import argparse import wandb import time import numpy as np import random import torch import torch.nn as nn import torch.optim as optim from AimbotEnv import Aimbot from tqdm import tqdm from torch.distributions.normal import Normal from torch.distributions.categorical import Categorical from distutils.util import strtobool from torch.utils.tensorboard import SummaryWriter bestReward = 0 DEFAULT_SEED = 9331 ENV_PATH = "../Build/Build-ParallelEnv-BigArea-6Enemy-EndBonus/Aimbot-ParallelEnv" WAND_ENTITY = "koha9" WORKER_ID = 1 BASE_PORT = 1000 # max round steps per agent is 2500, 25 seconds TOTAL_STEPS = 4000000 BATCH_SIZE = 512 MAX_TRAINNING_DATASETS = 8000 DECISION_PERIOD = 2 LEARNING_RATE = 7e-4 GAMMA = 0.99 GAE_LAMBDA = 0.95 EPOCHS = 4 CLIP_COEF = 0.1 POLICY_COEF = 1.0 ENTROPY_COEF = 0.01 CRITIC_COEF = 0.5 ANNEAL_LEARNING_RATE = False CLIP_VLOSS = True NORM_ADV = True TRAIN = False WANDB_TACK = False LOAD_DIR = None LOAD_DIR = "../PPO-Model/bigArea-4.pt" def parse_args(): # fmt: off # pytorch and environment parameters parser = argparse.ArgumentParser() parser.add_argument("--seed", type=int, default=DEFAULT_SEED, help="seed of the experiment") parser.add_argument("--path", type=str, default=ENV_PATH, help="enviroment path") parser.add_argument("--workerID", type=int, default=WORKER_ID, help="unity worker ID") parser.add_argument("--baseport", type=int, default=BASE_PORT, help="port to connect to Unity environment") parser.add_argument("--lr", type=float, default=LEARNING_RATE, help="the learning rate of optimizer") parser.add_argument("--cuda", type=lambda x: bool(strtobool(x)), default=True, nargs="?", const=True, help="if toggled, cuda will be enabled by default") parser.add_argument("--total-timesteps", type=int, default=TOTAL_STEPS, help="total timesteps of the experiments") # model parameters parser.add_argument("--train",type=lambda x: bool(strtobool(x)), default=TRAIN, nargs="?", const=True, help="Train Model or not") parser.add_argument("--datasetSize", type=int, default=MAX_TRAINNING_DATASETS, help="training dataset size,start training while dataset collect enough data") parser.add_argument("--minibatchSize", type=int, default=BATCH_SIZE, help="nimi batch size") parser.add_argument("--epochs", type=int, default=EPOCHS, help="the K epochs to update the policy") parser.add_argument("--annealLR", type=lambda x: bool(strtobool(x)), default=ANNEAL_LEARNING_RATE, nargs="?", const=True, help="Toggle learning rate annealing for policy and value networks") parser.add_argument("--wandb-track", type=lambda x: bool(strtobool(x)), default=WANDB_TACK, nargs="?", const=True, help="track on the wandb") parser.add_argument("--wandb-entity", type=str, default=WAND_ENTITY, help="the entity (team) of wandb's project") parser.add_argument("--load-dir", type=str, default=LOAD_DIR, help="load model directory") parser.add_argument("--decision-period", type=int, default=DECISION_PERIOD, help="the number of steps to run in each environment per policy rollout") # GAE loss parser.add_argument("--gae", type=lambda x: bool(strtobool(x)), default=True, nargs="?", const=True, help="Use GAE for advantage computation") parser.add_argument("--norm-adv", type=lambda x: bool(strtobool(x)), default=NORM_ADV, nargs="?", const=True, help="Toggles advantages normalization") parser.add_argument("--gamma", type=float, default=GAMMA, help="the discount factor gamma") parser.add_argument("--gaeLambda", type=float, default=GAE_LAMBDA, help="the lambda for the general advantage estimation") parser.add_argument("--clip-coef", type=float, default=CLIP_COEF, help="the surrogate clipping coefficient") parser.add_argument("--policy-coef", type=float, default=POLICY_COEF, help="coefficient of the policy") parser.add_argument("--ent-coef", type=float, default=ENTROPY_COEF, help="coefficient of the entropy") parser.add_argument("--critic-coef", type=float, default=CRITIC_COEF, help="coefficient of the value function") parser.add_argument("--clip-vloss", type=lambda x: bool(strtobool(x)), default=CLIP_VLOSS, nargs="?", const=True, help="Toggles whether or not to use a clipped loss for the value function, as per the paper.") parser.add_argument("--max-grad-norm", type=float, default=0.5, help="the maximum norm for the gradient clipping") parser.add_argument("--target-kl", type=float, default=None, help="the target KL divergence threshold") # fmt: on args = parser.parse_args() return args def layer_init(layer, std=np.sqrt(2), bias_const=0.0): torch.nn.init.orthogonal_(layer.weight, std) torch.nn.init.constant_(layer.bias, bias_const) return layer class PPOAgent(nn.Module): def __init__(self, env: Aimbot): super(PPOAgent, self).__init__() self.discrete_size = env.unity_discrete_size self.discrete_shape = list(env.unity_discrete_branches) self.continuous_size = env.unity_continuous_size self.network = nn.Sequential( layer_init(nn.Linear(np.array(env.unity_observation_shape).prod(), 384)), nn.ReLU(), layer_init(nn.Linear(384, 256)), nn.ReLU(), ) self.actor_dis = layer_init(nn.Linear(256, self.discrete_size), std=0.01) self.actor_mean = layer_init(nn.Linear(256, self.continuous_size), std=0.01) self.actor_logstd = nn.Parameter(torch.zeros(1, self.continuous_size)) self.critic = layer_init(nn.Linear(256, 1), std=1) def get_value(self, state: torch.Tensor): return self.critic(self.network(state)) def get_actions_value(self, state: torch.Tensor, actions=None): hidden = self.network(state) # discrete dis_logits = self.actor_dis(hidden) split_logits = torch.split(dis_logits, self.discrete_shape, dim=1) multi_categoricals = [Categorical(logits=thisLogits) for thisLogits in split_logits] # continuous actions_mean = self.actor_mean(hidden) action_logstd = self.actor_logstd.expand_as(actions_mean) action_std = torch.exp(action_logstd) con_probs = Normal(actions_mean, action_std) if actions is None: if args.train: # select actions base on probability distribution model disAct = torch.stack([ctgr.sample() for ctgr in multi_categoricals]) conAct = con_probs.sample() actions = torch.cat([disAct.T, conAct], dim=1) else: # select actions base on best probability distribution disAct = torch.stack([torch.argmax(logit, dim=1) for logit in split_logits]) conAct = actions_mean actions = torch.cat([disAct.T, conAct], dim=1) else: disAct = actions[:, 0 : env.unity_discrete_type].T conAct = actions[:, env.unity_discrete_type :] dis_log_prob = torch.stack( [ctgr.log_prob(act) for act, ctgr in zip(disAct, multi_categoricals)] ) dis_entropy = torch.stack([ctgr.entropy() for ctgr in multi_categoricals]) return ( actions, dis_log_prob.sum(0), dis_entropy.sum(0), con_probs.log_prob(conAct).sum(1), con_probs.entropy().sum(1), self.critic(hidden), ) def GAE(agent, args, rewards, dones, values, next_obs, next_done): # GAE with torch.no_grad(): next_value = agent.get_value(next_obs).reshape(1, -1) data_size = rewards.size()[0] if args.gae: advantages = torch.zeros_like(rewards).to(device) lastgaelam = 0 for t in reversed(range(data_size)): if t == data_size - 1: nextnonterminal = 1.0 - next_done nextvalues = next_value else: nextnonterminal = 1.0 - dones[t + 1] nextvalues = values[t + 1] delta = rewards[t] + args.gamma * nextvalues * nextnonterminal - values[t] advantages[t] = lastgaelam = ( delta + args.gamma * args.gaeLambda * nextnonterminal * lastgaelam ) returns = advantages + values else: returns = torch.zeros_like(rewards).to(device) for t in reversed(range(data_size)): if t == data_size - 1: nextnonterminal = 1.0 - next_done next_return = next_value else: nextnonterminal = 1.0 - dones[t + 1] next_return = returns[t + 1] returns[t] = rewards[t] + args.gamma * nextnonterminal * next_return advantages = returns - values return advantages, returns if __name__ == "__main__": args = parse_args() random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) device = torch.device("cuda" if torch.cuda.is_available() and args.cuda else "cpu") # Initialize environment anget optimizer env = Aimbot(envPath=args.path, workerID=args.workerID, basePort=args.baseport) if args.load_dir is None: agent = PPOAgent(env).to(device) else: agent = torch.load(args.load_dir) print("Load Agent", args.load_dir) print(agent.eval()) optimizer = optim.Adam(agent.parameters(), lr=args.lr, eps=1e-5) # Tensorboard and WandB Recorder game_name = "Aimbot-BigArea-6Enemy-EndBonus" run_name = f"{game_name}_{args.seed}_{int(time.time())}" if args.wandb_track: wandb.init( project=game_name, entity=args.wandb_entity, sync_tensorboard=True, config=vars(args), name=run_name, monitor_gym=True, save_code=True, ) writer = SummaryWriter(f"runs/{run_name}") writer.add_text( "hyperparameters", "|param|value|\n|-|-|\n%s" % ("\n".join([f"|{key}|{value}|" for key, value in vars(args).items()])), ) # Trajectory Buffer ob_bf = [[] for i in range(env.unity_agent_num)] act_bf = [[] for i in range(env.unity_agent_num)] dis_logprobs_bf = [[] for i in range(env.unity_agent_num)] con_logprobs_bf = [[] for i in range(env.unity_agent_num)] rewards_bf = [[] for i in range(env.unity_agent_num)] dones_bf = [[] for i in range(env.unity_agent_num)] values_bf = [[] for i in range(env.unity_agent_num)] # TRY NOT TO MODIFY: start the game total_update_step = args.total_timesteps // args.datasetSize global_step = 0 start_time = time.time() state, _, done = env.reset() # state = torch.Tensor(next_obs).to(device) # next_done = torch.zeros(env.unity_agent_num).to(device) for total_steps in range(total_update_step): # discunt learning rate, while step == total_update_step lr will be 0 print("new episode") if args.annealLR: frac = 1.0 - (total_steps - 1.0) / total_update_step lrnow = frac * args.lr optimizer.param_groups[0]["lr"] = lrnow # initialize empty training datasets obs = torch.tensor([]).to(device) # (n,env.unity_observation_size) actions = torch.tensor([]).to(device) # (n,env.unity_action_size) dis_logprobs = torch.tensor([]).to(device) # (n,1) con_logprobs = torch.tensor([]).to(device) # (n,1) rewards = torch.tensor([]).to(device) # (n,1) values = torch.tensor([]).to(device) # (n,1) advantages = torch.tensor([]).to(device) # (n,1) returns = torch.tensor([]).to(device) # (n,1) # MAIN LOOP: run agent in environment i = 0 training = False while True: if i % args.decision_period == 0: step = round(i / args.decision_period) # Choose action by agent global_step += 1 * env.unity_agent_num with torch.no_grad(): # predict actions action, dis_logprob, _, con_logprob, _, value = agent.get_actions_value( torch.Tensor(state).to(device) ) value = value.flatten() # variable from GPU to CPU action_cpu = action.cpu().numpy() dis_logprob_cpu = dis_logprob.cpu().numpy() con_logprob_cpu = con_logprob.cpu().numpy() value_cpu = value.cpu().numpy() # Environment step next_state, reward, next_done = env.step(action_cpu) # save memories for i in range(env.unity_agent_num): # save memories to buffers ob_bf[i].append(state[i]) act_bf[i].append(action_cpu[i]) dis_logprobs_bf[i].append(dis_logprob_cpu[i]) con_logprobs_bf[i].append(con_logprob_cpu[i]) rewards_bf[i].append(reward[i]) dones_bf[i].append(done[i]) values_bf[i].append(value_cpu[i]) if next_done[i] == True: # finished a round, send finished memories to training datasets # compute advantage and discounted reward adv, rt = GAE( agent, args, torch.tensor(rewards_bf[i]).to(device), torch.Tensor(dones_bf[i]).to(device), torch.tensor(values_bf[i]).to(device), torch.tensor(next_state[i]).to(device), torch.Tensor([next_done[i]]).to(device), ) # send memories to training datasets obs = torch.cat((obs, torch.tensor(ob_bf[i]).to(device)), 0) actions = torch.cat((actions, torch.tensor(act_bf[i]).to(device)), 0) dis_logprobs = torch.cat( (dis_logprobs, torch.tensor(dis_logprobs_bf[i]).to(device)), 0 ) con_logprobs = torch.cat( (con_logprobs, torch.tensor(con_logprobs_bf[i]).to(device)), 0 ) rewards = torch.cat((rewards, torch.tensor(rewards_bf[i]).to(device)), 0) values = torch.cat((values, torch.tensor(values_bf[i]).to(device)), 0) advantages = torch.cat((advantages, adv), 0) returns = torch.cat((returns, rt), 0) # clear buffers ob_bf[i] = [] act_bf[i] = [] dis_logprobs_bf[i] = [] con_logprobs_bf[i] = [] rewards_bf[i] = [] dones_bf[i] = [] values_bf[i] = [] print(f"train dataset:{obs.size()[0]}/{args.datasetSize}") if obs.size()[0] >= args.datasetSize: # start train NN break state, done = next_state, next_done else: # skip this step use last predict action next_obs, reward, done = env.step(action_cpu) # save memories for i in range(env.unity_agent_num): if next_done[i] == True: # save last memories to buffers ob_bf[i].append(state[i]) act_bf[i].append(action_cpu[i]) dis_logprobs_bf[i].append(dis_logprob_cpu[i]) con_logprobs_bf[i].append(con_logprob_cpu[i]) rewards_bf[i].append(reward[i]) dones_bf[i].append(done[i]) values_bf[i].append(value_cpu[i]) # finished a round, send finished memories to training datasets # compute advantage and discounted reward adv, rt = GAE( agent, args, torch.tensor(rewards_bf[i]).to(device), torch.Tensor(dones_bf[i]).to(device), torch.tensor(values_bf[i]).to(device), torch.tensor(next_state[i]).to(device), torch.Tensor([next_done[i]]).to(device), ) # send memories to training datasets obs = torch.cat((obs, torch.tensor(ob_bf[i]).to(device)), 0) actions = torch.cat((actions, torch.tensor(act_bf[i]).to(device)), 0) dis_logprobs = torch.cat( (dis_logprobs, torch.tensor(dis_logprobs_bf[i]).to(device)), 0 ) con_logprobs = torch.cat( (con_logprobs, torch.tensor(con_logprobs_bf[i]).to(device)), 0 ) rewards = torch.cat((rewards, torch.tensor(rewards_bf[i]).to(device)), 0) values = torch.cat((values, torch.tensor(values_bf[i]).to(device)), 0) advantages = torch.cat((advantages, adv), 0) returns = torch.cat((returns, rt), 0) # clear buffers ob_bf[i] = [] act_bf[i] = [] dis_logprobs_bf[i] = [] con_logprobs_bf[i] = [] rewards_bf[i] = [] dones_bf[i] = [] values_bf[i] = [] print(f"train dataset:{obs.size()[0]}/{args.datasetSize}") state, done = next_state, next_done i += 1 if args.train: # flatten the batch b_obs = obs.reshape((-1,) + env.unity_observation_shape) b_dis_logprobs = dis_logprobs.reshape(-1) b_con_logprobs = con_logprobs.reshape(-1) b_actions = actions.reshape((-1,) + (env.unity_action_size,)) b_advantages = advantages.reshape(-1) b_returns = returns.reshape(-1) b_values = values.reshape(-1) b_size = b_obs.size()[0] # Optimizing the policy and value network b_inds = np.arange(b_size) # clipfracs = [] for epoch in range(args.epochs): # shuffle all datasets np.random.shuffle(b_inds) for start in range(0, b_size, args.minibatchSize): end = start + args.minibatchSize mb_inds = b_inds[start:end] mb_advantages = b_advantages[mb_inds] # normalize advantages if args.norm_adv: mb_advantages = (mb_advantages - mb_advantages.mean()) / ( mb_advantages.std() + 1e-8 ) ( _, new_dis_logprob, dis_entropy, new_con_logprob, con_entropy, newvalue, ) = agent.get_actions_value(b_obs[mb_inds], b_actions[mb_inds]) # discrete ratio dis_logratio = new_dis_logprob - b_dis_logprobs[mb_inds] dis_ratio = dis_logratio.exp() # continuous ratio con_logratio = new_con_logprob - b_con_logprobs[mb_inds] con_ratio = con_logratio.exp() """ # early stop with torch.no_grad(): # calculate approx_kl http://joschu.net/blog/kl-approx.html old_approx_kl = (-logratio).mean() approx_kl = ((ratio - 1) - logratio).mean() clipfracs += [((ratio - 1.0).abs() > args.clip_coef).float().mean().item()] """ # discrete Policy loss dis_pg_loss_orig = -mb_advantages * dis_ratio dis_pg_loss_clip = -mb_advantages * torch.clamp( dis_ratio, 1 - args.clip_coef, 1 + args.clip_coef ) dis_pg_loss = torch.max(dis_pg_loss_orig, dis_pg_loss_clip).mean() # continuous Policy loss con_pg_loss_orig = -mb_advantages * con_ratio con_pg_loss_clip = -mb_advantages * torch.clamp( con_ratio, 1 - args.clip_coef, 1 + args.clip_coef ) con_pg_loss = torch.max(con_pg_loss_orig, con_pg_loss_clip).mean() # Value loss newvalue = newvalue.view(-1) if args.clip_vloss: v_loss_unclipped = (newvalue - b_returns[mb_inds]) ** 2 v_clipped = b_values[mb_inds] + torch.clamp( newvalue - b_values[mb_inds], -args.clip_coef, args.clip_coef, ) v_loss_clipped = (v_clipped - b_returns[mb_inds]) ** 2 v_loss_max = torch.max(v_loss_unclipped, v_loss_clipped) v_loss = 0.5 * v_loss_max.mean() else: v_loss = 0.5 * ((newvalue - b_returns[mb_inds]) ** 2).mean() # total loss entropy_loss = dis_entropy.mean() + con_entropy.mean() loss = ( dis_pg_loss * args.policy_coef + con_pg_loss * args.policy_coef - entropy_loss * args.ent_coef + v_loss * args.critic_coef ) optimizer.zero_grad() loss.backward() # Clips gradient norm of an iterable of parameters. nn.utils.clip_grad_norm_(agent.parameters(), args.max_grad_norm) optimizer.step() """ if args.target_kl is not None: if approx_kl > args.target_kl: break """ # record rewards for plotting purposes rewardsMean = np.mean(rewards.to("cpu").detach().numpy().copy()) writer.add_scalar("charts/learning_rate", optimizer.param_groups[0]["lr"], global_step) writer.add_scalar("losses/value_loss", v_loss.item(), global_step) writer.add_scalar("losses/dis_policy_loss", dis_pg_loss.item(), global_step) writer.add_scalar("losses/con_policy_loss", con_pg_loss.item(), global_step) writer.add_scalar("losses/total_loss", loss.item(), global_step) writer.add_scalar("losses/entropy_loss", entropy_loss.item(), global_step) # writer.add_scalar("losses/old_approx_kl", old_approx_kl.item(), global_step) # writer.add_scalar("losses/approx_kl", approx_kl.item(), global_step) # writer.add_scalar("losses/clipfrac", np.mean(clipfracs), global_step) # print("SPS:", int(global_step / (time.time() - start_time))) print("episode over mean reward:", rewardsMean) writer.add_scalar( "charts/SPS", int(global_step / (time.time() - start_time)), global_step ) writer.add_scalar("charts/Reward", rewardsMean, global_step) if rewardsMean > bestReward: bestReward = rewardsMean saveDir = "../PPO-Model/bigArea-384-128-hybrid-" + str(rewardsMean) + ".pt" torch.save(agent, saveDir) env.close() writer.close()