def main(): os.environ['OMP_NUM_THREADS'] = '1' num_updates = int(args.num_frames) // args.num_steps // args.num_processes logger = ModelLogger(args.log_dir, args.num_processes) envs = make_env_vec(args.num_processes, args.env_name, args.seed, args.log_dir, args.start_container, max_steps=1200, discrete_wrapper=args.discrete_actions) action_shape = 1 if envs.action_space.__class__.__name__ == "Discrete" else envs.action_space.shape[ 0] obs_shape = (envs.observation_space.shape[0] * args.num_stack, *envs.observation_space.shape[1:]) distribution = 'MixedDistribution' if args.use_mixed else 'DiagGaussian' actor_critic = CNNPolicy(obs_shape[0], envs.action_space, args.recurrent_policy, distribution=distribution) print_model_size(actor_critic) print(obs_shape) if args.cuda: actor_critic.cuda() if args.algo == 'a2c': optimizer = optim.RMSprop(actor_critic.parameters(), args.lr, eps=args.eps, alpha=args.alpha) rollouts = RolloutStorage(args.num_steps, args.num_processes, obs_shape, envs.action_space, actor_critic.state_size) current_obs = torch.zeros(args.num_processes, *obs_shape) obs = envs.reset() current_obs = update_current_obs(obs, current_obs, envs.observation_space.shape[0], args.num_stack) rollouts.observations[0].copy_(current_obs) if args.cuda: current_obs = current_obs.cuda() rollouts.cuda() start = time.time() for j in range(num_updates): #Running an episode for step in range(args.num_steps): # Sample actions value, action, action_log_prob, states = actor_critic.act( Variable(rollouts.observations[step]), Variable(rollouts.states[step]), Variable(rollouts.masks[step])) continous_action, discrete_action = action continous_action_log_prob, discrete_action_log_prob = action_log_prob cpu_actions = continous_action.data.squeeze(1).cpu().numpy() # Exploration epsilon greedy, ToDo better exploration policy #if np.random.random_sample() < args.exp_probability: #cpu_actions = [envs.action_space.sample() for _ in range(args.num_processes)] # Observation, reward and next obs obs, reward, done, info = envs.step(cpu_actions) #ToDo better collision strategy '''for i, flag in enumerate(done): if flag == True: envs.envs[i].user_tile_start = info[i]['Simulator']['tile_coords'] envs.envs[i].reset() envs.envs[i].user_tile_start = None''' slack = args.reward_slack scaled_reward = np.clip(reward + slack, a_min=-2.0**args.reward_pow, a_max=None) #scaled_reward = np.clip(reward + slack, a_min = -40.0, a_max=None) for i in range(args.num_processes): if scaled_reward[i] > 0: scaled_reward[i] = (1 + scaled_reward[i])**args.reward_pow - 1 scaled_reward = torch.from_numpy( np.expand_dims(np.stack(scaled_reward), 1)).float() '''if step != 0: cur_angle = continous_action[:, 1] scaled_reward -= (torch.abs(prev_angle - cur_angle).view(-1).data.cpu()**args.reward_facpow)*args.reward_factor prev_angle = continous_action[:, 1]''' reward = np.clip(reward, a_min=-4.0, a_max=None) + 1.0 reward = torch.from_numpy(np.expand_dims(np.stack(reward), 1)).float() masks = torch.FloatTensor([[0.0] if done_ else [1.0] for done_ in done]) logger.update_reward(reward, masks) if args.cuda: masks = masks.cuda() if current_obs.dim() == 4: current_obs *= masks.unsqueeze(2).unsqueeze(2) else: current_obs *= masks current_obs = update_current_obs(obs, current_obs, envs.observation_space.shape[0], args.num_stack) rollouts.insert(step, current_obs, states.data, discrete_action.data, discrete_action_log_prob.data, continous_action.data, continous_action_log_prob.data, value.data, scaled_reward, masks) next_value = actor_critic(Variable(rollouts.observations[-1]), Variable(rollouts.states[-1]), Variable(rollouts.masks[-1]))[0].data rollouts.compute_returns(next_value, args.use_gae, args.gamma, args.tau) #Performing Actor Critic Updates if args.algo in ['a2c']: recurrence_steps = 1 if args.recurrent_policy: recurrence_steps = args.num_recsteps indices = torch.arange(0, args.num_steps, recurrence_steps).long() if args.cuda: indices = indices.cuda() total_value_loss = 0 total_action_loss = 0 total_dist_entropy = 0 total_loss = 0 for rstep in range(recurrence_steps): values, action_log_probs, dist_entropy, states = actor_critic.evaluate_actions( Variable(rollouts.observations[:-1].index_select( 0, indices).view(-1, *obs_shape)), Variable(rollouts.states[:-1].index_select( 0, indices).view(-1, actor_critic.state_size)), Variable(rollouts.masks[:-1].index_select(0, indices).view( -1, 1)), [ Variable( rollouts.continuous_actions.index_select( 0, indices).view(-1, action_shape)), Variable( rollouts.discrete_actions.index_select( 0, indices).view(-1, 1)) ]) continuous_dist_entropy, discrete_dist_entropy = dist_entropy continuous_action_log_probs, discrete_action_log_probs = action_log_probs values = values.view(int(args.num_steps / recurrence_steps), args.num_processes, 1) continuous_action_log_probs = continuous_action_log_probs.view( int(args.num_steps / recurrence_steps), args.num_processes, 1) discrete_action_log_probs = discrete_action_log_probs.view( int(args.num_steps / recurrence_steps), args.num_processes, 1) advantages = Variable(rollouts.returns[:-1].index_select( 0, indices)) - values value_loss = advantages.pow(2).mean() '''action_loss = -(Variable(advantages.data) * (0.3*continuous_action_log_probs + 0.7*discrete_action_log_probs)).mean()''' '''action_loss = -(Variable(advantages.data) * ((j%2)*continuous_action_log_probs + ((j+1)%2)*discrete_action_log_probs)).mean()''' action_loss = -(Variable(advantages.data) * continuous_action_log_probs).mean() #loss = value_loss * args.value_loss_coef + action_loss - discrete_dist_entropy * args.entropy_coef loss = value_loss * args.value_loss_coef + action_loss total_value_loss += value_loss.data[0] total_action_loss += action_loss.data[0] #total_dist_entropy += discrete_dist_entropy.data[0] total_loss += loss indices += 1 total_value_loss /= recurrence_steps total_action_loss /= recurrence_steps #total_dist_entropy /= recurrence_steps total_loss /= recurrence_steps optimizer.zero_grad() total_loss.backward() if args.algo == 'a2c': nn.utils.clip_grad_norm(actor_critic.parameters(), args.max_grad_norm) optimizer.step() rollouts.after_update() #Saving the model if j % args.save_interval == 0 and args.save_dir != "": logger.save_model(args.save_dir, actor_critic, envs, args.algo, args.env_name, args.name, args.cuda) #Logging the model if j % args.log_interval == 0: logger.print_log(total_value_loss, total_action_loss, total_dist_entropy, args.num_processes, args.num_steps, j, start)
def main(): print("#######") print( "WARNING: All rewards are clipped so you need to use a monitor (see envs.py) or visdom plot to get true rewards" ) print("#######") os.environ['OMP_NUM_THREADS'] = '1' if args.vis: from visdom import Visdom viz = Visdom() win = None envs = SubprocVecEnv([ make_env(args.env_name, args.seed, i, args.log_dir) for i in range(args.num_processes) ]) obs_shape = envs.observation_space.shape obs_shape = (obs_shape[0] * args.num_stack, *obs_shape[1:]) if len(envs.observation_space.shape) == 3: actor_critic = CNNPolicy(obs_shape[0], envs.action_space) else: actor_critic = MLPPolicy(obs_shape[0], envs.action_space) if envs.action_space.__class__.__name__ == "Discrete": action_shape = 1 else: action_shape = envs.action_space.shape[0] if args.cuda: actor_critic.cuda() if args.algo == 'a2c': optimizer = optim.RMSprop(actor_critic.parameters(), args.lr, eps=args.eps, alpha=args.alpha) elif args.algo == 'ppo': optimizer = optim.Adam(actor_critic.parameters(), args.lr, eps=args.eps) elif args.algo == 'acktr': optimizer = KFACOptimizer(actor_critic) rollouts = RolloutStorage(args.num_steps, args.num_processes, obs_shape, envs.action_space) current_state = torch.zeros(args.num_processes, *obs_shape) def update_current_state(state): shape_dim0 = envs.observation_space.shape[0] state = torch.from_numpy(state).float() if args.num_stack > 1: current_state[:, :-shape_dim0] = current_state[:, shape_dim0:] current_state[:, -shape_dim0:] = state state = envs.reset() update_current_state(state) rollouts.states[0].copy_(current_state) # These variables are used to compute average rewards for all processes. episode_rewards = torch.zeros([args.num_processes, 1]) final_rewards = torch.zeros([args.num_processes, 1]) if args.cuda: current_state = current_state.cuda() rollouts.cuda() if args.algo == 'ppo': old_model = copy.deepcopy(actor_critic) for j in range(num_updates): for step in range(args.num_steps): # Sample actions value, action = actor_critic.act( Variable(rollouts.states[step], volatile=True)) cpu_actions = action.data.squeeze(1).cpu().numpy() # Obser reward and next state state, reward, done, info = envs.step(cpu_actions) reward = torch.from_numpy(np.expand_dims(np.stack(reward), 1)).float() episode_rewards += reward # If done then clean the history of observations. masks = torch.FloatTensor([[0.0] if done_ else [1.0] for done_ in done]) final_rewards *= masks final_rewards += (1 - masks) * episode_rewards episode_rewards *= masks if args.cuda: masks = masks.cuda() if current_state.dim() == 4: current_state *= masks.unsqueeze(2).unsqueeze(2) else: current_state *= masks update_current_state(state) rollouts.insert(step, current_state, action.data, value.data, reward, masks) next_value = actor_critic(Variable(rollouts.states[-1], volatile=True))[0].data if hasattr(actor_critic, 'obs_filter'): actor_critic.obs_filter.update(rollouts.states[:-1].view( -1, *obs_shape)) rollouts.compute_returns(next_value, args.use_gae, args.gamma, args.tau) if args.algo in ['a2c', 'acktr']: values, action_log_probs, dist_entropy = actor_critic.evaluate_actions( Variable(rollouts.states[:-1].view(-1, *obs_shape)), Variable(rollouts.actions.view(-1, action_shape))) values = values.view(args.num_steps, args.num_processes, 1) action_log_probs = action_log_probs.view(args.num_steps, args.num_processes, 1) advantages = Variable(rollouts.returns[:-1]) - values value_loss = advantages.pow(2).mean() action_loss = -(Variable(advantages.data) * action_log_probs).mean() if args.algo == 'acktr' and optimizer.steps % optimizer.Ts == 0: # Sampled fisher, see Martens 2014 actor_critic.zero_grad() pg_fisher_loss = -action_log_probs.mean() value_noise = Variable(torch.randn(values.size())) if args.cuda: value_noise = value_noise.cuda() sample_values = values + value_noise vf_fisher_loss = -(values - Variable(sample_values.data)).pow(2).mean() fisher_loss = pg_fisher_loss + vf_fisher_loss optimizer.acc_stats = True fisher_loss.backward(retain_graph=True) optimizer.acc_stats = False optimizer.zero_grad() (value_loss * args.value_loss_coef + action_loss - dist_entropy * args.entropy_coef).backward() if args.algo == 'a2c': nn.utils.clip_grad_norm(actor_critic.parameters(), args.max_grad_norm) optimizer.step() elif args.algo == 'ppo': advantages = rollouts.returns[:-1] - rollouts.value_preds[:-1] advantages = (advantages - advantages.mean()) / (advantages.std() + 1e-5) old_model.load_state_dict(actor_critic.state_dict()) if hasattr(actor_critic, 'obs_filter'): old_model.obs_filter = actor_critic.obs_filter for _ in range(args.ppo_epoch): sampler = BatchSampler(SubsetRandomSampler( range(args.num_processes * args.num_steps)), args.batch_size * args.num_processes, drop_last=False) for indices in sampler: indices = torch.LongTensor(indices) if args.cuda: indices = indices.cuda() states_batch = rollouts.states[:-1].view( -1, *obs_shape)[indices] actions_batch = rollouts.actions.view( -1, action_shape)[indices] return_batch = rollouts.returns[:-1].view(-1, 1)[indices] # Reshape to do in a single forward pass for all steps values, action_log_probs, dist_entropy = actor_critic.evaluate_actions( Variable(states_batch), Variable(actions_batch)) _, old_action_log_probs, _ = old_model.evaluate_actions( Variable(states_batch, volatile=True), Variable(actions_batch, volatile=True)) ratio = torch.exp(action_log_probs - Variable(old_action_log_probs.data)) adv_targ = Variable(advantages.view(-1, 1)[indices]) surr1 = ratio * adv_targ surr2 = torch.clamp(ratio, 1.0 - args.clip_param, 1.0 + args.clip_param) * adv_targ action_loss = -torch.min( surr1, surr2).mean() # PPO's pessimistic surrogate (L^CLIP) value_loss = (Variable(return_batch) - values).pow(2).mean() optimizer.zero_grad() (value_loss + action_loss - dist_entropy * args.entropy_coef).backward() optimizer.step() rollouts.states[0].copy_(rollouts.states[-1]) if j % args.save_interval == 0 and args.save_dir != "": save_path = os.path.join(args.save_dir, args.algo) try: os.makedirs(save_path) except OSError: pass # A really ugly way to save a model to CPU save_model = actor_critic if args.cuda: save_model = copy.deepcopy(actor_critic).cpu() torch.save(save_model, os.path.join(save_path, args.env_name + ".pt")) if j % args.log_interval == 0: print( "Updates {}, num frames {}, mean/median reward {:.1f}/{:.1f}, min/max reward {:.1f}/{:.1f}, entropy {:.5f}, value loss {:.5f}, policy loss {:.5f}" .format(j, (j + 1) * args.num_processes * args.num_steps, final_rewards.mean(), final_rewards.median(), final_rewards.min(), final_rewards.max(), -dist_entropy.data[0], value_loss.data[0], action_loss.data[0])) if j % args.vis_interval == 0: win = visdom_plot(viz, win, args.log_dir, args.env_name, args.algo)
def main(): print("#######") print( "WARNING: All rewards are clipped or normalized so you need to use a monitor (see envs.py) or visdom plot to get true rewards" ) print("#######") print(args) try: os.makedirs(args.log_dir) except OSError: files = glob.glob(os.path.join(args.log_dir, '*.monitor.csv')) for f in files: os.remove(f) torch.cuda.manual_seed(args.seed) torch.manual_seed(args.seed) np.random.seed(args.seed) random.seed(args.seed) for gamma in args.gamma: with open(args.log_dir + '/MSE_' + str(gamma) + '_monitor.csv', "wt") as monitor_file: monitor = csv.writer(monitor_file) monitor.writerow([ 'update', 'error', str(int(args.num_frames) // args.num_steps) ]) os.environ['OMP_NUM_THREADS'] = '1' print("Using env {}".format(args.env_name)) envs = [ make_env(args.env_name, args.seed, i, args.log_dir) for i in range(args.num_processes) ] if args.num_processes > 1: envs = SubprocVecEnv(envs) else: envs = DummyVecEnv(envs) if len(envs.observation_space.shape) == 1: envs = VecNormalize(envs) obs_shape = envs.observation_space.shape obs_shape = (obs_shape[0] * args.num_stack, *obs_shape[1:]) num_heads = len( args.gamma) if not args.reward_predictor else len(args.gamma) - 1 if len(envs.observation_space.shape) == 3: actor_critic = CNNPolicy(obs_shape[0], envs.action_space, num_heads=num_heads, hidden_size=args.hidden_size) else: actor_critic = MLPPolicy(obs_shape[0], envs.action_space, num_heads=num_heads, reward_predictor=args.reward_predictor, use_s=args.use_s, use_s_a=args.use_s_a, use_s_a_sprime=args.use_s_a_sprime) if envs.action_space.__class__.__name__ == "Discrete": action_shape = 1 else: action_shape = envs.action_space.shape[0] if args.cuda: actor_critic.cuda() lrs = [args.lr] * len(actor_critic.param_groups) if not args.reward_predictor: assert len(actor_critic.param_groups) == len(lrs) model_params = [{ 'params': model_p, 'lr': args.lr } for model_p, lr in zip(actor_critic.param_groups, lrs)] else: model_params = [{ 'params': model_p, 'lr': p_lr } for model_p, p_lr in zip(actor_critic.param_groups[:-1], lrs)] model_params.append({ 'params': actor_critic.param_groups[-1], 'lr': args.lr_rp }) if args.algo == 'a2c': optimizer = optim.RMSprop(model_params, args.lr, eps=args.eps, alpha=args.alpha) elif args.algo == 'ppo': optimizer = optim.Adam(model_params, args.lr, eps=args.eps) rollouts = RolloutStorage(args.num_steps, args.num_processes, obs_shape, envs.action_space, actor_critic.state_size, gamma=args.gamma, use_rp=args.reward_predictor) current_obs = torch.zeros(args.num_processes, *obs_shape) def update_current_obs(obs, obs_tensor): shape_dim0 = envs.observation_space.shape[0] obs = torch.from_numpy(obs).float() if args.num_stack > 1: obs_tensor[:, :-shape_dim0] = obs_tensor[:, shape_dim0:] obs_tensor[:, -shape_dim0:] = obs obs = envs.reset() update_current_obs(obs, current_obs) rollouts.observations[0].copy_(current_obs) # These variables are used to compute average rewards for all processes. episode_rewards = torch.zeros([args.num_processes, 1]) final_rewards = torch.zeros([args.num_processes, 1]) advantages_list = [] if args.cuda: current_obs = current_obs.cuda() rollouts.cuda() start = time.time() for j in range(num_updates): for step in range(args.num_steps): # Sample actions value, action, action_log_prob, states = actor_critic.act( Variable(rollouts.observations[step], volatile=True), Variable(rollouts.states[step], volatile=True), Variable(rollouts.masks[step], volatile=True)) cpu_actions = action.data.squeeze(1).cpu().numpy() cpu_actions = add_gaussian_noise(cpu_actions, args.action_noise) # Obser reward and next obs obs, raw_reward, done, info = envs.step(cpu_actions) reward = np.copy(raw_reward) reward = add_gaussian_noise(reward, args.reward_noise) reward = epsilon_greedy(reward, args.reward_epsilon, args.reward_high, args.reward_low) raw_reward = torch.from_numpy( np.expand_dims(np.stack(raw_reward), 1)).float() episode_rewards += raw_reward reward = torch.from_numpy(np.expand_dims(np.stack(reward), 1)).float() # If done then clean the history of observations. masks = torch.FloatTensor([[0.0] if done_ else [1.0] for done_ in done]) final_rewards *= masks final_rewards += (1 - masks) * episode_rewards episode_rewards *= masks if args.cuda: masks = masks.cuda() if current_obs.dim() == 4: current_obs *= masks.unsqueeze(2).unsqueeze(2) else: current_obs *= masks update_current_obs(obs, current_obs) if args.reward_predictor: r_hat = actor_critic.predict_reward( Variable(rollouts.observations[step], volatile=True), action, Variable(current_obs, volatile=True)) p_hat = min(args.rp_burn_in, j) / args.rp_burn_in estimate_reward = (1 - p_hat) * reward + p_hat * r_hat.data.cpu() reward = torch.cat([reward, estimate_reward], dim=-1) value = torch.cat([r_hat, value], dim=-1).data else: value = value.data rollouts.insert(step, current_obs, states.data, action.data, action_log_prob.data, value, reward, masks, raw_reward) next_value = actor_critic( Variable(rollouts.observations[-1], volatile=True), Variable(rollouts.states[-1], volatile=True), Variable(rollouts.masks[-1], volatile=True))[0].data if args.reward_predictor: if args.use_s or args.use_s_a: r_hat = actor_critic.predict_reward( Variable(rollouts.observations[-1], volatile=True), Variable(rollouts.actions[-1], volatile=True), None).data next_value = torch.cat([r_hat, next_value], dim=-1) else: next_value = torch.cat([ torch.zeros(list(next_value.size())[:-1] + [1]), next_value ], dim=-1) rollouts.compute_returns(next_value, args.use_gae, args.tau) if args.algo in ['a2c']: batch_states = Variable(rollouts.states[0].view( -1, actor_critic.state_size)) batch_masks = Variable(rollouts.masks[:-1].view(-1, 1)) batch_obs = Variable(rollouts.observations[:-1].view( -1, *obs_shape)) batch_actions = Variable(rollouts.actions.view(-1, action_shape)) values, action_log_probs, dist_entropy, states = actor_critic.evaluate_actions( batch_obs, batch_states, batch_masks, batch_actions) if args.reward_predictor: batch_obs_prime = Variable(rollouts.observations[1:].view( -1, *obs_shape)) values = torch.cat([ actor_critic.predict_reward(batch_obs, batch_actions, batch_obs_prime), values ], dim=-1) returns_as_variable = Variable(rollouts.returns[:-1]) batched_v_loss = 0 values = values.view(returns_as_variable.size()) action_log_probs = action_log_probs.view(args.num_steps, args.num_processes, 1) advantages = returns_as_variable - values value_loss = advantages.pow(2).sum(-1).mean() action_loss = -(Variable(advantages[:, :, -1].unsqueeze(-1).data) * action_log_probs).mean() if args.reward_predictor: rp_error = (values[:, :, 0].data - rollouts.raw_rewards).pow(2).mean() advantages_list.append([ rp_error, advantages[:, :, -1].pow(2).mean().data.cpu().numpy()[0] ]) else: advantages_list.append( advantages[:, :, -1].pow(2).mean().data.cpu().numpy()[0]) optimizer.zero_grad() (batched_v_loss + value_loss * args.value_loss_coef + action_loss - dist_entropy * args.entropy_coef).backward() if args.algo == 'a2c': nn.utils.clip_grad_norm(actor_critic.parameters(), args.max_grad_norm) optimizer.step() elif args.algo == 'ppo': advantages = rollouts.returns[:-1] - rollouts.value_preds[:-1] advantages = advantages[:, :, -1] advantages = (advantages - advantages.mean()) / (advantages.std() + 1e-5) for e in range(args.ppo_epoch): data_generator = rollouts.feed_forward_generator( advantages, args.num_mini_batch) for sample in data_generator: observations_batch, states_batch, actions_batch, \ return_batch, masks_batch, old_action_log_probs_batch, \ adv_targ, observations_batch_prime, true_rewards_batch, \ noisy_observations_batch, true_observations_batch = sample # Reshape to do in a single forward pass for all steps values, action_log_probs, dist_entropy, states = actor_critic.evaluate_actions( Variable(observations_batch), Variable(states_batch), Variable(masks_batch), Variable(actions_batch)) if args.reward_predictor: values = torch.cat([ actor_critic.predict_reward( Variable(observations_batch), Variable(actions_batch), Variable(observations_batch_prime)), values ], dim=-1) adv_targ = Variable(adv_targ) ratio = torch.exp(action_log_probs - Variable(old_action_log_probs_batch)) surr1 = ratio * adv_targ surr2 = torch.clamp(ratio, 1.0 - args.clip_param, 1.0 + args.clip_param) * adv_targ action_loss = -torch.min( surr1, surr2).mean() # PPO's pessimistic surrogate (L^CLIP) td = (Variable(return_batch) - values).pow(2) value_loss = td.sum(-1).mean() if args.reward_predictor: rp_error = (values[:, 0].data - true_rewards_batch).pow(2).mean() advantages_list.append( [rp_error, td[:, -1].mean(0).data.cpu().numpy()]) else: advantages_list.append( td[:, -1].mean(0).data.cpu().numpy()) optimizer.zero_grad() (value_loss + action_loss - dist_entropy * args.entropy_coef).backward() nn.utils.clip_grad_norm(actor_critic.parameters(), args.max_grad_norm) optimizer.step() rollouts.after_update() if j % args.save_interval == 0 and args.save_dir != "": save_path = os.path.join(args.save_dir, args.algo) try: os.makedirs(save_path) except OSError: pass # A really ugly way to save a model to CPU save_model = actor_critic if args.cuda: save_model = copy.deepcopy(actor_critic).cpu() save_model = [ save_model, hasattr(envs, 'ob_rms') and envs.ob_rms or None ] torch.save(save_model, os.path.join(save_path, args.env_name + ".pt")) if j % args.log_interval == 0: end = time.time() total_num_steps = (j + 1) * args.num_processes * args.num_steps print( "Updates {}, num timesteps {}, FPS {}, mean/median reward {:.1f}/{:.1f}, min/max reward {:.1f}/{:.1f}, " "entropy {:.5f}, value loss {:.5f}, policy loss {:.5f}".format( j, total_num_steps, int(total_num_steps / (end - start)), final_rewards.mean(), final_rewards.median(), final_rewards.min(), final_rewards.max(), dist_entropy.data[0], value_loss.data[0], action_loss.data[0])) if len(advantages_list) > 2: advantages_array = np.array(advantages_list).reshape( -1, len(args.gamma)).T for g, gamma in enumerate(args.gamma): with open( args.log_dir + '/MSE_' + str(gamma) + '_monitor.csv', "a") as monitor_file: monitor = csv.writer(monitor_file) monitor.writerow( [total_num_steps, np.mean(advantages_array[g])]) advantages_list = []
def main(): print("#######") print("WARNING: All rewards are clipped so you need to use a monitor (see envs.py) or visdom plot to get true rewards") print("#######") os.environ['OMP_NUM_THREADS'] = '1' print (args.cuda) print (args.num_steps) print (args.num_processes) print (args.lr) print (args.eps) print (args.alpha) print (args.use_gae) print (args.gamma) print (args.tau) print (args.value_loss_coef) print (args.entropy_coef) fdsafasd # if args.vis: # from visdom import Visdom # viz = Visdom() # win = None envs = SubprocVecEnv([ make_env(args.env_name, args.seed, i, args.log_dir) for i in range(args.num_processes) ]) # print('here3') # fdasf obs_shape = envs.observation_space.shape obs_shape = (obs_shape[0] * args.num_stack, *obs_shape[1:]) if len(envs.observation_space.shape) == 3: actor_critic = CNNPolicy(obs_shape[0], envs.action_space) else: actor_critic = MLPPolicy(obs_shape[0], envs.action_space) if envs.action_space.__class__.__name__ == "Discrete": action_shape = 1 else: action_shape = envs.action_space.shape[0] if args.cuda: actor_critic.cuda() if args.algo == 'a2c': optimizer = optim.RMSprop(actor_critic.parameters(), args.lr, eps=args.eps, alpha=args.alpha) # elif args.algo == 'ppo': # optimizer = optim.Adam(actor_critic.parameters(), args.lr, eps=args.eps) # elif args.algo == 'acktr': # optimizer = KFACOptimizer(actor_critic) rollouts = RolloutStorage(args.num_steps, args.num_processes, obs_shape, envs.action_space) #it has a self.state that is [steps, processes, obs] #steps is used to compute expected reward current_state = torch.zeros(args.num_processes, *obs_shape) def update_current_state(state): shape_dim0 = envs.observation_space.shape[0] state = torch.from_numpy(state).float() if args.num_stack > 1: current_state[:, :-shape_dim0] = current_state[:, shape_dim0:] current_state[:, -shape_dim0:] = state state = envs.reset() update_current_state(state) rollouts.states[0].copy_(current_state) #set the first state to current state # These variables are used to compute average rewards for all processes. episode_rewards = torch.zeros([args.num_processes, 1]) final_rewards = torch.zeros([args.num_processes, 1]) if args.cuda: current_state = current_state.cuda() rollouts.cuda() # if args.algo == 'ppo': # old_model = copy.deepcopy(actor_critic) start = time.time() for j in range(num_updates): for step in range(args.num_steps): # Sample actions value, action = actor_critic.act(Variable(rollouts.states[step], volatile=True)) # make prediction using state that you put into rollouts cpu_actions = action.data.squeeze(1).cpu().numpy() # Obser reward and next state state, reward, done, info = envs.step(cpu_actions) # print (state.shape) # [nProcesss, ndims, height, width] # fsdf reward = torch.from_numpy(np.expand_dims(np.stack(reward), 1)).float() episode_rewards += reward # If done then clean the history of observations. # these final rewards are only used for printing. but the mask is used in the storage, dont know why yet # oh its just clearing the env that finished, and resetting its episode_reward masks = torch.FloatTensor([[0.0] if done_ else [1.0] for done_ in done]) #if an env is done final_rewards *= masks final_rewards += (1 - masks) * episode_rewards episode_rewards *= masks if args.cuda: masks = masks.cuda() if current_state.dim() == 4: current_state *= masks.unsqueeze(2).unsqueeze(2) else: current_state *= masks update_current_state(state) rollouts.insert(step, current_state, action.data, value.data, reward, masks) # insert all that info into current step # not exactly why next_value = actor_critic(Variable(rollouts.states[-1], volatile=True))[0].data # use last state to make prediction of next value if hasattr(actor_critic, 'obs_filter'): actor_critic.obs_filter.update(rollouts.states[:-1].view(-1, *obs_shape)) #not sure what this is rollouts.compute_returns(next_value, args.use_gae, args.gamma, args.tau) # this computes R = r + r+ ...+ V(t) for each step if args.algo in ['a2c', 'acktr']: values, action_log_probs, dist_entropy = actor_critic.evaluate_actions(Variable(rollouts.states[:-1].view(-1, *obs_shape)), Variable(rollouts.actions.view(-1, action_shape))) # I think this aciton log prob could have been computed and stored earlier # and didnt we already store the value prediction??? values = values.view(args.num_steps, args.num_processes, 1) action_log_probs = action_log_probs.view(args.num_steps, args.num_processes, 1) advantages = Variable(rollouts.returns[:-1]) - values value_loss = advantages.pow(2).mean() action_loss = -(Variable(advantages.data) * action_log_probs).mean() # if args.algo == 'acktr' and optimizer.steps % optimizer.Ts == 0: # # Sampled fisher, see Martens 2014 # actor_critic.zero_grad() # pg_fisher_loss = -action_log_probs.mean() # value_noise = Variable(torch.randn(values.size())) # if args.cuda: # value_noise = value_noise.cuda() # sample_values = values + value_noise # vf_fisher_loss = -(values - Variable(sample_values.data)).pow(2).mean() # fisher_loss = pg_fisher_loss + vf_fisher_loss # optimizer.acc_stats = True # fisher_loss.backward(retain_graph=True) # optimizer.acc_stats = False optimizer.zero_grad() (value_loss * args.value_loss_coef + action_loss - dist_entropy * args.entropy_coef).backward() # if args.algo == 'a2c': # nn.utils.clip_grad_norm(actor_critic.parameters(), args.max_grad_norm) optimizer.step() # elif args.algo == 'ppo': # advantages = rollouts.returns[:-1] - rollouts.value_preds[:-1] # advantages = (advantages - advantages.mean()) / (advantages.std() + 1e-5) # old_model.load_state_dict(actor_critic.state_dict()) # if hasattr(actor_critic, 'obs_filter'): # old_model.obs_filter = actor_critic.obs_filter # for _ in range(args.ppo_epoch): # sampler = BatchSampler(SubsetRandomSampler(range(args.num_processes * args.num_steps)), args.batch_size * args.num_processes, drop_last=False) # for indices in sampler: # indices = torch.LongTensor(indices) # if args.cuda: # indices = indices.cuda() # states_batch = rollouts.states[:-1].view(-1, *obs_shape)[indices] # actions_batch = rollouts.actions.view(-1, action_shape)[indices] # return_batch = rollouts.returns[:-1].view(-1, 1)[indices] # # Reshape to do in a single forward pass for all steps # values, action_log_probs, dist_entropy = actor_critic.evaluate_actions(Variable(states_batch), Variable(actions_batch)) # _, old_action_log_probs, _ = old_model.evaluate_actions(Variable(states_batch, volatile=True), Variable(actions_batch, volatile=True)) # ratio = torch.exp(action_log_probs - Variable(old_action_log_probs.data)) # adv_targ = Variable(advantages.view(-1, 1)[indices]) # surr1 = ratio * adv_targ # surr2 = torch.clamp(ratio, 1.0 - args.clip_param, 1.0 + args.clip_param) * adv_targ # action_loss = -torch.min(surr1, surr2).mean() # PPO's pessimistic surrogate (L^CLIP) # value_loss = (Variable(return_batch) - values).pow(2).mean() # optimizer.zero_grad() # (value_loss + action_loss - dist_entropy * args.entropy_coef).backward() # optimizer.step() rollouts.states[0].copy_(rollouts.states[-1]) # the first state is now the last state of the previous # if j % args.save_interval == 0 and args.save_dir != "": # save_path = os.path.join(args.save_dir, args.algo) # try: # os.makedirs(save_path) # except OSError: # pass # # A really ugly way to save a model to CPU # save_model = actor_critic # if args.cuda: # save_model = copy.deepcopy(actor_critic).cpu() # torch.save(save_model, os.path.join(save_path, args.env_name + ".pt")) if j % args.log_interval == 0: end = time.time() total_num_steps = (j + 1) * args.num_processes * args.num_steps print("Updates {}, num timesteps {}, FPS {}, mean/median reward {:.1f}/{:.1f}, min/max reward {:.1f}/{:.1f}, entropy {:.5f}, value loss {:.5f}, policy loss {:.5f}". format(j, total_num_steps, int(total_num_steps / (end - start)), final_rewards.mean(), final_rewards.median(), final_rewards.min(), final_rewards.max(), -dist_entropy.data[0], value_loss.data[0], action_loss.data[0]))
def main(): print("#######") print( "WARNING: All rewards are clipped or normalized so you need to use a monitor (see envs.py) or visdom plot to get true rewards" ) print("#######") os.environ['OMP_NUM_THREADS'] = '1' if args.vis: from visdom import Visdom viz = Visdom() win = None envs = [ make_env(args.env_name, args.seed, i, args.log_dir) for i in range(args.num_processes) ] if args.num_processes > 1: envs = SubprocVecEnv(envs) else: envs = DummyVecEnv(envs) if len(envs.observation_space.shape) == 1: envs = VecNormalize(envs) obs_shape = envs.observation_space.shape obs_shape = (obs_shape[0] * args.num_stack, *obs_shape[1:]) obs_numel = reduce(operator.mul, obs_shape, 1) if len(obs_shape) == 3 and obs_numel > 1024: actor_critic = CNNPolicy(obs_shape[0], envs.action_space, args.recurrent_policy) else: assert not args.recurrent_policy, \ "Recurrent policy is not implemented for the MLP controller" actor_critic = MLPPolicy(obs_numel, envs.action_space) if envs.action_space.__class__.__name__ == "Discrete": action_shape = 1 else: action_shape = envs.action_space.shape[0] if args.cuda: actor_critic.cuda() if args.algo == 'a2c': optimizer = optim.RMSprop(actor_critic.parameters(), args.lr, eps=args.eps, alpha=args.alpha) elif args.algo == 'ppo': optimizer = optim.Adam(actor_critic.parameters(), args.lr, eps=args.eps) elif args.algo == 'acktr': optimizer = KFACOptimizer(actor_critic) rollouts = RolloutStorage(args.num_steps, args.num_processes, obs_shape, envs.action_space, actor_critic.state_size) current_obs = torch.zeros(args.num_processes, *obs_shape) def update_current_obs(obs): shape_dim0 = envs.observation_space.shape[0] obs = torch.from_numpy(obs).float() if args.num_stack > 1: current_obs[:, :-shape_dim0] = current_obs[:, shape_dim0:] current_obs[:, -shape_dim0:] = obs obs = envs.reset() update_current_obs(obs) rollouts.observations[0].copy_(current_obs) # These variables are used to compute average rewards for all processes. episode_rewards = torch.zeros([args.num_processes, 1]) final_rewards = torch.zeros([args.num_processes, 1]) if args.cuda: current_obs = current_obs.cuda() rollouts.cuda() start = time.time() for j in range(num_updates): for step in range(args.num_steps): # Sample actions value, action, action_log_prob, states = actor_critic.act( Variable(rollouts.observations[step], volatile=True), Variable(rollouts.states[step], volatile=True), Variable(rollouts.masks[step], volatile=True)) cpu_actions = action.data.squeeze(1).cpu().numpy() # Obser reward and next obs obs, reward, done, info = envs.step(cpu_actions) reward = torch.from_numpy(np.expand_dims(np.stack(reward), 1)).float() episode_rewards += reward # If done then clean the history of observations. masks = torch.FloatTensor([[0.0] if done_ else [1.0] for done_ in done]) final_rewards *= masks final_rewards += (1 - masks) * episode_rewards episode_rewards *= masks if args.cuda: masks = masks.cuda() if current_obs.dim() == 4: current_obs *= masks.unsqueeze(2).unsqueeze(2) else: current_obs *= masks update_current_obs(obs) rollouts.insert(step, current_obs, states.data, action.data, action_log_prob.data, value.data, reward, masks) next_value = actor_critic( Variable(rollouts.observations[-1], volatile=True), Variable(rollouts.states[-1], volatile=True), Variable(rollouts.masks[-1], volatile=True))[0].data rollouts.compute_returns(next_value, args.use_gae, args.gamma, args.tau) if args.algo in ['a2c', 'acktr']: values, action_log_probs, dist_entropy, states = actor_critic.evaluate_actions( Variable(rollouts.observations[:-1].view(-1, *obs_shape)), Variable(rollouts.states[0].view(-1, actor_critic.state_size)), Variable(rollouts.masks[:-1].view(-1, 1)), Variable(rollouts.actions.view(-1, action_shape))) values = values.view(args.num_steps, args.num_processes, 1) action_log_probs = action_log_probs.view(args.num_steps, args.num_processes, 1) advantages = Variable(rollouts.returns[:-1]) - values value_loss = advantages.pow(2).mean() action_loss = -(Variable(advantages.data) * action_log_probs).mean() if args.algo == 'acktr' and optimizer.steps % optimizer.Ts == 0: # Sampled fisher, see Martens 2014 actor_critic.zero_grad() pg_fisher_loss = -action_log_probs.mean() value_noise = Variable(torch.randn(values.size())) if args.cuda: value_noise = value_noise.cuda() sample_values = values + value_noise vf_fisher_loss = -(values - Variable(sample_values.data)).pow(2).mean() fisher_loss = pg_fisher_loss + vf_fisher_loss optimizer.acc_stats = True fisher_loss.backward(retain_graph=True) optimizer.acc_stats = False optimizer.zero_grad() (value_loss * args.value_loss_coef + action_loss - dist_entropy * args.entropy_coef).backward() if args.algo == 'a2c': nn.utils.clip_grad_norm(actor_critic.parameters(), args.max_grad_norm) optimizer.step() elif args.algo == 'ppo': advantages = rollouts.returns[:-1] - rollouts.value_preds[:-1] advantages = (advantages - advantages.mean()) / (advantages.std() + 1e-5) for e in range(args.ppo_epoch): if args.recurrent_policy: data_generator = rollouts.recurrent_generator( advantages, args.num_mini_batch) else: data_generator = rollouts.feed_forward_generator( advantages, args.num_mini_batch) for sample in data_generator: observations_batch, states_batch, actions_batch, \ return_batch, masks_batch, old_action_log_probs_batch, \ adv_targ = sample # Reshape to do in a single forward pass for all steps values, action_log_probs, dist_entropy, states = actor_critic.evaluate_actions( Variable(observations_batch), Variable(states_batch), Variable(masks_batch), Variable(actions_batch)) adv_targ = Variable(adv_targ) ratio = torch.exp(action_log_probs - Variable(old_action_log_probs_batch)) surr1 = ratio * adv_targ surr2 = torch.clamp(ratio, 1.0 - args.clip_param, 1.0 + args.clip_param) * adv_targ action_loss = -torch.min( surr1, surr2).mean() # PPO's pessimistic surrogate (L^CLIP) value_loss = (Variable(return_batch) - values).pow(2).mean() optimizer.zero_grad() (value_loss + action_loss - dist_entropy * args.entropy_coef).backward() nn.utils.clip_grad_norm(actor_critic.parameters(), args.max_grad_norm) optimizer.step() rollouts.after_update() if j % args.save_interval == 0 and args.save_dir != "": save_path = os.path.join(args.save_dir, args.algo) try: os.makedirs(save_path) except OSError: pass # A really ugly way to save a model to CPU save_model = actor_critic if args.cuda: save_model = copy.deepcopy(actor_critic).cpu() save_model = [ save_model, hasattr(envs, 'ob_rms') and envs.ob_rms or None ] torch.save(save_model, os.path.join(save_path, args.env_name + ".pt")) if j % args.log_interval == 0: end = time.time() total_num_steps = (j + 1) * args.num_processes * args.num_steps print( "Updates {}, num timesteps {}, FPS {}, mean/median reward {:.1f}/{:.1f}, min/max reward {:.1f}/{:.1f}, entropy {:.5f}, value loss {:.5f}, policy loss {:.5f}" .format(j, total_num_steps, int(total_num_steps / (end - start)), final_rewards.mean(), final_rewards.median(), final_rewards.min(), final_rewards.max(), dist_entropy.data[0], value_loss.data[0], action_loss.data[0])) if args.vis and j % args.vis_interval == 0: try: # Sometimes monitor doesn't properly flush the outputs win = visdom_plot(viz, win, args.log_dir, args.env_name, args.algo) except IOError: pass
def main(): print("#######") print( "WARNING: All rewards are clipped so you need to use a monitor (see envs.py) or visdom plot to get true rewards" ) print("#######") os.environ['OMP_NUM_THREADS'] = '1' # T choose whetehr to visualize if args.vis: from visdom import Visdom viz = Visdom() win = None envs = SubprocVecEnv([ make_env(args.env_name, args.seed, i, args.log_dir) for i in range(args.num_processes) ]) # T get shape of observation array of the environment obs_shape = envs.observation_space.shape # T adjusting the shape; not sure what the * is obs_shape = (obs_shape[0] * args.num_stack, *obs_shape[1:]) #T initialize the actor critic; MLP and CNN classes imported from model.py if len(envs.observation_space.shape) == 3: actor_critic = CNNPolicy(obs_shape[0], envs.action_space) else: actor_critic = MLPPolicy(obs_shape[0], envs.action_space) #T - some kind of setup with the actor_critic if args.finetune: checkpoint_path = save_path = os.path.join(args.save_dir, args.algo, args.checkpoint) state_dict = torch.load(checkpoint_path) print("Finetuning from checkpoint: %s, at step: %d" % (checkpoint_path, state_dict['update'])) actor_critic.load_state_dict(state_dict['model_state_dict']) keep_layers = [ 'v_fc3.weight', 'v_fc3.bias', 'a_fc2.weight', 'a_fc2.bias', 'dist.fc_mean.weight', 'dist.fc_mean.bias', 'dist.logstd._bias' ] for name, param in actor_critic.named_parameters(): if name not in keep_layers: param.requires_grad = False for name, param in actor_critic.named_parameters(): print('Param name: %s, requires_grad: %d' % (name, param.requires_grad)) # T set up dimensions of the action space if envs.action_space.__class__.__name__ == "Discrete": action_shape = 1 else: action_shape = envs.action_space.shape[0] # T all arguments imported from arguments.py # T enable cuda pythorch tensor support if args.cuda: actor_critic.cuda() # T - pull arguments and choose algorithm and optimizer if args.algo == 'a2c': optimizer = optim.RMSprop(filter(lambda p: p.requires_grad, actor_critic.parameters()), args.lr, eps=args.eps, alpha=args.alpha) elif args.algo == 'ppo': optimizer = optim.Adam(actor_critic.parameters(), args.lr, eps=args.eps) elif args.algo == 'acktr': optimizer = KFACOptimizer(actor_critic) #TO-DO figure out how to restore optimizer parameters when freezing some weights rollouts = RolloutStorage(args.num_steps, args.num_processes, obs_shape, envs.action_space) # return all zeros, so nothing observed current_obs = torch.zeros(args.num_processes, *obs_shape) # T-not sure what this function is doing?? def update_current_obs(obs): shape_dim0 = envs.observation_space.shape[0] obs = torch.from_numpy(obs).float() if args.num_stack > 1: current_obs[:, :-shape_dim0] = current_obs[:, shape_dim0:] current_obs[:, -shape_dim0:] = obs # T - reset the environment; call function to update observation obs = envs.reset() update_current_obs(obs) rollouts.observations[0].copy_(current_obs) # These variables are used to compute average rewards for all processes. # T - initialize rewards to be zero episode_rewards = torch.zeros([args.num_processes, 1]) final_rewards = torch.zeros([args.num_processes, 1]) if args.cuda: current_obs = current_obs.cuda() rollouts.cuda() if args.algo == 'ppo': old_model = copy.deepcopy(actor_critic) start = time.time() # T - begin iterative loop for j in range(num_updates): # T - take steps through single instance # T - this is the loop where action/critic happens for step in range(args.num_steps): # Sample actions # T - buried by the action method ultimately comes from torch.nn.Module value, action = actor_critic.act( Variable(rollouts.observations[step], volatile=True)) cpu_actions = action.data.squeeze(1).cpu().numpy() # Obser reward and next obs obs, reward, done, info = envs.step(cpu_actions) reward = torch.from_numpy(np.expand_dims(np.stack(reward), 1)).float() episode_rewards += reward # T done bool returned by steps; indicates if failure occurred (done) # If done then clean the history of observations. masks = torch.FloatTensor([[0.0] if done_ else [1.0] for done_ in done]) final_rewards *= masks final_rewards += (1 - masks) * episode_rewards episode_rewards *= masks if args.cuda: masks = masks.cuda() if current_obs.dim() == 4: current_obs *= masks.unsqueeze(2).unsqueeze(2) else: current_obs *= masks #T - now update the observation matrix update_current_obs(obs) #T - store what happened in this step rollouts.insert(step, current_obs, action.data, value.data, reward, masks) next_value = actor_critic( Variable(rollouts.observations[-1], volatile=True))[0].data if hasattr(actor_critic, 'obs_filter'): actor_critic.obs_filter.update(rollouts.observations[:-1].view( -1, *obs_shape)) rollouts.compute_returns(next_value, args.use_gae, args.gamma, args.tau) if args.algo in ['a2c', 'acktr']: values, action_log_probs, dist_entropy = actor_critic.evaluate_actions( Variable(rollouts.observations[:-1].view(-1, *obs_shape)), Variable(rollouts.actions.view(-1, action_shape))) values = values.view(args.num_steps, args.num_processes, 1) action_log_probs = action_log_probs.view(args.num_steps, args.num_processes, 1) advantages = Variable(rollouts.returns[:-1]) - values value_loss = advantages.pow(2).mean() action_loss = -(Variable(advantages.data) * action_log_probs).mean() if args.algo == 'acktr' and optimizer.steps % optimizer.Ts == 0: # Sampled fisher, see Martens 2014 actor_critic.zero_grad() pg_fisher_loss = -action_log_probs.mean() value_noise = Variable(torch.randn(values.size())) if args.cuda: value_noise = value_noise.cuda() sample_values = values + value_noise vf_fisher_loss = -(values - Variable(sample_values.data)).pow(2).mean() fisher_loss = pg_fisher_loss + vf_fisher_loss optimizer.acc_stats = True fisher_loss.backward(retain_graph=True) optimizer.acc_stats = False optimizer.zero_grad() (value_loss * args.value_loss_coef + action_loss - dist_entropy * args.entropy_coef).backward() if args.algo == 'a2c': nn.utils.clip_grad_norm(actor_critic.parameters(), args.max_grad_norm) optimizer.step() elif args.algo == 'ppo': advantages = rollouts.returns[:-1] - rollouts.value_preds[:-1] advantages = (advantages - advantages.mean()) / (advantages.std() + 1e-5) old_model.load_state_dict(actor_critic.state_dict()) for _ in range(args.ppo_epoch): sampler = BatchSampler(SubsetRandomSampler( range(args.num_processes * args.num_steps)), args.batch_size * args.num_processes, drop_last=False) for indices in sampler: indices = torch.LongTensor(indices) if args.cuda: indices = indices.cuda() observations_batch = rollouts.observations[:-1].view( -1, *obs_shape)[indices] actions_batch = rollouts.actions.view( -1, action_shape)[indices] return_batch = rollouts.returns[:-1].view(-1, 1)[indices] # Reshape to do in a single forward pass for all steps values, action_log_probs, dist_entropy = actor_critic.evaluate_actions( Variable(observations_batch), Variable(actions_batch)) _, old_action_log_probs, _ = old_model.evaluate_actions( Variable(observations_batch, volatile=True), Variable(actions_batch, volatile=True)) ratio = torch.exp(action_log_probs - Variable(old_action_log_probs.data)) adv_targ = Variable(advantages.view(-1, 1)[indices]) surr1 = ratio * adv_targ surr2 = torch.clamp(ratio, 1.0 - args.clip_param, 1.0 + args.clip_param) * adv_targ action_loss = -torch.min( surr1, surr2).mean() # PPO's pessimistic surrogate (L^CLIP) value_loss = (Variable(return_batch) - values).pow(2).mean() optimizer.zero_grad() (value_loss + action_loss - dist_entropy * args.entropy_coef).backward() optimizer.step() rollouts.observations[0].copy_(rollouts.observations[-1]) if j % args.save_interval == 0 and args.save_dir != "": save_path = os.path.join(args.save_dir, args.algo) try: os.makedirs(save_path) except OSError: pass # A really ugly way to save a model to CPU save_model = actor_critic if args.cuda: save_model = copy.deepcopy(actor_critic).cpu() file_name = FILE_PREFIX + '.pt' #torch.save(save_model, os.path.join(save_path, file_name)) data = { 'update': j, 'model_state_dict': save_model.state_dict(), 'optim_state_dict': optimizer.state_dict() } torch.save(data, os.path.join(save_path, file_name)) # T - write out some log information (not important for us) if j % args.log_interval == 0: end = time.time() total_num_steps = (j + 1) * args.num_processes * args.num_steps print( "Updates {}, num timesteps {}, FPS {}, mean/median reward {:.1f}/{:.1f}, min/max reward {:.1f}/{:.1f}, entropy {:.5f}, value loss {:.5f}, policy loss {:.5f}" .format(j, total_num_steps, int(total_num_steps / (end - start)), final_rewards.mean(), final_rewards.median(), final_rewards.min(), final_rewards.max(), -dist_entropy.data[0], value_loss.data[0], action_loss.data[0])) if args.vis and j % args.vis_interval == 0: try: # Sometimes monitor doesn't properly flush the outputs win = visdom_plot(viz, win, args.log_dir, args.env_name, args.algo) except IOError: pass
def main(): print("###############################################################") print("#################### VIZDOOM LEARNER START ####################") print("###############################################################") save_path = os.path.join(args.save_dir, "a2c") num_updates = int(args.num_frames) // args.num_steps // args.num_processes reward_name = "" if args.roe: reward_name = "_event" scenario_name = args.config_path.split("/")[1].split(".")[0] print("############### " + scenario_name + " ###############") log_file_name = "log/vizdoom_" + scenario_name + reward_name + "_" + str( args.agent_id) + ".log" log_event_file_name = "log/vizdoom_" + scenario_name + reward_name + "_" + str( args.agent_id) + ".eventlog" log_event_reward_file_name = "log/vizdoom_" + scenario_name + reward_name + "_" + str( args.agent_id) + ".eventrewardlog" start_updates = 0 start_step = 0 best_final_rewards = -1000000.0 os.environ['OMP_NUM_THREADS'] = '1' # ---- DOOM Environments # Host: game = DoomGame() # Use CIG example config or your own. game.load_config("../../scenarios/cig.cfg") game.set_doom_map("map01") # Limited deathmatch. #game.set_doom_map("map02") # Full deathmatch. # Host game with options that will be used in the competition. game.add_game_args( "-host 2 " # This machine will function as a host for a multiplayer game with this many players (including this machine). It will wait for other machines to connect using the -join parameter and then start the game when everyone is connected. "-deathmatch " # Deathmatch rules are used for the game. "+timelimit 10.0 " # The game (episode) will end after this many minutes have elapsed. "+sv_forcerespawn 1 " # Players will respawn automatically after they die. "+sv_noautoaim 1 " # Autoaim is disabled for all players. "+sv_respawnprotect 1 " # Players will be invulnerable for two second after spawning. "+sv_spawnfarthest 1 " # Players will be spawned as far as possible from any other players. "+sv_nocrouch 1 " # Disables crouching. "+viz_respawn_delay 10 " # Sets delay between respanws (in seconds). "+viz_nocheat 1" ) # Disables depth and labels buffer and the ability to use commands that could interfere with multiplayer game. # This can be used to host game without taking part in it (can be simply added as argument of vizdoom executable). game.add_game_args("viz_spectator 1") # During the competition, async mode will be forced for all agents. game.set_mode(Mode.ASYNC_PLAYER) # Start game game.init() # Clients: global envs es = [ make_cig_env(i, visual=args.visual) for i in range(args.num_processes) ] envs = VecEnv([es[i] for i in range(args.num_processes)]) obs_shape = envs.observation_space_shape obs_shape = (obs_shape[0] * args.num_stack, *obs_shape[1:]) if args.resume: actor_critic = torch.load( os.path.join(save_path, log_file_name + ".pt")) filename = glob.glob(os.path.join(args.log_dir, log_file_name))[0] #if args.roe: # TODO: Load event buffer with open(filename) as file: lines = file.readlines() start_updates = (int)(lines[-1].strip().split(",")[0]) start_steps = (int)(lines[-1].strip().split(",")[1]) num_updates += start_updates else: if not args.debug: try: os.makedirs(args.log_dir) except OSError: files = glob.glob(os.path.join(args.log_dir, log_file_name)) for f in files: os.remove(f) with open(log_file_name, "w") as myfile: myfile.write("") files = glob.glob( os.path.join(args.log_dir, log_event_file_name)) for f in files: os.remove(f) with open(log_event_file_name, "w") as myfile: myfile.write("") files = glob.glob( os.path.join(args.log_dir, log_event_reward_file_name)) for f in files: os.remove(f) with open(log_event_reward_file_name, "w") as myfile: myfile.write("") actor_critic = CNNPolicy(obs_shape[0], envs.action_space_shape) action_shape = 1 if args.cuda: actor_critic.cuda() optimizer = optim.RMSprop(actor_critic.parameters(), args.lr, eps=args.eps, alpha=args.alpha) rollouts = RolloutStorage(args.num_steps, args.num_processes, obs_shape, envs.action_space_shape) current_obs = torch.zeros(args.num_processes, *obs_shape) def update_current_obs(obs): shape_dim0 = envs.observation_space_shape[0] obs = torch.from_numpy(obs).float() if args.num_stack > 1: current_obs[:, :-shape_dim0] = current_obs[:, shape_dim0:] current_obs[:, -shape_dim0:] = obs obs = envs.reset() update_current_obs(obs) last_game_vars = [] for i in range(args.num_processes): last_game_vars.append(np.zeros(args.num_events)) rollouts.observations[0].copy_(current_obs) # These variables are used to compute average rewards for all processes. episode_rewards = torch.zeros([args.num_processes, 1]) final_rewards = torch.zeros([args.num_processes, 1]) episode_intrinsic_rewards = torch.zeros([args.num_processes, 1]) final_intrinsic_rewards = torch.zeros([args.num_processes, 1]) episode_events = torch.zeros([args.num_processes, args.num_events]) final_events = torch.zeros([args.num_processes, args.num_events]) if args.cuda: current_obs = current_obs.cuda() rollouts.cuda() # Create event buffer if args.qd: event_buffer = EventBufferSQLProxy(args.num_events, args.capacity, args.exp_id, args.agent_id) elif not args.resume: event_buffer = EventBuffer(args.num_events, args.capacity) else: event_buffer = pickle.load( open(log_file_name + "_event_buffer_temp.p", "rb")) event_episode_rewards = [] start = time.time() for j in np.arange(start_updates, num_updates): for step in range(args.num_steps): value, action = actor_critic.act( Variable(rollouts.observations[step], volatile=True)) cpu_actions = action.data.squeeze(1).cpu().numpy() obs, reward, done, info, events = envs.step(cpu_actions) intrinsic_reward = [] # Fix broken rewards - upscale for i in range(len(reward)): if scenario_name in ["deathmatch", "my_way_home"]: reward[i] *= 100 if scenario_name == "deadly_corridor": reward[i] = 1 if events[i][2] >= 1 else 0 for e in events: if args.roe: intrinsic_reward.append(event_buffer.intrinsic_reward(e)) else: r = reward[len(intrinsic_reward)] intrinsic_reward.append(r) reward = torch.from_numpy(np.expand_dims(np.stack(reward), 1)).float() intrinsic_reward = torch.from_numpy( np.expand_dims(np.stack(intrinsic_reward), 1)).float() #events = torch.from_numpy(np.expand_dims(np.stack(events), args.num_events)).float() events = torch.from_numpy(events).float() episode_rewards += reward episode_intrinsic_rewards += intrinsic_reward episode_events += events # Event stats event_rewards = [] for ei in range(0, args.num_events): ev = np.zeros(args.num_events) ev[ei] = 1 er = event_buffer.intrinsic_reward(ev) event_rewards.append(er) event_episode_rewards.append(event_rewards) # If done then clean the history of observations. masks = torch.FloatTensor([[0.0] if done_ else [1.0] for done_ in done]) final_rewards *= masks final_intrinsic_rewards *= masks final_events *= masks final_rewards += (1 - masks) * episode_rewards final_intrinsic_rewards += (1 - masks) * episode_intrinsic_rewards final_events += (1 - masks) * episode_events for i in range(args.num_processes): if done[i]: event_buffer.record_events(np.copy( final_events[i].numpy()), frame=j * args.num_steps) episode_rewards *= masks episode_intrinsic_rewards *= masks episode_events *= masks if args.cuda: masks = masks.cuda() if current_obs.dim() == 4: current_obs *= masks.unsqueeze(2).unsqueeze(2) else: current_obs *= masks update_current_obs(obs) rollouts.insert(step, current_obs, action.data, value.data, intrinsic_reward, masks) final_episode_reward = np.mean(event_episode_rewards, axis=0) event_episode_rewards = [] next_value = actor_critic( Variable(rollouts.observations[-1], volatile=True))[0].data if hasattr(actor_critic, 'obs_filter'): actor_critic.obs_filter.update(rollouts.observations[:-1].view( -1, *obs_shape)) rollouts.compute_returns(next_value, args.use_gae, args.gamma, args.tau) values, action_log_probs, dist_entropy = actor_critic.evaluate_actions( Variable(rollouts.observations[:-1].view(-1, *obs_shape)), Variable(rollouts.actions.view(-1, action_shape))) values = values.view(args.num_steps, args.num_processes, 1) action_log_probs = action_log_probs.view(args.num_steps, args.num_processes, 1) advantages = Variable(rollouts.returns[:-1]) - values value_loss = advantages.pow(2).mean() action_loss = -(Variable(advantages.data) * action_log_probs).mean() optimizer.zero_grad() (value_loss * args.value_loss_coef + action_loss - dist_entropy * args.entropy_coef).backward() nn.utils.clip_grad_norm(actor_critic.parameters(), args.max_grad_norm) optimizer.step() rollouts.observations[0].copy_(rollouts.observations[-1]) if final_rewards.mean() > best_final_rewards and not args.debug: try: os.makedirs(save_path) except OSError: pass best_final_rewards = final_rewards.mean() save_model = actor_critic if args.cuda: save_model = copy.deepcopy(actor_critic).cpu() torch.save( save_model, os.path.join(save_path, log_file_name.split(".log")[0] + ".pt")) if j % args.save_interval == 0 and args.save_dir != "" and not args.debug: try: os.makedirs(save_path) except OSError: pass # A really ugly way to save a model to CPU save_model = actor_critic if args.cuda: save_model = copy.deepcopy(actor_critic).cpu() torch.save(save_model, os.path.join(save_path, log_file_name + "_temp.pt")) if isinstance(event_buffer, EventBuffer): pickle.dump(event_buffer, open(log_file_name + "_event_buffer_temp.p", "wb")) if j % args.log_interval == 0: envs.log() end = time.time() total_num_steps = (j + 1) * args.num_processes * args.num_steps log = "Updates {}, num timesteps {}, FPS {}, mean/max reward {:.5f}/{:.5f}, mean/max intrinsic reward {:.5f}/{:.5f}"\ .format(j, total_num_steps, int(total_num_steps / (end - start)), final_rewards.mean(), final_rewards.max(), final_intrinsic_rewards.mean(), final_intrinsic_rewards.max() ) log_to_file = "{}, {}, {:.5f}, {:.5f}, {:.5f}, {:.5f}\n" \ .format(j, total_num_steps, final_rewards.mean(), final_rewards.std(), final_intrinsic_rewards.mean(), final_intrinsic_rewards.std()) log_to_event_file = ','.join( map(str, event_buffer.get_event_mean().tolist())) + "\n" log_to_event_reward_file = ','.join( map(str, event_buffer.get_event_rewards().tolist())) + "\n" print(log) print(log_to_event_file) # Save to files with open(log_file_name, "a") as myfile: myfile.write(log_to_file) with open(log_event_file_name, "a") as myfile: myfile.write(str(total_num_steps) + "," + log_to_event_file) with open(log_event_reward_file_name, "a") as myfile: myfile.write( str(total_num_steps) + "," + log_to_event_reward_file) envs.close() game.close() time.sleep(5)
def main(): print("#######") print( "WARNING: All rewards are clipped or normalized so you need to use a monitor (see envs.py) or visdom plot to get true rewards" ) print("#######") if args.run_index is not None: load_params(args) try: os.makedirs(args.log_dir) except OSError: files = glob.glob(os.path.join(args.log_dir, '*.monitor.csv')) for f in files: os.remove(f) torch.cuda.manual_seed(args.seed) torch.manual_seed(args.seed) np.random.seed(args.seed) random.seed(args.seed) os.environ['OMP_NUM_THREADS'] = '1' envs = [ make_env(args.env_name, args.seed, i, args.log_dir) for i in range(args.num_processes) ] if args.num_processes > 1: envs = SubprocVecEnv(envs) else: envs = DummyVecEnv(envs) if len(envs.observation_space.shape) == 1: envs = VecNormalize(envs) obs_shape = envs.observation_space.shape obs_shape = (obs_shape[0] * args.num_stack, *obs_shape[1:]) num_heads = 1 if args.reward_predictor else len(args.gamma) assert len(envs.observation_space.shape) == 3 actor_critic = CNNPolicy(obs_shape[0], envs.action_space, use_rp=args.reward_predictor, num_heads=num_heads) assert envs.action_space.__class__.__name__ == "Discrete" action_shape = 1 if args.cuda: actor_critic.cuda() if not args.reward_predictor: model_params = actor_critic.parameters() else: lrs = [args.lr_rp, args.lr] model_params = [{ 'params': model_p, 'lr': p_lr } for model_p, p_lr in zip(actor_critic.param_groups, lrs)] optimizer = optim.RMSprop(model_params, args.lr, eps=args.eps, alpha=args.alpha) rollouts = RolloutStorage(args.num_steps, args.num_processes, obs_shape, envs.action_space, actor_critic.state_size, gamma=args.gamma, use_rp=args.reward_predictor) current_obs = torch.zeros(args.num_processes, *obs_shape) def update_current_obs(obs): shape_dim0 = envs.observation_space.shape[0] obs = torch.from_numpy(obs).float() if args.num_stack > 1: current_obs[:, :-shape_dim0] = current_obs[:, shape_dim0:] current_obs[:, -shape_dim0:] = obs obs = envs.reset() update_current_obs(obs) rollouts.observations[0].copy_(current_obs) # These variables are used to compute average rewards for all processes. episode_rewards = torch.zeros([args.num_processes, 1]) final_rewards = torch.zeros([args.num_processes, 1]) if args.cuda: current_obs = current_obs.cuda() rollouts.cuda() start = time.time() for j in range(num_updates): for step in range(args.num_steps): # Sample actions value, action, action_log_prob, states = actor_critic.act( Variable(rollouts.observations[step], volatile=True), Variable(rollouts.states[step], volatile=True), Variable(rollouts.masks[step], volatile=True)) cpu_actions = action.data.squeeze(1).cpu().numpy() obs, raw_reward, done, info = envs.step(cpu_actions) if args.reward_noise > 0.0: stds = np.ones(raw_reward.shape) * args.reward_noise noise = np.random.normal(loc=0.0, scale=stds) reward = raw_reward + noise else: reward = raw_reward raw_reward = torch.from_numpy( np.expand_dims(np.stack(raw_reward), 1)).float() episode_rewards += raw_reward reward = torch.from_numpy(np.expand_dims(np.stack(reward), 1)).float() if args.reward_predictor: p_hat = min(args.rp_burn_in, j) / args.rp_burn_in estimate_reward = ( 1 - p_hat ) * reward + p_hat * value[:, 0].unsqueeze(-1).data.cpu() reward = torch.cat([reward, estimate_reward], dim=-1) value = value.data else: value = value.data # If done then clean the history of observations. masks = torch.FloatTensor([[0.0] if done_ else [1.0] for done_ in done]) final_rewards *= masks final_rewards += (1 - masks) * episode_rewards episode_rewards *= masks if args.cuda: masks = masks.cuda() if current_obs.dim() == 4: current_obs *= masks.unsqueeze(2).unsqueeze(2) else: current_obs *= masks update_current_obs(obs) rollouts.insert(step, current_obs, states.data, action.data, action_log_prob.data, value, reward, masks) next_value = actor_critic( Variable(rollouts.observations[-1], volatile=True), Variable(rollouts.states[-1], volatile=True), Variable(rollouts.masks[-1], volatile=True))[0].data rollouts.compute_returns(next_value) states = Variable(rollouts.states[0].view(-1, actor_critic.state_size)) masks = Variable(rollouts.masks[:-1].view(-1, 1)) obs = Variable(rollouts.observations[:-1].view(-1, *obs_shape)) actions = Variable(rollouts.actions.view(-1, action_shape)) values, action_log_probs, dist_entropy, states = actor_critic.evaluate_actions( obs, states, masks, actions) returns_as_variable = Variable(rollouts.returns[:-1]) values = values.view(returns_as_variable.size()) action_log_probs = action_log_probs.view(args.num_steps, args.num_processes, 1) advantages = returns_as_variable - values value_loss = advantages.pow(2).sum(-1).mean() action_loss = -(Variable(advantages[:, :, -1].unsqueeze(-1).data) * action_log_probs).mean() optimizer.zero_grad() (value_loss * args.value_loss_coef + action_loss - dist_entropy * args.entropy_coef).backward() nn.utils.clip_grad_norm(actor_critic.parameters(), args.max_grad_norm) optimizer.step() rollouts.after_update() if j % args.save_interval == 0 and args.save_dir != "": save_path = os.path.join(args.save_dir, 'a2c') try: os.makedirs(save_path) except OSError: pass # A really ugly way to save a model to CPU save_model = actor_critic if args.cuda: save_model = copy.deepcopy(actor_critic).cpu() save_model = [ save_model, hasattr(envs, 'ob_rms') and envs.ob_rms or None ] torch.save(save_model, os.path.join(save_path, args.env_name + ".pt")) if j % args.log_interval == 0: end = time.time() total_num_steps = (j + 1) * args.num_processes * args.num_steps print( "Updates {}, num timesteps {}, FPS {}, mean/median reward {:.1f}/{:.1f}, min/max reward {:.1f}/{:.1f}, " "entropy {:.5f}, value loss {:.5f}, policy loss {:.5f}".format( j, total_num_steps, int(total_num_steps / (end - start)), final_rewards.mean(), final_rewards.median(), final_rewards.min(), final_rewards.max(), dist_entropy.data[0], value_loss.data[0], action_loss.data[0]))
def main(): print("#######") print("WARNING: All rewards are clipped or normalized so you need to use a monitor (see envs.py) or visdom plot to get true rewards") print("#######") os.environ['OMP_NUM_THREADS'] = '1' if args.vis: from visdom import Visdom viz = Visdom(port=args.port) win = None names = getListOfGames("train") envs = [make_env_train(names[i], args.seed, i, args.log_dir) for i in range(len(names))] # TODO TODO TODO TODO TODO TODO TODO TODO TODO TODO TODO TODO TODO TODO TODO TODO TODO TODO args.num_processes = len(envs) # REMEMBER YOU CHENGED IT if len(envs) > 1: envs = SubprocVecEnv(envs) else: envs = DummyVecEnv(envs) if len(envs.observation_space.shape) == 1: envs = VecNormalize(envs) obs_shape = envs.observation_space.shape #print(obs_shape) obs_shape = (obs_shape[0], *obs_shape[1:]) #print(obs_shape) if len(envs.observation_space.shape) == 3: actor_critic = CNNPolicy(obs_shape[0], envs.action_space, args.recurrent_policy) else: assert not args.recurrent_policy, \ "Recurrent policy is not implemented for the MLP controller" actor_critic = MLPPolicy(obs_shape[0], envs.action_space) # Making it paralel actor_critic = torch.nn.parallel.DataParallel(actor_critic).module if envs.action_space.__class__.__name__ == "Discrete": action_shape = 1 else: action_shape = envs.action_space.shape[0] if args.cuda: actor_critic.cuda() if args.algo == 'a2c': agent = algo.A2C_ACKTR(actor_critic, args.value_loss_coef, args.entropy_coef, lr=args.lr, eps=args.eps, alpha=args.alpha, max_grad_norm=args.max_grad_norm) elif args.algo == 'ppo': agent = algo.PPO(actor_critic, args.clip_param, args.ppo_epoch, args.num_mini_batch, args.value_loss_coef, args.entropy_coef, lr=args.lr, eps=args.eps, max_grad_norm=args.max_grad_norm) # Make agent DataParallel agent = torch.nn.parallel.DataParallel(agent).module elif args.algo == 'acktr': agent = algo.A2C_ACKTR(actor_critic, args.value_loss_coef, args.entropy_coef, acktr=True) # Make rollouts DataParallel rollouts = torch.nn.parallel.DataParallel(RolloutStorage(args.num_steps, args.num_processes, obs_shape, envs.action_space, actor_critic.state_size)).module current_obs = torch.nn.parallel.DataParallel(torch.zeros(envs.nenvs, *obs_shape)).module def update_current_obs(obs): shape_dim0 = envs.observation_space.shape[0] obs = torch.from_numpy(obs).float() # if args.num_stack > 1: # current_obs[:, :-shape_dim0] = current_obs[:, shape_dim0:] current_obs[:, -shape_dim0:] = obs obs = envs.reset() update_current_obs(obs) rollouts.observations[0].copy_(current_obs) # These variables are used to compute average rewards for all processes. episode_rewards = torch.zeros([args.num_processes, 1]) final_rewards = torch.zeros([args.num_processes, 1]) if args.cuda: current_obs = current_obs.cuda() rollouts.cuda() start = time.time() for j in range(num_updates): for step in range(args.num_steps): # Sample actions value, action, action_log_prob, states = actor_critic.act( Variable(rollouts.observations[step], volatile=True), Variable(rollouts.states[step], volatile=True), Variable(rollouts.masks[step], volatile=True)) cpu_actions = action.data.squeeze(1).cpu().numpy() # Obser reward and next obs obs, reward, done, info = envs.step(cpu_actions) reward = torch.from_numpy(np.expand_dims(np.stack(reward), 1)).float() episode_rewards += reward # If done then clean the history of observations. masks = torch.FloatTensor([[0.0] if done_ else [1.0] for done_ in done]) final_rewards *= masks final_rewards += (1 - masks) * episode_rewards episode_rewards *= masks if args.cuda: masks = masks.cuda() if current_obs.dim() == 4: current_obs *= masks.unsqueeze(2).unsqueeze(2) else: current_obs *= masks update_current_obs(obs) rollouts.insert(current_obs, states.data, action.data, action_log_prob.data, value.data, reward, masks) next_value = actor_critic.get_value(Variable(rollouts.observations[-1], volatile=True), Variable(rollouts.states[-1], volatile=True), Variable(rollouts.masks[-1], volatile=True)).data rollouts.compute_returns(next_value, args.use_gae, args.gamma, args.tau) value_loss, action_loss, dist_entropy = agent.update(rollouts) rollouts.after_update() if j % args.save_interval == 0 and args.save_dir != "": save_path = os.path.join(args.save_dir, args.algo) try: os.makedirs(save_path) except OSError: pass # A really ugly way to save a model to CPU save_model = actor_critic if args.cuda: save_model = copy.deepcopy(actor_critic).cpu() save_model = [save_model, hasattr(envs, 'ob_rms') and envs.ob_rms or None] torch.save(save_model, os.path.join(save_path, args.env_name + ".pt")) if j % args.log_interval == 0: end = time.time() total_num_steps = (j + 1) * args.num_processes * args.num_steps print("Updates {}, num timesteps {}, FPS {}, mean/median reward {:.1f}/{:.1f}, min/max reward {:.1f}/{:.1f}, entropy {:.5f}, value loss {:.5f}, policy loss {:.5f}". format(j, total_num_steps, int(total_num_steps / (end - start)), final_rewards.mean(), final_rewards.median(), final_rewards.min(), final_rewards.max(), dist_entropy.data[0], value_loss.data[0], action_loss.data[0])) if args.vis and j % args.vis_interval == 0: try: # Sometimes monitor doesn't properly flush the outputs win = visdom_plot(viz, win, args.log_dir, args.env_name, args.algo, args.num_frames) except IOError: pass
state = envs.reset() update_current_state(state) rollouts.states[0].copy_(current_state) episode_rewards = torch.zeros([args.num_processes, 1]) final_rewards = torch.zeros([args.num_processes, 1]) if args.cuda: current_state = current_state.cuda() rollouts.cuda() for j in range(num_updates): for step in range(args.num_steps): value, action = actor_critic.act( Variable(rollouts.states[step], volatile=True)) cpu_actions = action.data.squeeze(1).cpu().numpy() state, reward, done, info = envs.step(cpu_actions) reward = torch.from_numpy(np.expand_dims(np.stack(reward), 1)).float() episode_rewards += reward masks = torch.FloatTensor([[0.0] if done_ else [1.0] for done_ in done]) final_rewards *= masks final_rewards += (1 - masks) * episode_rewards episode_rewards *= masks if args.cuda: masks = masks.cuda()
def main(): print("#######") print( "WARNING: All rewards are clipped or normalized so you need to use a monitor (see envs.py) or visdom plot to get true rewards" ) print("#######") os.environ['OMP_NUM_THREADS'] = '1' if args.vis: from visdom import Visdom viz = Visdom() win = None envs = [ make_env(args.env_name, args.seed, i, args.log_dir) for i in range(args.num_processes) ] if args.num_processes > 1: envs = SubprocVecEnv(envs) else: envs = DummyVecEnv(envs) if len(envs.observation_space.shape) == 1: envs = VecNormalize(envs) obs_shape = envs.observation_space.shape obs_shape = (obs_shape[0] * args.num_stack, *obs_shape[1:] ) # I guess the obs_shape[0] is channel number if len(envs.observation_space.shape) == 3: actor_critic = CNNPolicy(obs_shape[0], envs.action_space, args.recurrent_policy) else: assert not args.recurrent_policy, \ "Recurrent policy is not implemented for the MLP controller" actor_critic = MLPPolicy(obs_shape[0], envs.action_space) if envs.action_space.__class__.__name__ == "Discrete": action_shape = 1 else: action_shape = envs.action_space.shape[0] if args.cuda: actor_critic.cuda() if args.algo == 'a2c': optimizer = optim.RMSprop(actor_critic.parameters(), args.lr, eps=args.eps, alpha=args.alpha) elif args.algo == 'ppo': optimizer = optim.Adam(actor_critic.parameters(), args.lr, eps=args.eps) elif args.algo == 'acktr': optimizer = KFACOptimizer(actor_critic) rollouts = RolloutStorage(args.num_steps, args.num_processes, obs_shape, envs.action_space, actor_critic.state_size) current_obs = torch.zeros(args.num_processes, *obs_shape) def update_current_obs(obs): shape_dim0 = envs.observation_space.shape[0] obs = torch.from_numpy(obs).float() if args.num_stack > 1: current_obs[:, :-shape_dim0] = current_obs[:, shape_dim0:] current_obs[:, -shape_dim0:] = obs obs = envs.reset() update_current_obs(obs) rollouts.observations[0].copy_(current_obs) # These variables are used to compute average rewards for all processes. episode_rewards = torch.zeros([args.num_processes, 1]) final_rewards = torch.zeros([args.num_processes, 1]) if args.cuda: current_obs = current_obs.cuda() rollouts.cuda() start = time.time() for j in range(num_updates): for step in range(args.num_steps): # args.num_steps should be the length of interactions before each updating/training # Sample actions value, action, action_log_prob, states = actor_critic.act( Variable(rollouts.observations[step], volatile=True), Variable(rollouts.states[step], volatile=True), Variable(rollouts.masks[step], volatile=True)) cpu_actions = action.data.squeeze(1).cpu().numpy( ) # returns are state value, sampled action, act_log_prob, hidden states # Obser reward and next obs obs, reward, done, info = envs.step(cpu_actions) reward = torch.from_numpy(np.expand_dims(np.stack(reward), 1)).float() episode_rewards += reward # If done then clean the history of observations. masks = torch.FloatTensor([[0.0] if done_ else [1.0] for done_ in done]) final_rewards *= masks final_rewards += (1 - masks) * episode_rewards episode_rewards *= masks if args.cuda: masks = masks.cuda() if current_obs.dim() == 4: current_obs *= masks.unsqueeze(2).unsqueeze(2) else: current_obs *= masks update_current_obs(obs) rollouts.insert( step, current_obs, states.data, action.data, action_log_prob.data, value.data, reward, masks ) # so the rollout stores one batch of interaction sequences, each sequence has length of args.num_steps next_value = actor_critic( Variable(rollouts.observations[-1], volatile=True), Variable(rollouts.states[-1], volatile=True), Variable(rollouts.masks[-1], volatile=True))[0].data rollouts.compute_returns(next_value, args.use_gae, args.gamma, args.tau) if args.algo in ['a2c', 'acktr']: values, action_log_probs, dist_entropy, states = actor_critic.evaluate_actions( Variable(rollouts.observations[:-1].view(-1, *obs_shape)), Variable(rollouts.states[0].view(-1, actor_critic.state_size)), Variable(rollouts.masks[:-1].view(-1, 1)), Variable(rollouts.actions.view(-1, action_shape))) # values should be values of observations, states are the hidden states used in rnn module, by pwang8 values = values.view( args.num_steps, args.num_processes, 1) # values are estimated current state values action_log_probs = action_log_probs.view(args.num_steps, args.num_processes, 1) # rollouts.returns are current "Action" value calculted following Bellmans' eqaution gamma * State_value(t+1) + reward(t) advantages = Variable( rollouts.returns[:-1] ) - values # This is also the definition of advantage value (action_value - state_value). value_loss = advantages.pow( 2).mean() # values are estimated current state_value(t) action_loss = -(Variable(advantages.data) * action_log_probs).mean() # If ACKTR is utilized, it is not only a different optimizer is used, they also added some new loss source if args.algo == 'acktr' and optimizer.steps % optimizer.Ts == 0: # Sampled fisher, see Martens 2014 actor_critic.zero_grad() pg_fisher_loss = -action_log_probs.mean() value_noise = Variable(torch.randn(values.size())) if args.cuda: value_noise = value_noise.cuda() sample_values = values + value_noise vf_fisher_loss = -( values - Variable(sample_values.data) ).pow(2).mean( ) # don't know what is the difference between this and just randomly sample some noise fisher_loss = pg_fisher_loss + vf_fisher_loss optimizer.acc_stats = True fisher_loss.backward(retain_graph=True) optimizer.acc_stats = False optimizer.zero_grad() (value_loss * args.value_loss_coef + action_loss - dist_entropy * args.entropy_coef).backward() if args.algo == 'a2c': nn.utils.clip_grad_norm(actor_critic.parameters(), args.max_grad_norm) optimizer.step() elif args.algo == 'ppo': advantages = rollouts.returns[:-1] - rollouts.value_preds[: -1] # calculating the advantage value of an action advantages = (advantages - advantages.mean()) / (advantages.std() + 1e-5) # The difference from this ppo optimization to the optimization above is that: it updates params for # multiple epochs in ppo optimization. Because of this, it samples from the rollouts storage a minibatch # every time to calculate gradient. Sampling is conducted for optimization purpose. for e in range(args.ppo_epoch): if args.recurrent_policy: data_generator = rollouts.recurrent_generator( advantages, args.num_mini_batch) else: data_generator = rollouts.feed_forward_generator( advantages, args.num_mini_batch) for sample in data_generator: observations_batch, states_batch, actions_batch, \ return_batch, masks_batch, old_action_log_probs_batch, \ adv_targ = sample # Reshape to do in a single forward pass for all steps values, action_log_probs, dist_entropy, states = actor_critic.evaluate_actions( Variable(observations_batch), Variable(states_batch), Variable(masks_batch), Variable(actions_batch)) # For the 1st epoch of updating, I guess the action_log_probls is the same as old_action_log_probs_batch # because params of the NN have not been updated at that time. But later, in other updating epochs, # this ratio will generate some error. The old_action_log_probs_batch will not be updated during # these param updating epochs. # action_log_probs is the log prob of that action taken by the agent. So it's one value here, not # log_prob for all actions with certain input observation/state. By pwang8, Dec 31, 2017 adv_targ = Variable(adv_targ) ratio = torch.exp(action_log_probs - Variable(old_action_log_probs_batch)) surr1 = ratio * adv_targ surr2 = torch.clamp(ratio, 1.0 - args.clip_param, 1.0 + args.clip_param) * adv_targ action_loss = -torch.min( surr1, surr2).mean() # PPO's pessimistic surrogate (L^CLIP) # compared to a2c, the major difference for ppo is that action_loss is calculated in controlled way value_loss = (Variable(return_batch) - values).pow(2).mean() optimizer.zero_grad() (value_loss + action_loss - dist_entropy * args.entropy_coef).backward() nn.utils.clip_grad_norm(actor_critic.parameters(), args.max_grad_norm) optimizer.step() rollouts.after_update() if j % args.save_interval == 0 and args.save_dir != "": save_path = os.path.join(args.save_dir, args.algo) try: os.makedirs(save_path) except OSError: pass # A really ugly way to save a model to CPU save_model = actor_critic if args.cuda: save_model = copy.deepcopy(actor_critic).cpu() save_model = [ save_model, hasattr(envs, 'ob_rms') and envs.ob_rms or None ] torch.save(save_model, os.path.join(save_path, args.env_name + ".pt")) if j % args.log_interval == 0: end = time.time() total_num_steps = (j + 1) * args.num_processes * args.num_steps print( "Updates {}, num timesteps {}, FPS {}, mean/median reward {:.1f}/{:.1f}, min/max reward {:.1f}/{:.1f}, entropy {:.5f}, value loss {:.5f}, policy loss {:.5f}" .format(j, total_num_steps, int(total_num_steps / (end - start)), final_rewards.mean(), final_rewards.median(), final_rewards.min(), final_rewards.max(), dist_entropy.data[0], value_loss.data[0], action_loss.data[0])) if args.vis and j % args.vis_interval == 0: try: # Sometimes monitor doesn't properly flush the outputs win = visdom_plot(viz, win, args.log_dir, args.env_name, args.algo) except IOError: pass
def main(): envs = [make_env(env_name, seed, rank, log_dir) for rank in range(num_processes)] envs = SubprocVecEnv(envs) obs_shape = envs.observation_space.shape obs_shape = [obs_shape[0]*num_stack, *obs_shape[1:]] actor_critic = CNNPolicy(obs_shape[0], envs.action_space, False) if cuda: actor_critic.cuda() optimizer = optim.RMSprop(actor_critic.parameters(), lr, eps=eps, alpha=alpha) rollouts = RolloutStorage(num_steps, num_processes, obs_shape, envs.action_space, actor_critic.state_size) current_obs = torch.zeros(num_processes, *obs_shape) def update_current_obs(obs): shape_dim0 = envs.observation_space.shape[0] obs = torch.from_numpy(obs).float() if num_stack > 1: current_obs[:, :-shape_dim0] = current_obs[:, shape_dim0:] current_obs[:, -shape_dim0:] = obs obs = envs.reset() update_current_obs(obs) rollouts.observations[0].copy_(current_obs) episode_rewards = torch.zeros([num_processes,1]) final_rewards = torch.zeros([num_processes,1]) if cuda: rollouts.cuda() current_obs = current_obs.cuda() if envs.action_space.__class__.__name__ == "Discrete": action_shape = 1 else: action_shape = envs.action_space.shape[0] # test start = time.time() for j in range(num_updates): for step in range(num_steps): # Sample actions value, action, action_log_prob, states = actor_critic.act(Variable(rollouts.observations[step], volatile=True), Variable(rollouts.states[step], volatile=True), Variable(rollouts.masks[step], volatile=True)) cpu_actions = action.data.squeeze().cpu().numpy() #print(cpu_action) # obser reward and next obs obs, reward, done, info = envs.step(cpu_actions) # stack: make sure that reward is a numpy array(convert list to ndarray) reward = torch.from_numpy(np.expand_dims(np.stack(reward), 1)).float() episode_rewards += reward # If done then clean the history of observations. masks = torch.FloatTensor([[0.0] if done_ else [1.0] for done_ in done]) final_rewards *= masks final_rewards += (1 - masks) * episode_rewards episode_rewards *= masks if cuda: masks = masks.cuda() if current_obs.dim() == 4: current_obs *= masks.unsqueeze(2).unsqueeze(2) else: current_obs *= masks # update obs nad rollouts update_current_obs(obs) rollouts.insert(step, current_obs, states.data, action.data, action_log_prob.data, value.data, reward, masks) # compute current update's return next_value = actor_critic(Variable(rollouts.observations[-1], volatile=True), Variable(rollouts.states[-1], volatile=True), Variable(rollouts.masks[-1], volatile=True))[0].data rollouts.compute_returns(next_value, False, gamma, tau) # in a2c the values were calculated twice # the data in rollouts must be viewed, because the shape in rollouts is [num_steps, num_processes, x] which is [num,x] in actor_critic values, action_log_probs, dist_entropy, states = actor_critic.evaluate_actions(Variable(rollouts.observations[:-1].view(-1, *obs_shape)), Variable(rollouts.states[0].view(-1, actor_critic.state_size)), Variable(rollouts.masks[:-1].view(-1, 1)), Variable(rollouts.actions.view(-1, action_shape))) # compute the loss values = values.view(num_steps, num_processes, 1) action_log_probs = action_log_probs.view(num_steps, num_processes, 1) advantages = Variable(rollouts.returns[:-1]) - values value_loss = advantages.pow(2).mean() action_loss = -(Variable(advantages.data) * action_log_probs).mean() # update model optimizer.zero_grad() loss = value_loss * value_loss_coef + action_loss - dist_entropy * entropy_coef loss.backward() nn.utils.clip_grad_norm(actor_critic.parameters(), max_grad_norm) optimizer.step() rollouts.after_update() if j % log_interval == 0: end = time.time() total_num_steps = (j + 1) * num_processes * num_steps print("Updates {}, num timesteps {}, FPS {}, mean/median reward {:.1f}/{:.1f}, min/max reward {:.1f}/{:.1f}, entropy {:.5f}, value loss {:.5f}, policy loss {:.5f}". format(j, total_num_steps, int(total_num_steps / (end - start)), final_rewards.mean(), final_rewards.median(), final_rewards.min(), final_rewards.max(), dist_entropy.data[0], value_loss.data[0], action_loss.data[0])) # todo: test save_url torch.save(actor_critic,save_url)
def main(): print("###############################################################") print("#################### VIZDOOM LEARNER START ####################") print("###############################################################") save_path = os.path.join(args.save_dir, str(args.exp_id)) log_path = os.path.join(args.log_dir, str(args.exp_id)) num_updates = int(args.num_frames) // args.num_steps // args.num_processes reward_name = "" if args.roe: reward_name = "_event" scenario_name = args.config_path.split("/")[1].split(".")[0] print("############### " + scenario_name + " ###############") log_file_name = "vizdoom_" + scenario_name + reward_name + "_" + str( args.exp_id) + "_" + str(args.agent_id) + ".log" #log_event_file_name = "vizdoom_" + scenario_name + reward_name + "_" + str(args.exp_id) + "_" + str(args.agent_id) + ".eventlog" #log_event_reward_file_name = "vizdoom_" + scenario_name + reward_name + "_" + str(args.exp_id) + "_" + str(args.agent_id) + ".eventrewardlog" start_updates = 0 start_step = 0 best_final_rewards = -1000000.0 os.environ['OMP_NUM_THREADS'] = '1' cig = "cig" in args.config_path global envs es = [ make_env(i, args.config_path, visual=args.visual, cig=cig) for i in range(args.num_processes) ] envs = VecEnv([es[i] for i in range(args.num_processes)]) obs_shape = envs.observation_space_shape obs_shape = (obs_shape[0] * args.num_stack, *obs_shape[1:]) if args.resume: actor_critic = torch.load( os.path.join(save_path, f"{args.agent_id}.pt")) filename = glob.glob(os.path.join(log_path, log_file_name))[0] with open(filename) as file: lines = file.readlines() start_updates = (int)(lines[-1].strip().split(",")[0]) start_steps = (int)(lines[-1].strip().split(",")[1]) num_updates += start_updates else: try: os.makedirs(save_path) except OSError: pass try: os.makedirs(log_path) except OSError: files = glob.glob(os.path.join(args.log_dir, log_file_name)) for f in files: os.remove(f) #with open(log_file_name, "w") as myfile: # myfile.write("") #files = glob.glob(os.path.join(args.log_dir, log_event_file_name)) #for f in files: # os.remove(f) #with open(log_event_file_name, "w") as myfile: # myfile.write("") #files = glob.glob(os.path.join(args.log_dir, log_event_reward_file_name)) #for f in files: # os.remove(f) #with open(log_event_reward_file_name, "w") as myfile: # myfile.write("") actor_critic = CNNPolicy(obs_shape[0], envs.action_space_shape) action_shape = 1 if args.cuda: actor_critic.cuda() optimizer = optim.RMSprop(actor_critic.parameters(), args.lr, eps=args.eps, alpha=args.alpha) rollouts = RolloutStorage(args.num_steps, args.num_processes, obs_shape, envs.action_space_shape) current_obs = torch.zeros(args.num_processes, *obs_shape) def update_current_obs(obs): shape_dim0 = envs.observation_space_shape[0] obs = torch.from_numpy(obs).float() if args.num_stack > 1: current_obs[:, :-shape_dim0] = current_obs[:, shape_dim0:] current_obs[:, -shape_dim0:] = obs obs = envs.reset() update_current_obs(obs) last_game_vars = [] for i in range(args.num_processes): last_game_vars.append(np.zeros(args.num_events)) rollouts.observations[0].copy_(current_obs) # These variables are used to compute average rewards for all processes. episode_rewards = torch.zeros([args.num_processes, 1]) final_rewards = torch.zeros([args.num_processes, 1]) episode_intrinsic_rewards = torch.zeros([args.num_processes, 1]) final_intrinsic_rewards = torch.zeros([args.num_processes, 1]) episode_events = torch.zeros([args.num_processes, args.num_events]) final_events = torch.zeros([args.num_processes, args.num_events]) if args.cuda: current_obs = current_obs.cuda() rollouts.cuda() def mean_distance_to_nearest_neighbor(elite_events): d = [] nearest = None for a in range(len(elite_events)): for b in range(len(elite_events)): if a != b: elite_a = elite_events[a] elite_b = elite_events[b] dist = np.linalg.norm(elite_a - elite_b) if nearest is None or dist < nearest: nearest = dist if nearest is not None: d.append(nearest) nearest = None return np.mean(d) def distance_to_nearest_neighbor(elite_events, events): nearest = None for elite_a in elite_events: dist = np.linalg.norm(elite_a - events) if nearest is None or dist < nearest: nearest = dist return nearest def add_to_archive(frame, episode_length): #print("Final rewards: ", final_rewards.numpy()) fitness = final_rewards.numpy().mean() #print("raw: ", final_events.numpy()) behavior = final_events.numpy().mean(axis=0) #print("Fitness:", fitness) #print("Behavior:", behavior) neighbors = event_buffer.get_neighbors(behavior, args.niche_divs, episode_length) add = len(neighbors) == 0 for neighbor in neighbors: if fitness > neighbor.fitness: add = True else: add = False break if add: if len(neighbors) > 0: event_buffer.remove_elites(neighbors) #print(f"- Removing elites {[neighbor.elite_id for neighbor in neighbors]}") for neighbor in neighbors: try: #print(f"- Deleting model {neighbor.elite_id}") os.remove( os.path.join(save_path, f"{neighbor.elite_id}.pt")) #print("Successfully deleted model with id : ", neighbor.elite_id) except: print("Error while deleting model with id : ", neighbor.elite_id) name = str(uuid.uuid1()) #print("Adding elite") event_buffer.add_elite(name, behavior, fitness, frame, episode_length) save_model = actor_critic if args.cuda: save_model = copy.deepcopy(actor_critic).cpu() torch.save(save_model, os.path.join(save_path, f"{name}.pt")) # Create event buffer event_buffer = EventBufferSQLProxy(args.num_events, args.capacity, args.exp_id, args.agent_id, qd=args.qd, per_step=args.per_step) event_episode_rewards = [] episode_finished = np.zeros(args.num_processes) start = time.time() for j in np.arange(start_updates, num_updates): for step in range(args.num_steps): value, action = actor_critic.act( Variable(rollouts.observations[step], volatile=True)) cpu_actions = action.data.squeeze(1).cpu().numpy() obs, reward, done, info, events = envs.step(cpu_actions) intrinsic_reward = [] # Fix broken rewards - upscale for i in range(len(reward)): if scenario_name in ["deathmatch", "my_way_home"]: reward[i] *= 100 if scenario_name == "deadly_corridor": reward[i] = 1 if events[i][2] >= 1 else 0 for e in events: if args.roe: ir = event_buffer.intrinsic_reward(e) if args.per_step: ir = ir / 4200 intrinsic_reward.append(ir) else: r = reward[len(intrinsic_reward)] intrinsic_reward.append(r) reward = torch.from_numpy(np.expand_dims(np.stack(reward), 1)).float() intrinsic_reward = torch.from_numpy( np.expand_dims(np.stack(intrinsic_reward), 1)).float() #events = torch.from_numpy(np.expand_dims(np.stack(events), args.num_events)).float() events = torch.from_numpy(events).float() episode_rewards += reward episode_intrinsic_rewards += intrinsic_reward episode_events += events # Event stats ''' event_rewards = [] for ei in range(0,args.num_events): ev = np.zeros(args.num_events) ev[ei] = 1 er = event_buffer.intrinsic_reward(ev) if args.per_step: er = er / 4200 er = event_buffer.intrinsic_reward(ev) event_rewards.append(er) event_episode_rewards.append(event_rewards) ''' # If done then clean the history of observations. masks = torch.FloatTensor([[0.0] if done_ else [1.0] for done_ in done]) final_rewards *= masks final_intrinsic_rewards *= masks final_events *= masks final_rewards += (1 - masks) * episode_rewards final_intrinsic_rewards += (1 - masks) * episode_intrinsic_rewards final_events += (1 - masks) * episode_events for i in range(args.num_processes): if done[i]: #event_buffer.record_events(np.copy(final_events[i].numpy()), frame=j*args.num_steps*args.num_processes) episode_length = (step + j * args.num_steps) - episode_finished[i] episode_finished[i] = episode_length + episode_finished[i] add_to_archive( step * args.num_processes + j * args.num_steps * args.num_processes, episode_length) episode_rewards *= masks episode_intrinsic_rewards *= masks episode_events *= masks if args.cuda: masks = masks.cuda() if current_obs.dim() == 4: current_obs *= masks.unsqueeze(2).unsqueeze(2) else: current_obs *= masks update_current_obs(obs) rollouts.insert(step, current_obs, action.data, value.data, intrinsic_reward, masks) #final_episode_reward = np.mean(event_episode_rewards, axis=0) #event_episode_rewards = [] next_value = actor_critic( Variable(rollouts.observations[-1], volatile=True))[0].data if hasattr(actor_critic, 'obs_filter'): actor_critic.obs_filter.update(rollouts.observations[:-1].view( -1, *obs_shape)) rollouts.compute_returns(next_value, args.use_gae, args.gamma, args.tau) values, action_log_probs, dist_entropy = actor_critic.evaluate_actions( Variable(rollouts.observations[:-1].view(-1, *obs_shape)), Variable(rollouts.actions.view(-1, action_shape))) values = values.view(args.num_steps, args.num_processes, 1) action_log_probs = action_log_probs.view(args.num_steps, args.num_processes, 1) advantages = Variable(rollouts.returns[:-1]) - values value_loss = advantages.pow(2).mean() action_loss = -(Variable(advantages.data) * action_log_probs).mean() optimizer.zero_grad() (value_loss * args.value_loss_coef + action_loss - dist_entropy * args.entropy_coef).backward() nn.utils.clip_grad_norm(actor_critic.parameters(), args.max_grad_norm) optimizer.step() rollouts.observations[0].copy_(rollouts.observations[-1]) if j % args.log_interval == 0: envs.log() end = time.time() total_num_steps = (j + 1) * args.num_processes * args.num_steps log = "Updates {}, num timesteps {}, FPS {}, mean/max reward {:.5f}/{:.5f}, mean/max intrinsic reward {:.5f}/{:.5f}"\ .format(j, total_num_steps, int(total_num_steps / (end - start)), final_rewards.mean(), final_rewards.max(), final_intrinsic_rewards.mean(), final_intrinsic_rewards.max() ) log_to_file = "{}, {}, {:.5f}, {:.5f}, {:.5f}, {:.5f}\n" \ .format(j, total_num_steps, final_rewards.mean(), final_rewards.std(), final_intrinsic_rewards.mean(), final_intrinsic_rewards.std()) with open(os.path.join(log_path, log_file_name), "a") as myfile: myfile.write(log_to_file) save_model = actor_critic if args.cuda: save_model = copy.deepcopy(actor_critic).cpu() torch.save(save_model, os.path.join(save_path, f"{args.agent_id}.pt")) print(log) envs.close() time.sleep(5)
def main(): print("###############################################################") print("#################### VISDOOM LEARNER START ####################") print("###############################################################") os.environ['OMP_NUM_THREADS'] = '1' if args.vis: from visdom import Visdom viz = Visdom() win = None global envs envs = VecEnv( [make_env(i, args.config_path) for i in range(args.num_processes)], logging=True, log_dir=args.log_dir) obs_shape = envs.observation_space_shape obs_shape = (obs_shape[0] * args.num_stack, *obs_shape[1:]) if args.algo == 'a2c' or args.algo == 'acktr': actor_critic = CNNPolicy(obs_shape[0], envs.action_space_shape) elif args.algo == 'a2t': source_models = [] files = glob.glob(os.path.join(args.source_models_path, '*.pt')) for file in files: print(file, 'loading model...') source_models.append(torch.load(file)) actor_critic = A2TPolicy(obs_shape[0], envs.action_space_shape, source_models) elif args.algo == 'resnet': # args.num_stack = 3 actor_critic = ResnetPolicy(obs_shape[0], envs.action_space_shape) action_shape = 1 if args.cuda: actor_critic.cuda() if args.algo == 'a2c' or args.algo == 'resnet': optimizer = optim.RMSprop(actor_critic.parameters(), args.lr, eps=args.eps, alpha=args.alpha) elif args.algo == 'a2t': a2t_params = [p for p in actor_critic.parameters() if p.requires_grad] optimizer = optim.RMSprop(a2t_params, args.lr, eps=args.eps, alpha=args.alpha) elif args.algo == 'acktr': optimizer = KFACOptimizer(actor_critic) rollouts = RolloutStorage(args.num_steps, args.num_processes, obs_shape, envs.action_space_shape) current_obs = torch.zeros(args.num_processes, *obs_shape) def update_current_obs(obs): shape_dim0 = envs.observation_space_shape[0] obs = torch.from_numpy(obs).float() if args.num_stack > 1: current_obs[:, :-shape_dim0] = current_obs[:, shape_dim0:] current_obs[:, -shape_dim0:] = obs obs = envs.reset() update_current_obs(obs) rollouts.observations[0].copy_(current_obs) # These variables are used to compute average rewards for all processes. episode_rewards = torch.zeros([args.num_processes, 1]) final_rewards = torch.zeros([args.num_processes, 1]) if args.cuda: current_obs = current_obs.cuda() rollouts.cuda() start = time.time() for j in range(num_updates): for step in range(args.num_steps): # Sample actions value, action = actor_critic.act( Variable(rollouts.observations[step], volatile=True)) cpu_actions = action.data.squeeze(1).cpu().numpy() # Obser reward and next obs obs, reward, done, info = envs.step(cpu_actions) # print ('Actions:', cpu_actions, 'Rewards:', reward) reward = torch.from_numpy(np.expand_dims(np.stack(reward), 1)).float() episode_rewards += reward # If done then clean the history of observations. masks = torch.FloatTensor([[0.0] if done_ else [1.0] for done_ in done]) final_rewards *= masks final_rewards += (1 - masks) * episode_rewards episode_rewards *= masks if args.cuda: masks = masks.cuda() if current_obs.dim() == 4: current_obs *= masks.unsqueeze(2).unsqueeze(2) else: current_obs *= masks update_current_obs(obs) rollouts.insert(step, current_obs, action.data, value.data, reward, masks) next_value = actor_critic( Variable(rollouts.observations[-1], volatile=True))[0].data if hasattr(actor_critic, 'obs_filter'): actor_critic.obs_filter.update(rollouts.observations[:-1].view( -1, *obs_shape)) rollouts.compute_returns(next_value, args.use_gae, args.gamma, args.tau) if args.algo in ['a2c', 'acktr', 'a2t', 'resnet']: values, action_log_probs, dist_entropy = actor_critic.evaluate_actions( Variable(rollouts.observations[:-1].view(-1, *obs_shape)), Variable(rollouts.actions.view(-1, action_shape))) values = values.view(args.num_steps, args.num_processes, 1) action_log_probs = action_log_probs.view(args.num_steps, args.num_processes, 1) advantages = Variable(rollouts.returns[:-1]) - values value_loss = advantages.pow(2).mean() action_loss = -(Variable(advantages.data) * action_log_probs).mean() if args.algo == 'acktr' and optimizer.steps % optimizer.Ts == 0: # Sampled fisher, see Martens 2014 actor_critic.zero_grad() pg_fisher_loss = -action_log_probs.mean() value_noise = Variable(torch.randn(values.size())) if args.cuda: value_noise = value_noise.cuda() sample_values = values + value_noise vf_fisher_loss = -(values - Variable(sample_values.data)).pow(2).mean() fisher_loss = pg_fisher_loss + vf_fisher_loss optimizer.acc_stats = True fisher_loss.backward(retain_graph=True) optimizer.acc_stats = False optimizer.zero_grad() (value_loss * args.value_loss_coef + action_loss - dist_entropy * args.entropy_coef).backward() if args.algo == 'a2c' or args.algo == 'resnet': nn.utils.clip_grad_norm(actor_critic.parameters(), args.max_grad_norm) elif args.algo == 'a2t': nn.utils.clip_grad_norm(a2t_params, args.max_grad_norm) optimizer.step() rollouts.observations[0].copy_(rollouts.observations[-1]) if j % args.save_interval == 0 and args.save_dir != "": save_path = os.path.join(args.save_dir, args.algo) try: os.makedirs(save_path) except OSError: pass # A really ugly way to save a model to CPU save_model = actor_critic if args.cuda: save_model = copy.deepcopy(actor_critic).cpu() torch.save(save_model, os.path.join(save_path, args.env_name + ".pt")) if j % args.log_interval == 0: envs.log() end = time.time() total_num_steps = (j + 1) * args.num_processes * args.num_steps print( "Updates {}, num timesteps {}, FPS {}, mean/median reward {:.1f}/{:.1f}, min/max reward {:.1f}/{:.1f}, entropy {:.5f}, value loss {:.5f}, policy loss {:.5f}" .format(j, total_num_steps, int(total_num_steps / (end - start)), final_rewards.mean(), final_rewards.median(), final_rewards.min(), final_rewards.max(), -dist_entropy.data[0], value_loss.data[0], action_loss.data[0])) if args.vis and j % args.vis_interval == 0: try: # Sometimes monitor doesn't properly flush the outputs win = visdom_plot(viz, win, args.log_dir, 'VizDoom', args.algo) except IOError: pass envs.close() time.sleep(5)
def main(): print("#######") print("WARNING: All rewards are clipped so you need to use a monitor (see envs.py) or visdom plot to get true rewards") print("#######") os.environ['OMP_NUM_THREADS'] = '1' if args.vis: from visdom import Visdom viz = Visdom() win = [] win_dic ={} for i in range(len(mt_env_id_dic_selected)): win += [None] win_afs_per_m = None win_afs_loss = None win_basic_loss = None plot_dic = {} envs = [] ''' Because the oral program has only one game per model, so Song add loop i So whatever you wanna run , just put in SubprocVecEnvMt! ''' for i in range(len(mt_env_id_dic_selected)): log_dir = args.log_dir+mt_env_id_dic_selected[i]+'/' for j in range(args.num_processes): envs += [make_env(mt_env_id_dic_selected[i], args.seed, j, log_dir)] ''' This envs is an intergration of all the running env''' envs = SubprocVecEnvMt(envs) num_processes_total = args.num_processes * len(mt_env_id_dic_selected) '''(1,128,128)''' obs_shape = envs.observation_space.shape #num_stack :number of frames to stack obs_shape = (obs_shape[0] * args.num_stack, *obs_shape[1:]) from arguments import is_restore if is_restore and args.save_dir: load_path = os.path.join(args.save_dir, args.algo) actor_critic =torch.load(os.path.join(load_path, args.env_name + ".pt")) # print ("restored previous model!") # print (actor_critic.Variable) # print (sss) else: if len(envs.observation_space.shape) == 3: actor_critic = CNNPolicy(obs_shape[0], envs.action_space) else: actor_critic = MLPPolicy(obs_shape[0], envs.action_space) if envs.action_space.__class__.__name__ == "Discrete": action_shape = 1 else: action_shape = envs.action_space.shape[0] if args.cuda: actor_critic.cuda() if args.algo == 'a2c': optimizer = optim.RMSprop(actor_critic.parameters(), args.lr, eps=args.eps, alpha=args.alpha) elif args.algo == 'ppo': optimizer = optim.Adam(actor_critic.parameters(), args.lr, eps=args.eps) elif args.algo == 'acktr': optimizer = KFACOptimizer(actor_critic) #'args.num_steps: number of forward steps in A2C #rollouts is an intergration of state\ reward\ next state\action and so on rollouts = RolloutStorage(args.num_steps, num_processes_total, obs_shape, envs.action_space) current_state = torch.zeros(num_processes_total, *obs_shape) ''' not sure about it''' def update_current_state(state): shape_dim0 = envs.observation_space.shape[0] # print (shape_dim0) # print (sss) state = torch.from_numpy(state).float() if args.num_stack > 1: current_state[:, :-shape_dim0] = current_state[:, shape_dim0:] current_state[:, -shape_dim0:] = state state = envs.reset() update_current_state(state) rollouts.states[0].copy_(current_state) # These variables are used to compute average rewards for all processes. episode_rewards = torch.zeros([num_processes_total, 1]) final_rewards = torch.zeros([num_processes_total, 1]) if args.cuda: current_state = current_state.cuda() rollouts.cuda() if args.algo == 'ppo': old_model = copy.deepcopy(actor_critic) from arguments import ewc, ewc_lambda, ewc_interval afs_per_m = [] afs_offset = [0.0]*gtn_M afs_loss_list = [] basic_loss_list = [] episode_reward_rec = 0.0 one = torch.FloatTensor([1]).cuda() mone = one * -1 '''for one whole game ''' for j in range(num_updates): for step in range(args.num_steps): if ewc == 1: try: states_store = torch.cat([states_store, rollouts.states[step].clone()], 0) except Exception as e: states_store = rollouts.states[step].clone() # Sample actions '''act fun refer to "observe it!"''' value, action = actor_critic.act(Variable(rollouts.states[step], volatile=True)) cpu_actions = action.data.squeeze(1).cpu().numpy() # Obser reward and next state state, reward, done = envs.step(cpu_actions) '''record the last 100 episodes rewards''' episode_reward_rec += reward episode_reward_rec = rec_last_100_epi_reward(episode_reward_rec,done) reward = torch.from_numpy(np.expand_dims(np.stack(reward), 1)).float() '''reward is shape of process_num_total, not batch-size''' # print ((reward).size()) # print (done) # print (sss) episode_rewards += reward ################ # rec_last_100_epi_reward(reward,done) # episode_reward_ppo += reward[0] # If done then clean the history of observations. final_rewards is used for compute after one whole num_step masks = torch.FloatTensor([[0.0] if done_ else [1.0] for done_ in done]) final_rewards *= masks final_rewards += (1 - masks) * episode_rewards episode_rewards *= masks if args.cuda: masks = masks.cuda() if current_state.dim() == 4: current_state *= masks.unsqueeze(2).unsqueeze(2) else: current_state *= masks update_current_state(state) rollouts.insert(step, current_state, action.data, value.data, reward, masks) next_value = actor_critic(Variable(rollouts.states[-1], volatile=True))[0].data if hasattr(actor_critic, 'obs_filter'): actor_critic.obs_filter.update(rollouts.states[:-1].view(-1, *obs_shape)) rollouts.compute_returns(next_value, args.use_gae, args.gamma, args.tau) if args.algo in ['a2c', 'acktr']: # reset gradient optimizer.zero_grad() # forward values, action_log_probs, dist_entropy, conv_list = actor_critic.evaluate_actions(Variable(rollouts.states[:-1].view(-1, *obs_shape)), Variable(rollouts.actions.view(-1, action_shape))) # pre-process values = values.view(args.num_steps, num_processes_total, 1) action_log_probs = action_log_probs.view(args.num_steps, num_processes_total, 1) # compute afs loss afs_per_m_temp, afs_loss = actor_critic.get_afs_per_m( action_log_probs=action_log_probs, conv_list=conv_list, ) if len(afs_per_m_temp)>0: afs_per_m += [afs_per_m_temp] if (afs_loss is not None) and (afs_loss.data.cpu().numpy()[0]!=0.0): afs_loss.backward(mone, retain_graph=True) afs_loss_list += [afs_loss.data.cpu().numpy()[0]] advantages = Variable(rollouts.returns[:-1]) - values value_loss = advantages.pow(2).mean() action_loss = -(Variable(advantages.data) * action_log_probs).mean() final_loss_basic = value_loss * args.value_loss_coef + action_loss - dist_entropy * args.entropy_coef ewc_loss = None if j != 0: if ewc == 1: ewc_loss = actor_critic.get_ewc_loss(lam=ewc_lambda) if ewc_loss is None: final_loss = final_loss_basic else: final_loss = final_loss_basic + ewc_loss # print (final_loss_basic.data.cpu().numpy()[0]) # final_loss_basic basic_loss_list += [final_loss_basic.data.cpu().numpy()[0]] final_loss.backward() if args.algo == 'a2c': nn.utils.clip_grad_norm(actor_critic.parameters(), args.max_grad_norm) optimizer.step() elif args.algo == 'ppo': advantages = rollouts.returns[:-1] - rollouts.value_preds[:-1] advantages = (advantages - advantages.mean()) / (advantages.std() + 1e-5) old_model.load_state_dict(actor_critic.state_dict()) if hasattr(actor_critic, 'obs_filter'): old_model.obs_filter = actor_critic.obs_filter for _ in range(args.ppo_epoch): sampler = BatchSampler(SubsetRandomSampler(range(num_processes_total * args.num_steps)), args.batch_size * num_processes_total, drop_last=False) for indices in sampler: indices = torch.LongTensor(indices) if args.cuda: indices = indices.cuda() states_batch = rollouts.states[:-1].view(-1, *obs_shape)[indices] actions_batch = rollouts.actions.view(-1, action_shape)[indices] return_batch = rollouts.returns[:-1].view(-1, 1)[indices] # Reshape to do in a single forward pass for all steps values, action_log_probs, dist_entropy, conv_list = actor_critic.evaluate_actions(Variable(states_batch), Variable(actions_batch)) _, old_action_log_probs, _, old_conv_list= old_model.evaluate_actions(Variable(states_batch, volatile=True), Variable(actions_batch, volatile=True)) ratio = torch.exp(action_log_probs - Variable(old_action_log_probs.data)) adv_targ = Variable(advantages.view(-1, 1)[indices]) surr1 = ratio * adv_targ surr2 = torch.clamp(ratio, 1.0 - args.clip_param, 1.0 + args.clip_param) * adv_targ action_loss = -torch.min(surr1, surr2).mean() # PPO's pessimistic surrogate (L^CLIP) value_loss = (Variable(return_batch) - values).pow(2).mean() optimizer.zero_grad() final_loss_basic = (value_loss + action_loss - dist_entropy * args.entropy_coef) basic_loss_list += [final_loss_basic.data.cpu().numpy()[0]] final_loss_basic.backward() optimizer.step() rollouts.states[0].copy_(rollouts.states[-1]) # if j % int(num_updates/2-10) == 0 and args.save_dir != "": if j % args.save_interval == 0 and args.save_dir != "": save_path = os.path.join(args.save_dir, args.algo) try: os.makedirs(save_path) except OSError: pass # A really ugly way to save a model to CPU save_model = actor_critic if args.cuda: save_model = copy.deepcopy(actor_critic).cpu() torch.save(save_model, os.path.join(save_path, args.env_name + ".pt")) import pickle with open(os.path.join(save_path, args.env_name + "_last_100_reward"), "wb") as f: pickle.dump(reward_dict, f) if j % args.log_interval == 0: print("Updates {}, num frames {}, mean/median reward {:.1f}/{:.1f}, min/max reward {:.1f}/{:.1f}, entropy {:.5f}, value loss {:.5f}, policy loss {:.5f}". format(j, (j + 1) * args.num_processes * args.num_steps, final_rewards.mean(), final_rewards.median(), final_rewards.min(), final_rewards.max(), -dist_entropy.data[0], value_loss.data[0], action_loss.data[0])) try: print("ewc loss {:.5f}". format(ewc_loss.data.cpu().numpy()[0])) except Exception as e: pass if j > 5 and j % args.vis_interval == 0 and args.vis: ''' load from the folder''' for ii in range(len(mt_env_id_dic_selected)): log_dir = args.log_dir+mt_env_id_dic_selected[ii]+'/' win[ii] = visdom_plot(viz, win[ii], log_dir, mt_env_id_dic_selected[ii], args.algo) plot_dic = reward_dict for plot_name in plot_dic.keys(): # if plot_name not in win_dic: # win_dic[plot_name] = None if plot_name in win_dic.keys(): if len(plot_dic[plot_name]) > 0: win_dic[plot_name] = viz.line( torch.from_numpy(np.asarray(plot_dic[plot_name])), win=win_dic[plot_name], opts=dict(title=break_line_html(exp+'>>'+plot_name)) ) else: win_dic[plot_name] = None if len(afs_per_m)>0: win_afs_per_m = viz.line( torch.from_numpy(np.asarray(afs_per_m)), win=win_afs_per_m, opts=dict(title=title_html+'>>afs') ) # print (basic_loss_list) '''a2c:len(basic_loss_list) is vis_interval+1. because j start from 0 ppo:len(basic_loss_list) is (vis_interval+1)*ppo_epoch_4*len(BatchSampler) ''' # print (len(basic_loss_list)) # print (ss) win_basic_loss = viz.line( torch.from_numpy(np.asarray(basic_loss_list)), win=win_basic_loss, opts=dict(title=title_html+'>>basic_loss') ) if len(afs_loss_list) > 0: win_afs_loss = viz.line( torch.from_numpy(np.asarray(afs_loss_list)), win=win_afs_loss, opts=dict(title=title_html+'>>afs_loss') ) from arguments import parameter_noise, parameter_noise_interval if parameter_noise == 1: if j % parameter_noise_interval == 0: actor_critic.parameter_noise() if ewc == 1: if j % ewc_interval == 0 or j==0: actor_critic.compute_fisher(states_store) states_store = None actor_critic.star()
def main(): print("#######") print("WARNING: All rewards are clipped or normalized so you need to use a monitor (see envs.py) or visdom plot to get true rewards") print("#######") os.environ['OMP_NUM_THREADS'] = '1' #os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID" #os.environ['CUDA_VISIBLE_DEVICES'] = "9" if args.vis: from visdom import Visdom viz = Visdom(port=args.port) win = None envs = [make_env(args.env_name, args.seed, i, args.log_dir, args.add_timestep) for i in range(args.num_processes)] if args.num_processes > 1: envs = SubprocVecEnv(envs) else: envs = DummyVecEnv(envs) if len(envs.observation_space.shape) == 1: envs = VecNormalize(envs) obs_shape = envs.observation_space.shape obs_shape = (obs_shape[0] * args.num_stack, *obs_shape[1:]) if len(envs.observation_space.shape) == 3: actor_critic = CNNPolicy(obs_shape[0], envs.action_space,args.hid_size, args.feat_size,args.recurrent_policy) else: assert not args.recurrent_policy, \ "Recurrent policy is not implemented for the MLP controller" actor_critic = MLPPolicy(obs_shape[0], envs.action_space) if envs.action_space.__class__.__name__ == "Discrete": action_shape = 1 else: action_shape = envs.action_space.shape[0] if args.use_cell: hs = HistoryCell(obs_shape[0], actor_critic.feat_size, 2*actor_critic.hidden_size, 1) ft = FutureCell(obs_shape[0], actor_critic.feat_size, 2 * actor_critic.hidden_size, 1) else: hs = History(obs_shape[0], actor_critic.feat_size, actor_critic.hidden_size, 2, 1) ft = Future(obs_shape[0], actor_critic.feat_size, actor_critic.hidden_size, 2, 1) if args.cuda: actor_critic=actor_critic.cuda() hs = hs.cuda() ft = ft.cuda() if args.algo == 'a2c': agent = algo.A2C_ACKTR(actor_critic, args.value_loss_coef, args.entropy_coef, lr=args.lr, eps=args.eps, alpha=args.alpha, max_grad_norm=args.max_grad_norm) elif args.algo == 'ppo': agent = algo.PPO(actor_critic, hs,ft,args.clip_param, args.ppo_epoch, args.num_mini_batch, args.value_loss_coef, args.entropy_coef, args.hf_loss_coef,ac_lr=args.lr,hs_lr=args.lr,ft_lr=args.lr, eps=args.eps, max_grad_norm=args.max_grad_norm, num_processes=args.num_processes, num_steps=args.num_steps, use_cell=args.use_cell, lenhs=args.lenhs,lenft=args.lenft, plan=args.plan, ac_intv=args.ac_interval, hs_intv=args.hs_interval, ft_intv=args.ft_interval ) elif args.algo == 'acktr': agent = algo.A2C_ACKTR(actor_critic, args.value_loss_coef, args.entropy_coef, acktr=True) rollouts = RolloutStorage(args.num_steps, args.num_processes, obs_shape, envs.action_space, actor_critic.state_size, feat_size=512) current_obs = torch.zeros(args.num_processes, *obs_shape) def update_current_obs(obs): shape_dim0 = envs.observation_space.shape[0] obs = torch.from_numpy(obs).float() if args.num_stack > 1: current_obs[:, :-shape_dim0] = current_obs[:, shape_dim0:] current_obs[:, -shape_dim0:] = obs obs = envs.reset() update_current_obs(obs) rollouts.observations[0].copy_(current_obs) if args.cuda: current_obs = current_obs.cuda() rollouts.cuda() rec_x = [] rec_y = [] file = open('./rec/' + args.env_name + '_' + args.method_name + '.txt', 'w') hs_info = torch.zeros(args.num_processes, 2 * actor_critic.hidden_size).cuda() hs_ind = torch.IntTensor(args.num_processes, 1).zero_() epinfobuf = deque(maxlen=100) start_time = time.time() for j in range(num_updates): print('begin sample, time {}'.format(time.strftime("%Hh %Mm %Ss", time.gmtime(time.time() - start_time)))) for step in range(args.num_steps): # Sample actions with torch.no_grad(): rollouts.feat[step]=actor_critic.get_feat(rollouts.observations[step]) if args.use_cell: for i in range(args.num_processes): h = torch.zeros(1, 2 * actor_critic.hid_size).cuda() c = torch.zeros(1, 2 * actor_critic.hid_size).cuda() start_ind = max(hs_ind[i],step+1-args.lenhs) for ind in range(start_ind,step+1): h,c=hs(rollouts.feat[ind,i].unsqueeze(0),h,c) hs_info[i,:]=h.view(1,2*actor_critic.hid_size) del h,c gc.collect() else: for i in range(args.num_processes): start_ind = max(hs_ind[i], step + 1 - args.lenhs) hs_info[i,:]=hs(rollouts.feat[start_ind:step+1,i]) hidden_feat=actor_critic.cat(rollouts.feat[step],hs_info) value, action, action_log_prob, states = actor_critic.act( hidden_feat, rollouts.states[step]) cpu_actions = action.data.squeeze(1).cpu().numpy() # Obser reward and next obs obs, reward, done, infos = envs.step(cpu_actions) for info in infos: maybeepinfo = info.get('episode') if maybeepinfo: epinfobuf.extend([maybeepinfo['r']]) reward = torch.from_numpy(np.expand_dims(np.stack(reward), 1)).float() masks = torch.FloatTensor([[0.0] if done_ else [1.0] for done_ in done]) hs_ind = ((1-masks)*(step+1)+masks*hs_ind.float()).int() if args.cuda: masks = masks.cuda() if current_obs.dim() == 4: current_obs *= masks.unsqueeze(2).unsqueeze(2) else: current_obs *= masks update_current_obs(obs) rollouts.insert(current_obs, hs_ind,states.data, action.data, action_log_prob.data, value.data, reward, masks) with torch.no_grad(): rollouts.feat[-1] = actor_critic.get_feat(rollouts.observations[-1]) if args.use_cell: for i in range(args.num_processes): h = torch.zeros(1, 2 * actor_critic.hid_size).cuda() c = torch.zeros(1, 2 * actor_critic.hid_size).cuda() start = max(hs_ind[i], step + 1 - args.lenhs) for ind in range(start, step + 1): h, c = hs(rollouts.feat[ind, i].unsqueeze(0), h, c) hs_info[i, :] = h.view(1, 2 * actor_critic.hid_size) del h,c else: for i in range(args.num_processes): start_ind = max(hs_ind[i], step + 1 - args.lenhs) hs_info[i, :] = hs(rollouts.feat[start_ind:step + 1, i]) hidden_feat = actor_critic.cat(rollouts.feat[-1],hs_info) next_value = actor_critic.get_value(hidden_feat).detach() rollouts.compute_returns(next_value, args.use_gae, args.gamma, args.tau) rollouts.compute_ft_ind() print('begin update, time {}'.format(time.strftime("%Hh %Mm %Ss", time.gmtime(time.time() - start_time)))) value_loss, action_loss, dist_entropy = agent.update(rollouts) print('end update, time {}'.format(time.strftime("%Hh %Mm %Ss", time.gmtime(time.time() - start_time)))) rollouts.after_update() if j % args.save_interval == 0 and args.save_dir != "": save_path = os.path.join(args.save_dir, args.algo) try: os.makedirs(save_path) except OSError: pass # A really ugly way to save a model to CPU save_model = actor_critic if args.cuda: save_model = copy.deepcopy(actor_critic).cpu() save_model = [save_model, hasattr(envs, 'ob_rms') and envs.ob_rms or None] torch.save(save_model, os.path.join(save_path, args.env_name + ".pt")) if j % args.log_interval == 0: end = time.time() total_num_steps = (j + 1) * args.num_processes * args.num_steps v_mean,v_median,v_min,v_max = safe(epinfobuf) print("Updates {}, num timesteps {},time {}, FPS {}, mean/median reward {:.1f}/{:.1f}, min/max reward {:.1f}/{:.1f}, entropy {:.5f}, value loss {:.5f}, policy loss {:.5f}". format(j, total_num_steps, time.strftime("%Hh %Mm %Ss", time.gmtime(time.time() - start_time)), int(total_num_steps / (end - start_time)), v_mean, v_median, v_min, v_max, dist_entropy, value_loss, action_loss)) if not (v_mean==np.nan): rec_x.append(total_num_steps) rec_y.append(v_mean) file.write(str(total_num_steps)) file.write(' ') file.writelines(str(v_mean)) file.write('\n') if args.vis and j % args.vis_interval == 0: try: # Sometimes monitor doesn't properly flush the outputs win = visdom_plot(viz, win, args.log_dir, args.env_name, args.algo, args.num_frames) except IOError: pass plot_line(rec_x, rec_y, './imgs/' + args.env_name + '_' + args.method_name + '.png', args.method_name, args.env_name, args.num_frames) file.close()
def main(): print("###############################################################") print("#################### VIZDOOM LEARNER START ####################") print("###############################################################") save_path = os.path.join(args.save_dir, "a2c") num_updates = int(args.num_frames) // args.num_steps // args.num_processes reward_name = "" if args.roe: reward_name = "_event" scenario_name = args.config_path.split("/")[1].split(".")[0] print("############### " + scenario_name + " ###############") log_file_name = "vizdoom_" + scenario_name + reward_name + "_" + str( args.agent_id) + ".log" log_event_file_name = "vizdoom_" + scenario_name + reward_name + "_" + str( args.agent_id) + ".eventlog" log_event_reward_file_name = "vizdoom_" + scenario_name + reward_name + "_" + str( args.agent_id) + ".eventrewardlog" start_updates = 0 start_step = 0 best_final_rewards = -1000000.0 os.environ['OMP_NUM_THREADS'] = '1' global envs es = [ make_env(i, args.config_path, visual=args.visual, bots=args.bots) for i in range(args.num_processes) ] envs = VecEnv([es[i] for i in range(args.num_processes)]) obs_shape = envs.observation_space_shape obs_shape = (obs_shape[0] * args.num_stack, *obs_shape[1:]) if args.resume: actor_critic = torch.load( os.path.join(save_path, log_file_name + ".pt")) filename = glob.glob(os.path.join(args.log_dir, log_file_name))[0] if args.roe: e # TODO: Load event buffer with open(filename) as file: lines = file.readlines() start_updates = (int)(lines[-1].strip().split(",")[0]) start_steps = (int)(lines[-1].strip().split(",")[1]) num_updates += start_updates else: if not args.debug: try: os.makedirs(args.log_dir) except OSError: files = glob.glob(os.path.join(args.log_dir, log_file_name)) for f in files: os.remove(f) with open(log_file_name, "w") as myfile: myfile.write("") files = glob.glob( os.path.join(args.log_dir, log_event_file_name)) for f in files: os.remove(f) with open(log_event_file_name, "w") as myfile: myfile.write("") files = glob.glob( os.path.join(args.log_dir, log_event_reward_file_name)) for f in files: os.remove(f) with open(log_event_reward_file_name, "w") as myfile: myfile.write("") actor_critic = CNNPolicy(obs_shape[0], envs.action_space_shape) action_shape = 1 if args.cuda: actor_critic.cuda() optimizer = optim.RMSprop(actor_critic.parameters(), args.lr, eps=args.eps, alpha=args.alpha) rollouts = RolloutStorage(args.num_steps, args.num_processes, obs_shape, envs.action_space_shape) current_obs = torch.zeros(args.num_processes, *obs_shape) def update_current_obs(obs): shape_dim0 = envs.observation_space_shape[0] obs = torch.from_numpy(obs).float() if args.num_stack > 1: current_obs[:, :-shape_dim0] = current_obs[:, shape_dim0:] current_obs[:, -shape_dim0:] = obs obs = envs.reset() update_current_obs(obs) last_game_vars = [] for i in range(args.num_processes): last_game_vars.append(np.zeros(args.num_events)) rollouts.observations[0].copy_(current_obs) # These variables are used to compute average rewards for all processes. episode_rewards = torch.zeros([args.num_processes, 1]) final_rewards = torch.zeros([args.num_processes, 1]) episode_intrinsic_rewards = torch.zeros([args.num_processes, 1]) final_intrinsic_rewards = torch.zeros([args.num_processes, 1]) episode_events = torch.zeros([args.num_processes, args.num_events]) final_events = torch.zeros([args.num_processes, args.num_events]) if args.cuda: current_obs = current_obs.cuda() rollouts.cuda() # Create event buffer if args.qd: event_buffer = EventBufferSQLProxy(args.num_events, args.capacity, args.exp_id, args.agent_id) elif not args.resume: event_buffer = EventBuffer(args.num_events, args.capacity) else: event_buffer = pickle.load( open(log_file_name + "_event_buffer_temp.p", "rb")) event_episode_rewards = [] start = time.time() for j in np.arange(start_updates, num_updates): for step in range(args.num_steps): value, action = actor_critic.act( Variable(rollouts.observations[step], volatile=True)) cpu_actions = action.data.squeeze(1).cpu().numpy() obs, reward, done, info, events = envs.step(cpu_actions) intrinsic_reward = [] # Fix broken rewards - upscale for i in range(len(reward)): if scenario_name in ["deathmatch", "my_way_home"]: reward[i] *= 100 if scenario_name == "deadly_corridor": reward[i] = 1 if events[i][2] >= 1 else 0 for e in events: if args.roe: intrinsic_reward.append(event_buffer.intrinsic_reward(e)) else: r = reward[len(intrinsic_reward)] intrinsic_reward.append(r) reward = torch.from_numpy(np.expand_dims(np.stack(reward), 1)).float() intrinsic_reward = torch.from_numpy( np.expand_dims(np.stack(intrinsic_reward), 1)).float() #events = torch.from_numpy(np.expand_dims(np.stack(events), args.num_events)).float() events = torch.from_numpy(events).float() episode_rewards += reward episode_intrinsic_rewards += intrinsic_reward episode_events += events # Event stats event_rewards = [] for ei in range(0, args.num_events): ev = np.zeros(args.num_events) ev[ei] = 1 er = event_buffer.intrinsic_reward(ev) event_rewards.append(er) event_episode_rewards.append(event_rewards) # If done then clean the history of observations. masks = torch.FloatTensor([[0.0] if done_ else [1.0] for done_ in done]) final_rewards *= masks final_intrinsic_rewards *= masks final_events *= masks final_rewards += (1 - masks) * episode_rewards final_intrinsic_rewards += (1 - masks) * episode_intrinsic_rewards final_events += (1 - masks) * episode_events for i in range(args.num_processes): if done[i]: event_buffer.record_events(np.copy( final_events[i].numpy()), frame=j * args.num_steps) episode_rewards *= masks episode_intrinsic_rewards *= masks episode_events *= masks if args.cuda: masks = masks.cuda() if current_obs.dim() == 4: current_obs *= masks.unsqueeze(2).unsqueeze(2) else: current_obs *= masks update_current_obs(obs) rollouts.insert(step, current_obs, action.data, value.data, intrinsic_reward, masks) final_episode_reward = np.mean(event_episode_rewards, axis=0) event_episode_rewards = [] next_value = actor_critic( Variable(rollouts.observations[-1], volatile=True))[0].data if hasattr(actor_critic, 'obs_filter'): actor_critic.obs_filter.update(rollouts.observations[:-1].view( -1, *obs_shape)) rollouts.compute_returns(next_value, args.use_gae, args.gamma, args.tau) values, action_log_probs, dist_entropy = actor_critic.evaluate_actions( Variable(rollouts.observations[:-1].view(-1, *obs_shape)), Variable(rollouts.actions.view(-1, action_shape))) values = values.view(args.num_steps, args.num_processes, 1) action_log_probs = action_log_probs.view(args.num_steps, args.num_processes, 1) advantages = Variable(rollouts.returns[:-1]) - values value_loss = advantages.pow(2).mean() action_loss = -(Variable(advantages.data) * action_log_probs).mean() optimizer.zero_grad() (value_loss * args.value_loss_coef + action_loss - dist_entropy * args.entropy_coef).backward() nn.utils.clip_grad_norm(actor_critic.parameters(), args.max_grad_norm) optimizer.step() rollouts.observations[0].copy_(rollouts.observations[-1]) if final_rewards.mean() > best_final_rewards and not args.debug: try: os.makedirs(save_path) except OSError: pass best_final_rewards = final_rewards.mean() save_model = actor_critic if args.cuda: save_model = copy.deepcopy(actor_critic).cpu() torch.save( save_model, os.path.join(save_path, log_file_name.split(".log")[0] + ".pt")) if j % args.save_interval == 0 and args.save_dir != "" and not args.debug: try: os.makedirs(save_path) except OSError: pass # A really ugly way to save a model to CPU save_model = actor_critic if args.cuda: save_model = copy.deepcopy(actor_critic).cpu() torch.save(save_model, os.path.join(save_path, log_file_name + "_temp.pt")) if isinstance(event_buffer, EventBuffer): pickle.dump(event_buffer, open(log_file_name + "_event_buffer_temp.p", "wb")) if j % args.log_interval == 0: envs.log() end = time.time() total_num_steps = (j + 1) * args.num_processes * args.num_steps log = "Updates {}, num timesteps {}, FPS {}, mean/max reward {:.5f}/{:.5f}, mean/max intrinsic reward {:.5f}/{:.5f}"\ .format(j, total_num_steps, int(total_num_steps / (end - start)), final_rewards.mean(), final_rewards.max(), final_intrinsic_rewards.mean(), final_intrinsic_rewards.max() ) log_to_file = "{}, {}, {:.5f}, {:.5f}, {:.5f}, {:.5f}\n" \ .format(j, total_num_steps, final_rewards.mean(), final_rewards.std(), final_intrinsic_rewards.mean(), final_intrinsic_rewards.std()) log_to_event_file = ','.join( map(str, event_buffer.get_event_mean().tolist())) + "\n" log_to_event_reward_file = ','.join( map(str, event_buffer.get_event_rewards().tolist())) + "\n" print(log) print(log_to_event_file) # Save to files with open(log_file_name, "a") as myfile: myfile.write(log_to_file) with open(log_event_file_name, "a") as myfile: myfile.write(str(total_num_steps) + "," + log_to_event_file) with open(log_event_reward_file_name, "a") as myfile: myfile.write( str(total_num_steps) + "," + log_to_event_reward_file) envs.close() time.sleep(5)
def main(): print("######") print("HELLO! Returns start with infinity values") print("######") os.environ['OMP_NUM_THREADS'] = '1' if args.random_task: env_params = { 'wt': np.round(np.random.uniform(0.5, 1.0), 2), 'x': np.round(np.random.uniform(-0.1, 0.1), 2), 'y': np.round(np.random.uniform(-0.1, 0.1), 2), 'z': np.round(np.random.uniform(0.15, 0.2), 2), } else: env_params = { 'wt': args.euclidean_weight, 'x': args.goal_x, 'y': args.goal_y, 'z': args.goal_z, } envs = [make_env(args.env_name, args.seed, i, args.log_dir, **env_params) for i in range(args.num_processes)] if args.num_processes > 1: envs = SubprocVecEnv(envs) else: envs = DummyVecEnv(envs) envs = VecNormalize(envs, ob=False) obs_shape = envs.observation_space.shape obs_shape = (obs_shape[0] * args.num_stack, *obs_shape[1:]) if len(envs.observation_space.shape) == 3: actor_critic = CNNPolicy(obs_shape[0], envs.action_space, args.recurrent_policy) else: assert not args.recurrent_policy, \ "Recurrent policy is not implemented for the MLP controller" actor_critic = MLPPolicy(obs_shape[0], envs.action_space) if envs.action_space.__class__.__name__ == "Discrete": action_shape = 1 else: action_shape = envs.action_space.shape[0] if args.cuda: actor_critic.cuda() if args.algo == 'a2c': optimizer = optim.RMSprop(actor_critic.parameters(), args.lr, eps=args.eps, alpha=args.alpha) elif args.algo == 'ppo': optimizer = optim.Adam(actor_critic.parameters(), args.lr, eps=args.eps) elif args.algo == 'acktr': optimizer = KFACOptimizer(actor_critic) rollouts = RolloutStorage(args.num_steps, args.num_processes, obs_shape, envs.action_space, actor_critic.state_size) current_obs = torch.zeros(args.num_processes, *obs_shape) def update_current_obs(obs): shape_dim0 = envs.observation_space.shape[0] obs = torch.from_numpy(obs).float() if args.num_stack > 1: current_obs[:, :-shape_dim0] = current_obs[:, shape_dim0:] current_obs[:, -shape_dim0:] = obs obs = envs.reset() update_current_obs(obs) rollouts.observations[0].copy_(current_obs) # These variables are used to compute average rewards for all processes. episode_rewards = torch.zeros([args.num_processes, 1]) final_rewards = torch.zeros([args.num_processes, 1]) if args.cuda: current_obs = current_obs.cuda() rollouts.cuda() actor_critic.input_norm.update(rollouts.observations[0]) last_return = -np.inf best_return = -np.inf best_models = None start = time.time() for j in range(num_updates): for step in range(args.num_steps): # Sample actions value, action, action_log_prob, states = actor_critic.act(Variable(rollouts.observations[step], volatile=True), Variable(rollouts.states[step], volatile=True), Variable(rollouts.masks[step], volatile=True)) cpu_actions = action.data.squeeze(1).cpu().numpy() # Obser reward and next obs obs, reward, done, info = envs.step(cpu_actions) reward = torch.from_numpy(np.expand_dims(np.stack(reward), 1)).float() episode_rewards += reward # If done then clean the history of observations. masks = torch.FloatTensor([[0.0] if done_ else [1.0] for done_ in done]) final_rewards *= masks final_rewards += (1 - masks) * episode_rewards episode_rewards *= masks if args.cuda: masks = masks.cuda() if current_obs.dim() == 4: current_obs *= masks.unsqueeze(2).unsqueeze(2) else: current_obs *= masks update_current_obs(obs) rollouts.insert(step, current_obs, states.data, action.data, action_log_prob.data, value.data, reward, masks) actor_critic.input_norm.update(rollouts.observations[step + 1]) next_value = actor_critic(Variable(rollouts.observations[-1], volatile=True), Variable(rollouts.states[-1], volatile=True), Variable(rollouts.masks[-1], volatile=True))[0].data rollouts.compute_returns(next_value, args.use_gae, args.gamma, args.tau) if args.algo in ['a2c', 'acktr']: values, action_log_probs, dist_entropy, states = actor_critic.evaluate_actions(Variable(rollouts.observations[:-1].view(-1, *obs_shape)), Variable(rollouts.states[0].view(-1, actor_critic.state_size)), Variable(rollouts.masks[:-1].view(-1, 1)), Variable(rollouts.actions.view(-1, action_shape))) values = values.view(args.num_steps, args.num_processes, 1) action_log_probs = action_log_probs.view(args.num_steps, args.num_processes, 1) advantages = Variable(rollouts.returns[:-1]) - values value_loss = advantages.pow(2).mean() action_loss = -(Variable(advantages.data) * action_log_probs).mean() if args.algo == 'acktr' and optimizer.steps % optimizer.Ts == 0: # Sampled fisher, see Martens 2014 actor_critic.zero_grad() pg_fisher_loss = -action_log_probs.mean() value_noise = Variable(torch.randn(values.size())) if args.cuda: value_noise = value_noise.cuda() sample_values = values + value_noise vf_fisher_loss = -(values - Variable(sample_values.data)).pow(2).mean() fisher_loss = pg_fisher_loss + vf_fisher_loss optimizer.acc_stats = True fisher_loss.backward(retain_graph=True) optimizer.acc_stats = False optimizer.zero_grad() (value_loss * args.value_loss_coef + action_loss - dist_entropy * args.entropy_coef).backward() if args.algo == 'a2c': nn.utils.clip_grad_norm(actor_critic.parameters(), args.max_grad_norm) optimizer.step() elif args.algo == 'ppo': advantages = rollouts.returns[:-1] - rollouts.value_preds[:-1] advantages = (advantages - advantages.mean()) / (advantages.std() + 1e-5) for e in range(args.ppo_epoch): if args.recurrent_policy: data_generator = rollouts.recurrent_generator(advantages, args.num_mini_batch) else: data_generator = rollouts.feed_forward_generator(advantages, args.num_mini_batch) for sample in data_generator: observations_batch, states_batch, actions_batch, \ return_batch, masks_batch, old_action_log_probs_batch, \ adv_targ = sample # Reshape to do in a single forward pass for all steps values, action_log_probs, dist_entropy, states = actor_critic.evaluate_actions(Variable(observations_batch), Variable(states_batch), Variable(masks_batch), Variable(actions_batch)) adv_targ = Variable(adv_targ) ratio = torch.exp(action_log_probs - Variable(old_action_log_probs_batch)) surr1 = ratio * adv_targ surr2 = torch.clamp(ratio, 1.0 - args.clip_param, 1.0 + args.clip_param) * adv_targ action_loss = -torch.min(surr1, surr2).mean() # PPO's pessimistic surrogate (L^CLIP) value_loss = (Variable(return_batch) - values).pow(2).mean() optimizer.zero_grad() (value_loss + action_loss - dist_entropy * args.entropy_coef).backward() nn.utils.clip_grad_norm(actor_critic.parameters(), args.max_grad_norm) optimizer.step() rollouts.after_update() if args.vis and j % args.vis_interval == 0: last_return = plot(logger, args.log_dir) if last_return > best_return: best_return = last_return try: os.makedirs(os.path.dirname(args.save_path)) except OSError: pass info = { 'return': best_return, 'reward_norm': np.sqrt(envs.ret_rms.var + envs.epsilon) } # A really ugly way to save a model to CPU save_model = actor_critic if args.cuda: save_model = copy.deepcopy(actor_critic).cpu() torch.save((save_model, env_params, info), args.save_path) if j % args.log_interval == 0: end = time.time() total_num_steps = (j + 1) * args.num_processes * args.num_steps print("Updates {}, num timesteps {}, FPS {}, average return {:.5f}, best_return {:.5f}, value loss {:.5f}, policy loss {:.5f}". format(j, total_num_steps, int(total_num_steps / (end - start)), last_return, best_return, value_loss.data[0], action_loss.data[0]))
def main(): print("#######") print( "WARNING: All rewards are clipped or normalized so you need to use a monit`or (see envs.py) or visdom plot to get true rewards" ) print("#######") os.environ['OMP_NUM_THREADS'] = '1' # logger = Logger(algorithm_name = args.algo, environment_name = args.env_name, folder = args.folder) # logger.save_args(args) # print ("---------------------------------------") # print ('Saving to', logger.save_folder) # print ("---------------------------------------") if args.vis: from visdom import Visdom viz = Visdom(port=args.port) win = None envs = [ make_env(args.env_name, args.seed, i, args.log_dir) for i in range(args.num_processes) ] if args.num_processes > 1: envs = SubprocVecEnv(envs) else: envs = DummyVecEnv(envs) if len(envs.observation_space.shape) == 1: envs = VecNormalize(envs) obs_shape = envs.observation_space.shape obs_shape = (obs_shape[0] * args.num_stack, *obs_shape[1:]) if len(envs.observation_space.shape) == 3: actor_critic = CNNPolicy(obs_shape[0], envs.action_space, args.recurrent_policy) target_actor_critic = CNNPolicy(obs_shape[0], envs.action_space, args.recurrent_policy) else: actor_critic = MLPPolicy(obs_shape[0], envs.action_space) target_actor_critic = MLPPolicy(obs_shape[0], envs.action_space) for param, target_param in zip(actor_critic.parameters(), target_actor_critic.parameters()): target_param.data.copy_(param.data) if envs.action_space.__class__.__name__ == "Discrete": action_shape = 1 else: action_shape = envs.action_space.shape[0] if args.cuda: actor_critic.cuda() actor_regularizer_criterion = nn.KLDivLoss() optimizer = optim.RMSprop(actor_critic.parameters(), args.lr, eps=args.eps, alpha=args.alpha) rollouts = RolloutStorage(args.num_steps, args.num_processes, obs_shape, envs.action_space, actor_critic.state_size) current_obs = torch.zeros(args.num_processes, *obs_shape) def update_current_obs(obs): shape_dim0 = envs.observation_space.shape[0] obs = torch.from_numpy(obs).float() if args.num_stack > 1: current_obs[:, :-shape_dim0] = current_obs[:, shape_dim0:] current_obs[:, -shape_dim0:] = obs obs = envs.reset() update_current_obs(obs) rollouts.observations[0].copy_(current_obs) # These variables are used to compute average rewards for all processes. episode_rewards = torch.zeros([args.num_processes, 1]) final_rewards = torch.zeros([args.num_processes, 1]) if args.cuda: current_obs = current_obs.cuda() rollouts.cuda() start = time.time() for j in range(num_updates): for step in range(args.num_steps): # Sample actions value, action, action_log_prob, states = actor_critic.act( Variable(rollouts.observations[step], volatile=True), Variable(rollouts.states[step], volatile=True), Variable(rollouts.masks[step], volatile=True)) cpu_actions = action.data.squeeze(1).cpu().numpy() # Obser reward and next obs obs, reward, done, info = envs.step(cpu_actions) reward = torch.from_numpy(np.expand_dims(np.stack(reward), 1)).float() episode_rewards += reward # If done then clean the history of observations. masks = torch.FloatTensor([[0.0] if done_ else [1.0] for done_ in done]) final_rewards *= masks final_rewards += (1 - masks) * episode_rewards episode_rewards *= masks if args.cuda: masks = masks.cuda() if current_obs.dim() == 4: current_obs *= masks.unsqueeze(2).unsqueeze(2) else: current_obs *= masks update_current_obs(obs) rollouts.insert(step, current_obs, states.data, action.data, action_log_prob.data, value.data, reward, masks) next_value = actor_critic( Variable(rollouts.observations[-1], volatile=True), Variable(rollouts.states[-1], volatile=True), Variable(rollouts.masks[-1], volatile=True))[0].data rollouts.compute_returns(next_value, args.use_gae, args.gamma, args.tau) values, action_log_probs, dist_entropy, states = actor_critic.evaluate_actions( Variable(rollouts.observations[:-1].view(-1, *obs_shape)), Variable(rollouts.states[0].view(-1, actor_critic.state_size)), Variable(rollouts.masks[:-1].view(-1, 1)), Variable(rollouts.actions.view(-1, action_shape))) """ Used for KL Constraint in case of Continuous Action Stochastic Policies """ # target_values, target_action_log_probs, target_dist_entropy, target_states, target_action_mean, target_action_std = target_actor_critic.evaluate_actions_mean_and_std(Variable(rollouts.observations[:-1].view(-1, *obs_shape)), # Variable(rollouts.states[0].view(-1, actor_critic.state_size)), # Variable(rollouts.masks[:-1].view(-1, 1)), # Variable(rollouts.actions.view(-1, action_shape))) # actor_regularizer_loss = (torch.log(action_std/target_action_std) + (action_std.pow(2) + (action_mean - target_action_mean).pow(2))/(2*target_action_std.pow(2)) - 0.5) values = values.view(args.num_steps, args.num_processes, 1) action_log_probs = action_log_probs.view(args.num_steps, args.num_processes, 1) advantages = Variable(rollouts.returns[:-1]) - values value_loss = advantages.pow(2).mean() ### Loss with regularizer added ##action_loss = -(Variable(advantages.data) * action_log_probs).mean() + args.actor_lambda * actor_regularizer_loss.mean(0).sum() action_loss = -(Variable(advantages.data) * action_log_probs).mean() optimizer.zero_grad() total_loss = value_loss * args.value_loss_coef + action_loss - dist_entropy * args.entropy_coef total_loss.backward() nn.utils.clip_grad_norm(actor_critic.parameters(), args.max_grad_norm) optimizer.step() ## Exponential average for target updates #if (j%args.target_update_interval == 0): # for param, target_param in zip(actor_critic.parameters(), target_actor_critic.parameters()): # target_param.data.copy_(args.target_tau * param.data + (1 - args.target_tau) * target_param.data) rollouts.after_update() if j % args.save_interval == 0 and args.save_dir != "": save_path = os.path.join(args.save_dir, args.algo) try: os.makedirs(save_path) except OSError: pass # A really ugly way to save a model to CPU save_model = actor_critic if args.cuda: save_model = copy.deepcopy(actor_critic).cpu() save_model = [ save_model, hasattr(envs, 'ob_rms') and envs.ob_rms or None ] torch.save(save_model, os.path.join(save_path, args.env_name + ".pt")) if j % args.log_interval == 0: end = time.time() total_num_steps = (j + 1) * args.num_processes * args.num_steps print( "Updates {}, num timesteps {}, FPS {}, mean/median reward {:.1f}/{:.1f}, min/max reward {:.1f}/{:.1f}, entropy {:.5f}, value loss {:.5f}, policy loss {:.5f}" .format(j, total_num_steps, int(total_num_steps / (end - start)), final_rewards.mean(), final_rewards.median(), final_rewards.min(), final_rewards.max(), dist_entropy.data[0], value_loss.data[0], action_loss.data[0])) final_rewards_mean = [final_rewards.mean()] final_rewards_median = [final_rewards.median()] final_rewards_min = [final_rewards.min()] final_rewards_max = [final_rewards.max()] # logger.record_data(final_rewards_mean, final_rewards_median, final_rewards_min, final_rewards_max) # logger.save() if args.vis and j % args.vis_interval == 0: try: # Sometimes monitor doesn't properly flush the outputs win = visdom_plot(viz, win, args.log_dir, args.env_name, args.algo) except IOError: pass
def main(): print("#######") print("WARNING: All rewards are clipped so you need to use a monitor (see envs.py) or visdom plot to get true rewards") print("#######") os.environ['OMP_NUM_THREADS'] = '1' print (args.cuda) print (args.num_steps) print (args.num_processes) print (args.lr) print (args.eps) print (args.alpha) print (args.use_gae) print (args.gamma) print (args.tau) print (args.value_loss_coef) print (args.entropy_coef) # fsdaf # Create environment envs = SubprocVecEnv([ make_env(args.env_name, args.seed, i, args.log_dir) for i in range(args.num_processes) ]) obs_shape = envs.observation_space.shape obs_shape = (obs_shape[0] * args.num_stack, *obs_shape[1:]) if len(envs.observation_space.shape) == 3: actor_critic = CNNPolicy(obs_shape[0], envs.action_space) else: actor_critic = MLPPolicy(obs_shape[0], envs.action_space) if envs.action_space.__class__.__name__ == "Discrete": action_shape = 1 else: action_shape = envs.action_space.shape[0] # action_shape = action_shape # shape_dim0 = envs.observation_space.shape[0] # if args.cuda: # dtype = torch.cuda.FloatTensor # else: # dtype = torch.FloatTensor hparams = {'cuda':args.cuda, 'num_steps':args.num_steps, 'num_processes':args.num_processes, 'obs_shape':obs_shape, 'lr':args.lr, 'eps':args.eps, 'alpha':args.alpha, 'use_gae':args.use_gae, 'gamma':args.gamma, 'tau':args.tau, 'value_loss_coef':args.value_loss_coef, 'entropy_coef':args.entropy_coef} # Create agent # agent = a2c(envs, hparams) # rollouts = RolloutStorage(self.num_steps, self.num_processes, self.obs_shape, envs.action_space) #it has a self.state that is [steps, processes, obs] #steps is used to compute expected reward if args.cuda: actor_critic.cuda() # rollouts.cuda() optimizer = optim.RMSprop(actor_critic.parameters(), hparams['lr'], eps=hparams['eps'], alpha=hparams['alpha']) rollouts = RolloutStorage(args.num_steps, args.num_processes, obs_shape, envs.action_space) # Init state current_state = torch.zeros(args.num_processes, *obs_shape)#.type(dtype) def update_current_state(state):#, shape_dim0): shape_dim0 = envs.observation_space.shape[0] state = torch.from_numpy(state).float() if args.num_stack > 1: current_state[:, :-shape_dim0] = current_state[:, shape_dim0:] current_state[:, -shape_dim0:] = state # return current_state state = envs.reset() update_current_state(state)#, shape_dim0) # agent.insert_first_state(current_state) rollouts.states[0].copy_(current_state) #set the first state to current state # These are used to compute average rewards for all processes. episode_rewards = torch.zeros([args.num_processes, 1]) final_rewards = torch.zeros([args.num_processes, 1]) if args.cuda: current_state = current_state.cuda()#type(dtype) # if args.cuda: rollouts.cuda() #Begin training start = time.time() for j in range(num_updates): for step in range(args.num_steps): # Act # action, value = agent.act(Variable(agent.rollouts.states[step], volatile=True)) value, action = actor_critic.act(Variable(rollouts.states[step], volatile=True)) cpu_actions = action.data.squeeze(1).cpu().numpy() # Observe reward and next state state, reward, done, info = envs.step(cpu_actions) # state:[nProcesss, ndims, height, width] # Record rewards # reward, masks, final_rewards, episode_rewards, current_state = update_rewards(reward, done, final_rewards, episode_rewards, current_state) reward = torch.from_numpy(np.expand_dims(np.stack(reward), 1)).float() episode_rewards += reward # If done then clean the history of observations. # these final rewards are only used for printing. but the mask is used in the storage, dont know why yet # oh its just clearing the env that finished, and resetting its episode_reward masks = torch.FloatTensor([[0.0] if done_ else [1.0] for done_ in done]) #if an env is done final_rewards *= masks final_rewards += (1 - masks) * episode_rewards episode_rewards *= masks if args.cuda: masks = masks.cuda() if current_state.dim() == 4: current_state *= masks.unsqueeze(2).unsqueeze(2) else: current_state *= masks # return reward, masks, final_rewards, episode_rewards, current_state # Update state update_current_state(state)#, shape_dim0) # Agent record step # agent.insert_data(step, current_state, action.data, value.data, reward, masks) rollouts.insert(step, current_state, action.data, value.data, reward, masks) #Optimize agent # agent.update() next_value = actor_critic(Variable(rollouts.states[-1], volatile=True))[0].data # use last state to make prediction of next value if hasattr(actor_critic, 'obs_filter'): actor_critic.obs_filter.update(rollouts.states[:-1].view(-1, *obs_shape)) #not sure what this is rollouts.compute_returns(next_value, args.use_gae, args.gamma, args.tau) # this computes R = r + r+ ...+ V(t) for each step values, action_log_probs, dist_entropy = actor_critic.evaluate_actions( Variable(rollouts.states[:-1].view(-1, *obs_shape)), Variable(rollouts.actions.view(-1, action_shape))) # I think this aciton log prob could have been computed and stored earlier # and didnt we already store the value prediction??? values = values.view(args.num_steps, args.num_processes, 1) action_log_probs = action_log_probs.view(args.num_steps, args.num_processes, 1) advantages = Variable(rollouts.returns[:-1]) - values value_loss = advantages.pow(2).mean() action_loss = -(Variable(advantages.data) * action_log_probs).mean() optimizer.zero_grad() (value_loss * args.value_loss_coef + action_loss - dist_entropy * args.entropy_coef).backward() optimizer.step() rollouts.states[0].copy_(rollouts.states[-1]) # the first state is now the last state of the previous # #Save model # if j % args.save_interval == 0 and args.save_dir != "": # save_path = os.path.join(args.save_dir, args.algo) # try: # os.makedirs(save_path) # except OSError: # pass # # A really ugly way to save a model to CPU # save_model = actor_critic # if args.cuda: # save_model = copy.deepcopy(actor_critic).cpu() # torch.save(save_model, os.path.join(save_path, args.env_name + ".pt")) #Print updates if j % args.log_interval == 0: end = time.time() total_num_steps = (j + 1) * args.num_processes * args.num_steps # print("Updates {}, n_timesteps {}, FPS {}, mean/median R {:.1f}/{:.1f}, min/max R {:.1f}/{:.1f}, T:{:.4f}".#, entropy {:.5f}, value loss {:.5f}, policy loss {:.5f}". # format(j, total_num_steps, # int(total_num_steps / (end - start)), # final_rewards.mean(), # final_rewards.median(), # final_rewards.min(), # final_rewards.max(), # end - start))#, -dist_entropy.data[0], # # value_loss.data[0], action_loss.data[0])) # print("Upts {}, n_timesteps {}, min/med/mean/max {:.1f}/{:.1f}/{:.1f}/{:.1f}, FPS {}, T:{:.1f}". # format(j, total_num_steps, # final_rewards.min(), # final_rewards.median(), # final_rewards.mean(), # final_rewards.max(), # int(total_num_steps / (end - start)), # end - start)) if j % (args.log_interval*30) == 0: print("Upts, n_timesteps, min/med/mean/max, FPS, Time") print("{}, {}, {:.1f}/{:.1f}/{:.1f}/{:.1f}, {}, {:.1f}". format(j, total_num_steps, final_rewards.min(), final_rewards.median(), final_rewards.mean(), final_rewards.max(), int(total_num_steps / (end - start)), end - start))
def main(): os.environ['OMP_NUM_THREADS'] = '1' if args.vis: from visdom import Visdom viz = Visdom() win = None envs = [ make_env(args.env_name, args.seed, i, args.log_dir, args.start_container) for i in range(args.num_processes) ] test_envs = [ make_env(args.env_name, args.seed, i, args.log_dir, args.start_container) for i in range(args.num_processes) ] if args.num_processes > 1: envs = SubprocVecEnv(envs) test_envs = SubprocVecEnv(test_envs) else: envs = DummyVecEnv(envs) test_envs = DummyVecEnv(test_envs) obs_shape = envs.observation_space.shape obs_shape = (obs_shape[0] * args.num_stack, *obs_shape[1:]) if args.saved_encoder_model: obs_shape = (args.num_stack, args.latent_space_size) obs_numel = reduce(operator.mul, obs_shape, 1) if len(obs_shape) == 3 and obs_numel > 1024: actor_critic = CNNPolicy(obs_shape[0], envs.action_space, args.recurrent_policy) else: assert not args.recurrent_policy, \ "Recurrent policy is not implemented for the MLP controller" actor_critic = MLPPolicy(obs_numel, envs.action_space) modelSize = 0 for p in actor_critic.parameters(): pSize = reduce(operator.mul, p.size(), 1) modelSize += pSize print(str(actor_critic)) print('Total model size: %d' % modelSize) if envs.action_space.__class__.__name__ == "Discrete": action_shape = 1 else: action_shape = envs.action_space.shape[0] if args.resume_experiment: print("\n############## Loading saved model ##############\n") actor_critic, ob_rms = torch.load( os.path.join(save_path, args.env_name + args.save_tag + ".pt")) tr.load(os.path.join(log_path, args.env_name + args.save_tag + ".p")) if args.cuda: actor_critic.cuda() if args.algo == 'a2c': optimizer = optim.RMSprop(actor_critic.parameters(), args.lr, eps=args.eps, alpha=args.alpha) elif args.algo == 'ppo': optimizer = optim.Adam(actor_critic.parameters(), args.lr, eps=args.eps) elif args.algo == 'acktr': optimizer = KFACOptimizer(actor_critic) print(obs_shape) rollouts = RolloutStorage(args.num_steps, args.num_processes, obs_shape, envs.action_space, actor_critic.state_size) rollouts_test = RolloutStorage(args.num_steps_test, args.num_processes, obs_shape, envs.action_space, actor_critic.state_size) current_obs = torch.zeros(args.num_processes, *obs_shape) current_obs_test = torch.zeros(args.num_processes, *obs_shape) def update_current_obs(obs, test=False): shape_dim0 = envs.observation_space.shape[0] if args.saved_encoder_model: shape_dim0 = 1 obs, _ = vae.encode(Variable(torch.cuda.FloatTensor(obs))) obs = obs.data.cpu().numpy() obs = torch.from_numpy(obs).float() if not test: if args.num_stack > 1: current_obs[:, :-shape_dim0] = current_obs[:, shape_dim0:] current_obs[:, -shape_dim0:] = obs else: if args.num_stack > 1: current_obs_test[:, : -shape_dim0] = current_obs_test[:, shape_dim0:] current_obs_test[:, -shape_dim0:] = obs obs = envs.reset() update_current_obs(obs) rollouts.observations[0].copy_(current_obs) # These variables are used to compute average rewards for all processes. episode_rewards = torch.zeros([args.num_processes, 1]) final_rewards = torch.zeros([args.num_processes, 1]) reward_avg = 0 if args.cuda: current_obs = current_obs.cuda() current_obs_test = current_obs_test.cuda() rollouts.cuda() rollouts_test.cuda() start = time.time() for j in range(num_updates): for step in range(args.num_steps): # Sample actions value, action, action_log_prob, states = actor_critic.act( Variable(rollouts.observations[step], volatile=True), Variable(rollouts.states[step], volatile=True), Variable(rollouts.masks[step], volatile=True)) cpu_actions = action.data.squeeze(1).cpu().numpy() # Observation, reward and next obs obs, reward, done, info = envs.step(cpu_actions) # Maxime: clip the reward within [0,1] for more reliable training # This code deals poorly with large reward values reward = np.clip(reward, a_min=0, a_max=None) / 400 reward = torch.from_numpy(np.expand_dims(np.stack(reward), 1)).float() episode_rewards += reward # If done then clean the history of observations. masks = torch.FloatTensor([[0.0] if done_ else [1.0] for done_ in done]) final_rewards *= masks final_rewards += (1 - masks) * episode_rewards episode_rewards *= masks tr.episodes_done += args.num_processes - masks.sum() if args.cuda: masks = masks.cuda() if current_obs.dim() == 4: current_obs *= masks.unsqueeze(2).unsqueeze(2) else: current_obs *= masks update_current_obs(obs) rollouts.insert(step, current_obs, states.data, action.data, action_log_prob.data, value.data, reward, masks) next_value = actor_critic( Variable(rollouts.observations[-1], volatile=True), Variable(rollouts.states[-1], volatile=True), Variable(rollouts.masks[-1], volatile=True))[0].data rollouts.compute_returns(next_value, args.use_gae, args.gamma, args.tau) tr.iterations_done += 1 if args.algo in ['a2c', 'acktr']: values, action_log_probs, dist_entropy, states = actor_critic.evaluate_actions( Variable(rollouts.observations[:-1].view(-1, *obs_shape)), Variable(rollouts.states[0].view(-1, actor_critic.state_size)), Variable(rollouts.masks[:-1].view(-1, 1)), Variable(rollouts.actions.view(-1, action_shape))) values = values.view(args.num_steps, args.num_processes, 1) action_log_probs = action_log_probs.view(args.num_steps, args.num_processes, 1) advantages = Variable(rollouts.returns[:-1]) - values value_loss = advantages.pow(2).mean() action_loss = -(Variable(advantages.data) * action_log_probs).mean() if args.algo == 'acktr' and optimizer.steps % optimizer.Ts == 0: # Sampled fisher, see Martens 2014 actor_critic.zero_grad() pg_fisher_loss = -action_log_probs.mean() value_noise = Variable(torch.randn(values.size())) if args.cuda: value_noise = value_noise.cuda() sample_values = values + value_noise vf_fisher_loss = -(values - Variable(sample_values.data)).pow(2).mean() fisher_loss = pg_fisher_loss + vf_fisher_loss optimizer.acc_stats = True fisher_loss.backward(retain_graph=True) optimizer.acc_stats = False optimizer.zero_grad() (value_loss * args.value_loss_coef + action_loss - dist_entropy * args.entropy_coef).backward() if args.algo == 'a2c': nn.utils.clip_grad_norm(actor_critic.parameters(), args.max_grad_norm) optimizer.step() elif args.algo == 'ppo': advantages = rollouts.returns[:-1] - rollouts.value_preds[:-1] advantages = (advantages - advantages.mean()) / (advantages.std() + 1e-5) for e in range(args.ppo_epoch): if args.recurrent_policy: data_generator = rollouts.recurrent_generator( advantages, args.num_mini_batch) else: data_generator = rollouts.feed_forward_generator( advantages, args.num_mini_batch) for sample in data_generator: observations_batch, states_batch, actions_batch, \ return_batch, masks_batch, old_action_log_probs_batch, \ adv_targ = sample # Reshape to do in a single forward pass for all steps values, action_log_probs, dist_entropy, states = actor_critic.evaluate_actions( Variable(observations_batch), Variable(states_batch), Variable(masks_batch), Variable(actions_batch)) adv_targ = Variable(adv_targ) ratio = torch.exp(action_log_probs - Variable(old_action_log_probs_batch)) surr1 = ratio * adv_targ surr2 = torch.clamp(ratio, 1.0 - args.clip_param, 1.0 + args.clip_param) * adv_targ action_loss = -torch.min( surr1, surr2).mean() # PPO's pessimistic surrogate (L^CLIP) value_loss = (Variable(return_batch) - values).pow(2).mean() optimizer.zero_grad() (value_loss + action_loss - dist_entropy * args.entropy_coef).backward() nn.utils.clip_grad_norm(actor_critic.parameters(), args.max_grad_norm) optimizer.step() rollouts.after_update() if j % args.save_interval == 0 and args.save_dir != "": # A really ugly way to save a model to CPU save_model = actor_critic if args.cuda: save_model = copy.deepcopy(actor_critic).cpu() save_model = [ save_model, hasattr(envs, 'ob_rms') and envs.ob_rms or None ] torch.save( save_model, os.path.join(save_path, args.env_name + args.save_tag + ".pt")) total_test_reward_list = [] step_test_list = [] for _ in range(args.num_tests): test_obs = test_envs.reset() update_current_obs(test_obs, test=True) rollouts_test.observations[0].copy_(current_obs_test) step_test = 0 total_test_reward = 0 while step_test < args.num_steps_test: value_test, action_test, action_log_prob_test, states_test = actor_critic.act( Variable(rollouts_test.observations[step_test], volatile=True), Variable(rollouts_test.states[step_test], volatile=True), Variable(rollouts_test.masks[step_test], volatile=True)) cpu_actions_test = action_test.data.squeeze( 1).cpu().numpy() # Observation, reward and next obs obs_test, reward_test, done_test, info_test = test_envs.step( cpu_actions_test) # masks here doesn't really matter, but still masks_test = torch.FloatTensor( [[0.0] if done_test_ else [1.0] for done_test_ in done_test]) # Maxime: clip the reward within [0,1] for more reliable training # This code deals poorly with large reward values reward_test = np.clip(reward_test, a_min=0, a_max=None) / 400 total_test_reward += reward_test[0] reward_test = torch.from_numpy( np.expand_dims(np.stack(reward_test), 1)).float() update_current_obs(obs_test) rollouts_test.insert(step_test, current_obs_test, states_test.data, action_test.data, action_log_prob_test.data,\ value_test.data, reward_test, masks_test) step_test += 1 if done_test: break #rollouts_test.reset() # Need to reinitialise with .cuda(); don't forget total_test_reward_list.append(total_test_reward) step_test_list.append(step_test) append_to(tr.test_reward, tr, sum(total_test_reward_list) / args.num_tests) append_to(tr.test_episode_len, tr, sum(step_test_list) / args.num_tests) logger.log_scalar_rl( "test_reward", tr.test_reward[0], args.sliding_wsize, [tr.episodes_done, tr.global_steps_done, tr.iterations_done]) logger.log_scalar_rl( "test_episode_len", tr.test_episode_len[0], args.sliding_wsize, [tr.episodes_done, tr.global_steps_done, tr.iterations_done]) # Saving all the MyContainer variables tr.save( os.path.join(log_path, args.env_name + args.save_tag + ".p")) if j % args.log_interval == 0: reward_avg = 0.99 * reward_avg + 0.01 * final_rewards.mean() end = time.time() tr.global_steps_done = (j + 1) * args.num_processes * args.num_steps print( "Updates {}, num timesteps {}, FPS {}, running avg reward {:.3f}, entropy {:.5f}, value loss {:.5f}, policy loss {:.5f}" .format(j, tr.global_steps_done, int(tr.global_steps_done / (end - start)), reward_avg, dist_entropy.data[0], value_loss.data[0], action_loss.data[0])) append_to(tr.pg_loss, tr, action_loss.data[0]) append_to(tr.val_loss, tr, value_loss.data[0]) append_to(tr.entropy_loss, tr, dist_entropy.data[0]) append_to(tr.train_reward_avg, tr, reward_avg) logger.log_scalar_rl( "train_pg_loss", tr.pg_loss[0], args.sliding_wsize, [tr.episodes_done, tr.global_steps_done, tr.iterations_done]) logger.log_scalar_rl( "train_val_loss", tr.val_loss[0], args.sliding_wsize, [tr.episodes_done, tr.global_steps_done, tr.iterations_done]) logger.log_scalar_rl( "train_entropy_loss", tr.entropy_loss[0], args.sliding_wsize, [tr.episodes_done, tr.global_steps_done, tr.iterations_done]) logger.log_scalar_rl( "train_reward_avg", tr.train_reward_avg[0], args.sliding_wsize, [tr.episodes_done, tr.global_steps_done, tr.iterations_done]) """ print("Updates {}, num timesteps {}, FPS {}, mean/median reward {:.1f}/{:.1f}, min/max reward {:.1f}/{:.1f}, entropy {:.5f}, value loss {:.5f}, policy loss {:.5f}". format( j, total_num_steps, int(total_num_steps / (end - start)), final_rewards.mean(), final_rewards.median(), final_rewards.min(), final_rewards.max(), dist_entropy.data[0], value_loss.data[0], action_loss.data[0]) ) """ if args.vis and j % args.vis_interval == 0: try: # Sometimes monitor doesn't properly flush the outputs win = visdom_plot(viz, win, args.log_dir, args.env_name, args.algo) except IOError: pass