_modules): imagine_horizon = self.args.imagine_horizon #np.minimum(self.args.imagine_horizon, self.args.episode_length-1-i) imagined_rollout = self.transition_model.imagine_rollout(self.current_states_dict["sample"].detach(), self.next_action, self.history.detach(), imagine_horizon) # decoder loss horizon = np.minimum(self.args.imagine_horizon, self.args.episode_length-1-i) obs_dist = self.obs_decoder(imagined_rollout["sample"][:horizon]) decoder_loss = -torch.mean(obs_dist.log_prob(next_observations[i:i+horizon][:,:,:3,:,:])) # reward loss reward_dist = self.reward_model(self.current_states_dict["sample"]) reward_loss = -torch.mean(reward_dist.log_prob(rewards[:-1])) # update models world_model_loss = encoder_loss + 100 * ub_loss + lb_loss + reward_loss + decoder_loss * 1e-2 self.world_model_opt.zero_grad() world_model_loss.backward() nn.utils.clip_grad_norm_(self.world_model_parameters, self.args.grad_clip_norm) self.world_model_opt.step() # update momentum encoder soft_update_params(self.obs_encoder, self.obs_encoder_momentum, self.args.encoder_tau) # update momentum projection head soft_update_params(self.prjoection_head, self.prjoection_head_momentum, self.args.encoder_tau) # actor loss with FreezeParameters(self.world_model_modules + self.value_modules): imag_rew_dist = self.reward_model(imagined_rollout["sample"]) target_imag_val_dist = self.target_value_model(imagined_rollout["sample"]) imag_rews = imag_rew_dist.mean target_imag_vals = target_imag_val_dist.mean discounts = self.args.discount * torch.ones_like(imag_rews).detach() self.target_returns = self._compute_lambda_return(imag_rews[:-1], target_imag_vals[:-1], discounts[:-1] , self.args.td_lambda, target_imag_vals[-1]) discounts = torch.cat([torch.ones_like(discounts[:1]), discounts[1:-1]], 0) self.discounts = torch.cumprod(discounts, 0).detach() actor_loss = -torch.mean(self.discounts * self.target_returns) # update actor self.actor_opt.zero_grad() actor_loss.backward() nn.utils.clip_grad_norm_(self.actor_model.parameters(), self.args.grad_clip_norm) self.actor_opt.step() # value loss with torch.no_grad(): value_feat = imagined_rollout["sample"][:-1].detach() value_targ = self.target_returns.detach() value_dist = self.value_model(value_feat) value_loss = -torch.mean(self.discounts * value_dist.log_prob(value_targ).unsqueeze(-1)) # update value self.value_opt.zero_grad() value_loss.backward() nn.utils.clip_grad_norm_(self.value_model.parameters(), self.args.grad_clip_norm) self.value_opt.step() # update target value if step % self.args.value_target_update_freq == 0: self.target_value_model = copy.deepcopy(self.value_model) # counter for reward count = np.arange((counter-1) * (self.args.batch_size), (counter) * (self.args.batch_size)) if step % self.args.logging_freq: writer.add_scalar('World Loss/World Loss', world_model_loss.detach().item(), step) writer.add_scalar('Main Models Loss/Encoder Loss', encoder_loss.detach().item(), step) writer.add_scalar('Main Models Loss/Decoder Loss', decoder_loss, step) writer.add_scalar('Actor Critic Loss/Actor Loss', actor_loss.detach().item(), step) writer.add_scalar('Actor Critic Loss/Value Loss', value_loss.detach().item(), step) writer.add_scalar('Actor Critic Loss/Reward Loss', reward_loss.detach().item(), step) writer.add_scalar('Bound Loss/Upper Bound Loss', ub_loss.detach().item(), step) writer.add_scalar('Bound Loss/Lower Bound Loss', lb_loss.detach().item(), step) step += 1 if step>total_steps: print("Training finished") break # save model if step % self.args.saving_interval == 0: path = os.path.dirname(os.path.realpath(__file__)) + "/saved_models/models.pth" self.save_models(path) #torch.cuda.empty_cache() # memory leak issues for j in range(len(all_rews)): writer.add_scalar('Rewards/Rewards', all_rews[j], count[j]) def evaluate(self, env, eval_episodes, render=False): episode_rew = np.zeros((eval_episodes)) video_images = [[] for _ in range(eval_episodes)] for i in range(eval_episodes): obs = env.reset() done = False prev_state = self.rssm.init_state(1, self.device) prev_action = torch.zeros(1, self.action_size).to(self.device) while not done: with torch.no_grad(): posterior, action = self.act_with_world_model(obs, prev_state, prev_action) action = action[0].cpu().numpy() next_obs, rew, done, _ = env.step(action) prev_state = posterior prev_action = torch.tensor(action, dtype=torch.float32).to(self.device).unsqueeze(0) episode_rew[i] += rew if render: video_images[i].append(obs['image'].transpose(1,2,0).copy()) obs = next_obs return episode_rew, np.array(video_images[:self.args.max_videos_to_save]) def _upper_bound_minimization(self, last_states, current_states, negative_current_states, predicted_current_states): club_sample = CLUBSample(last_states, current_states, negative_current_states, predicted_current_states) likelihood_loss = club_sample.learning_loss() club_loss = club_sample() return likelihood_loss, club_loss def _past_encoder_loss(self, curr_states_dict, predicted_curr_states_dict): # current state distribution curr_states_dist = curr_states_dict["distribution"] # predicted current state distribution predicted_curr_states_dist = predicted_curr_states_dict["distribution"] # KL divergence loss loss = torch.distributions.kl.kl_divergence(curr_states_dist, predicted_curr_states_dist).mean() return loss def get_features(self, x, momentum=False): if self.args.aug: x = T.RandomCrop((80, 80))(x) # (None,80,80,4) x = T.functional.pad(x, (4, 4, 4, 4), "symmetric") # (None,88,88,4) x = T.RandomCrop((84, 84))(x) # (None,84,84,4) with torch.no_grad(): if momentum: x = self.obs_encoder_momentum(x) else: x = self.obs_encoder(x) return x def _compute_lambda_return(self, rewards, values, discounts, td_lam, last_value): next_values = torch.cat([values[1:], last_value.unsqueeze(0)],0) targets = rewards + discounts * next_values * (1-td_lam) rets =[] last_rew = last_value for t in range(rewards.shape[0]-1, -1, -1): last_rew = targets[t] + discounts[t] * td_lam *(last_rew) rets.append(last_rew) returns = torch.flip(torch.stack(rets), [0]) return returns def save_models(self, save_path): torch.save( {'rssm' : self.transition_model.state_dict(), 'actor': self.actor_model.state_dict(), 'reward_model': self.reward_model.state_dict(), 'obs_encoder': self.obs_encoder.state_dict(), 'obs_decoder': self.obs_decoder.state_dict(), 'actor_optimizer': self.actor_opt.state_dict(), 'value_optimizer': self.value_opt.state_dict(), 'world_model_optimizer': self.world_model_opt.state_dict(),}, save_path) if __name__ == '__main__': args = parse_args() writer = SummaryWriter() device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu') step = 0 total_steps = 10000 dpi = DPI(args) dpi.train(step,total_steps)