SAC_apex.py

import argparse
import json
import os
import gym
import time
import torch
import copy
import numpy as np
import torch.nn as nn
import torch.multiprocessing as mp
from torch.optim import Adam
from copy import deepcopy
from torch.utils.tensorboard import SummaryWriter
from OppModeling.atari_wrappers import make_ftg_ram, make_ftg_ram_nonstation
from OppModeling.utils import Counter
from OppModeling.SAC import MLPActorCritic
from OppModeling.CPC import CPC
from OppModeling.ReplayBuffer import ReplayBufferOppo
from games import Soccer, SoccerPLUS
from OppModeling.train_apex import sac
from OppModeling.evaluation_apex import test_func

if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    # running setting
    parser.add_argument('--cuda', default=False, action='store_true')
    parser.add_argument('--seed', '-s', type=int, default=0)
    parser.add_argument('--n_process', type=int, default=4)  # better to set as the cpu core - opp_num - 1
    # basic env setting
    parser.add_argument('--env', type=str, default="FightingiceDataFrameskip-v0")   # not used in Soccer Environment
    parser.add_argument('--p2', type=str, default="Toothless")  # not used in Soccer Environment
    # non station agent settings
    parser.add_argument('--non_station', default=False, action='store_true')  # not used in Soccer Environment
    parser.add_argument('--stable', default=False, action='store_true')  # not used in Soccer Environment
    parser.add_argument('--opp_freq', type=int, default=1,help="opponent change frequency")
    parser.add_argument('--opp_list', nargs='+',help="the list of different opponent type")
    parser.add_argument('--opp_num', type=int, default=2, help="different opponent types")
    parser.add_argument('--opp1_per', type=float, default=0.5, help="batch percentage of this opponent")  # name does not mean the opp type
    parser.add_argument('--opp3_per', type=float, default=0.5, help="batch percentage of this opponent")  # name does not mean the opp type
    # sac setting
    parser.add_argument('--replay_size', type=int, default=50000)
    parser.add_argument('--batch_size', type=int, default=5000)
    parser.add_argument('--hid', type=int, default=256)
    parser.add_argument('--l', type=int, default=2, help="layers")
    parser.add_argument('--episode', type=int, default=100000)   # changed to training steps, though the name is still episode
    parser.add_argument('--update_after', type=int, default=1000)
    parser.add_argument('--update_every', type=int, default=1)
    parser.add_argument('--max_ep_len', type=int, default=1000)
    parser.add_argument('--min_alpha', type=float, default=0.05)
    parser.add_argument('--dynamic_alpha', default=False, action="store_true")
    parser.add_argument('--gamma', type=float, default=0.95)
    parser.add_argument('--lr', type=float, default=1e-3)
    parser.add_argument('--polyak', type=float, default=0.995)
    # CPC setting
    parser.add_argument('--cpc', default=False, action="store_true")
    parser.add_argument('--cpc_batch', type=int, default=256)
    parser.add_argument('--z_dim', type=int, default=32)
    parser.add_argument('--c_dim', type=int, default=16)
    parser.add_argument('--timestep', type=int, default=10)
    parser.add_argument('--cpc_update_freq', type=int, default=1, )
    parser.add_argument('--forget_percent', type=float, default=0.2, )

    # evaluation settings
    parser.add_argument('--test_episode', type=int, default=100)
    parser.add_argument('--test_every', type=int, default=100)
    # only for percentage experiment
    # Saving settings
    parser.add_argument('--save_freq', type=int, default=500)
    parser.add_argument('--exp_name', type=str, default='percent5-5')
    parser.add_argument('--save-dir', type=str, default="./exp_po45")
    parser.add_argument('--traj_dir', type=str, default="./experiments_per")
    parser.add_argument('--model_para', type=str, default="sac.torch")
    parser.add_argument('--cpc_para', type=str, default="cpc.torch")
    parser.add_argument('--numpy_para', type=str, default="model.numpy")
    parser.add_argument('--train_indicator', type=str, default="test.data")
    args = parser.parse_args()

    # Basic Settings
    mp.set_start_method("forkserver")
    os.environ['OMP_NUM_THREADS'] = '1'
    os.environ['MKL_NUM_THREADS'] = '1'
    torch.set_num_threads(torch.get_num_threads())
    experiment_dir = os.path.join(args.save_dir, args.exp_name)
    if not os.path.exists(experiment_dir):
        os.makedirs(experiment_dir)
    tensorboard_dir = os.path.join(experiment_dir, "runs")
    if not os.path.exists(tensorboard_dir):
        os.makedirs(tensorboard_dir)
    main_dir = os.path.join(tensorboard_dir, "train")
    if not os.path.exists(main_dir):
        os.makedirs(main_dir)
    writer = SummaryWriter(log_dir=main_dir)
    with open(os.path.join(experiment_dir, "arguments"), 'w') as f:
        json.dump(args.__dict__, f, indent=2)
    device = torch.device("cuda") if args.cuda else torch.device("cpu")
    # env and model setup
    ac_kwargs = dict(hidden_sizes=[args.hid] * args.l)

    # if args.exp_name == "test":
    #     env = gym.make("CartPole-v0")
    # elif args.non_station:
    #     env = make_ftg_ram_nonstation(args.env, p2_list=args.opp_list, total_episode=args.opp_freq,stable=args.stable)
    # else:
    #     env = make_ftg_ram(args.env, p2=args.p2)
    env = SoccerPLUS()
    obs_dim = env.n_features
    act_dim = env.n_actions
    # create model
    global_ac = MLPActorCritic(obs_dim, act_dim, **ac_kwargs)
    if args.cpc:
        global_cpc = CPC(timestep=args.timestep, obs_dim=obs_dim, hidden_sizes=[args.hid] * args.l, z_dim=args.z_dim,
                         c_dim=args.c_dim, device=device)
    else:
        global_cpc = None
    # create shared model for actor
    global_ac_targ = deepcopy(global_ac)
    shared_ac = deepcopy(global_ac).cpu()
    # create optimizer
    pi_optimizer = Adam(global_ac.pi.parameters(), lr=args.lr, eps=1e-4)
    q1_optimizer = Adam(global_ac.q1.parameters(), lr=args.lr, eps=1e-4)
    q2_optimizer = Adam(global_ac.q2.parameters(), lr=args.lr, eps=1e-4)
    alpha_optim = Adam([global_ac.log_alpha], lr=args.lr, eps=1e-4)
    if args.cpc:
        cpc_optimizer = Adam(global_cpc.parameters(), lr=args.lr, eps=1e-4)
    env.close()
    del env

    # training setup
    T = Counter()  # training steps
    E = Counter()  # training episode
    # used for training single agent, similar setting for the following code
    # replay_buffer = ReplayBufferOppo(obs_dim=obs_dim, max_size=args.replay_size, cpc=args.cpc,
    #                                  cpc_model=global_cpc, writer=writer,E=E)

    # used for training multiple agent, now use two. , similar setting for the following code, you can change between these two setting
    bufferopp1 = ReplayBufferOppo(obs_dim=obs_dim, max_size=args.replay_size, cpc=args.cpc,
                                     cpc_model=global_cpc, writer=writer, E=E)

    bufferopp3 = ReplayBufferOppo(obs_dim=obs_dim, max_size=args.replay_size, cpc=args.cpc,
                                     cpc_model=global_cpc, writer=writer, E=E)


    if os.path.exists(os.path.join(args.save_dir, args.exp_name, args.model_para)):
        global_ac.load_state_dict(torch.load(os.path.join(args.save_dir, args.exp_name, args.model_para)))
        print("load sac model")
        if args.cpc:
            global_cpc.load_state_dict(torch.load(os.path.join(args.save_dir, args.exp_name, args.cpc_para)))
            print("load cpc model")
    if os.path.exists(os.path.join(args.save_dir, args.exp_name, args.train_indicator)):
        (e, t) = torch.load(os.path.join(args.save_dir, args.exp_name, args.train_indicator))
        T.set(t)
        E.set(e)
        print("load training indicator")

    last_updated = 0
    last_deliver = 0
    last_saved = 0
    test_t = 0
    if args.cuda:
        global_ac.to(device)
        global_ac_targ.to(device)
        if args.cpc:
            global_cpc.to(device)

    for p in global_ac_targ.parameters():
        p.requires_grad = False

    buffer_q = mp.SimpleQueue()
    model_q = [mp.SimpleQueue() for _ in range(args.n_process + args.opp_num)]  # + n opp test process
    evaluation_queue = list()
    processes = []
    # Process n for evaluation
    for rank in range(args.n_process + args.opp_num):  # + n opp test process
        # Test during training
        if rank < args.opp_num:
            # test processes
            p = mp.Process(target=test_func, args=(rank, E, T, args, model_q[rank], torch.device("cpu"), tensorboard_dir))
        else:
            # actor processes
            model_q[rank].put(shared_ac.state_dict())
            p = mp.Process(target=sac,
                           args=(rank, E, T,args, model_q[rank], buffer_q, torch.device("cpu"), tensorboard_dir))
        p.start()
        # time.sleep(5)
        processes.append(p)

    target_entropy = -np.log((1.0 / act_dim)) * 0.5
    alpha = max(global_ac.log_alpha.exp().item(), args.min_alpha) if args.dynamic_alpha else args.min_alpha
    # alpha = args.min_alpha
    e = E.value()
    # changed to T
    while T.value() < args.episode:
        t = T.value()

        # If do experiments, need to block the receive if now new data, otherwise will impact the result
        # if not buffer_q.empty():
            # print("Going to read data from ACTOR...")
            # before_rece = time.time()
        received_data = buffer_q.get()
        # wait_time = time.time() - before_rece
        # print("waited {}s Reading data from ACTOR!!!".format(wait_time))
        (trajectory, meta) = copy.deepcopy(received_data)
        del received_data
        if args.cpc and len(trajectory) <= args.timestep:
            continue
        if meta[0][0] == "4":
            bufferopp1.store(trajectory, meta=meta)
        elif meta[0][0] == "5":
            bufferopp3.store(trajectory, meta=meta)
        # replay_buffer.store(trajectory, meta=meta)
        # writer.add_scalar("learner/buffer_size", replay_buffer.size, e)
        writer.add_scalar("learner/buffer1_size", bufferopp1.size, e)
        writer.add_scalar("learner/buffer2_size", bufferopp3.size, e)
        E.increment()
        e = E.value()

        # SAC Update handling
        if e >= args.update_after:
            # if the batch size is very large, can train sac once per round
            for _ in range(args.update_every):
                T.increment()
                t = T.value()
                batch1 = bufferopp1.sample_trans(int(args.batch_size*args.opp1_per), device=device)
                batch3 = bufferopp1.sample_trans(int(args.batch_size*args.opp3_per), device=device)
                batch = {k: torch.cat((batch1[k], batch3[k]), dim=0) for k, v in batch1.items()}
                # batch = replay_buffer.sample_trans(args.batch_size, device=device)
                # First run one gradient descent step for Q1 and Q2
                q1_optimizer.zero_grad()
                q2_optimizer.zero_grad()

                loss_q = global_ac.compute_loss_q(batch, global_ac_targ, args.gamma, alpha)
                loss_q.backward()
                nn.utils.clip_grad_norm_(global_ac.parameters(), max_norm=20, norm_type=2)
                q1_optimizer.step()
                q2_optimizer.step()

                # Next run one gradient descent step for pi.
                pi_optimizer.zero_grad()
                loss_pi, entropy = global_ac.compute_loss_pi(batch, alpha)
                loss_pi.backward()
                nn.utils.clip_grad_norm_(global_ac.parameters(), max_norm=20, norm_type=2)
                pi_optimizer.step()

                alpha_optim.zero_grad()
                alpha_loss = -(global_ac.log_alpha * (entropy + target_entropy).detach()).mean()
                alpha_loss.backward(retain_graph=False)
                nn.utils.clip_grad_norm_(global_ac.parameters(), max_norm=20, norm_type=2)
                alpha = max(global_ac.log_alpha.exp().item(), args.min_alpha) if args.dynamic_alpha else args.min_alpha
                alpha_optim.step()

                # Finally, update target networks by polyak averaging.
                with torch.no_grad():
                    for p, p_targ in zip(global_ac.parameters(), global_ac_targ.parameters()):
                        p_targ.data.copy_((1 - args.polyak) * p.data + args.polyak * p_targ.data)

                writer.add_scalar("learner/pi_loss", loss_pi.detach().item(), t)
                writer.add_scalar("learner/q_loss", loss_q.detach().item(), t)
                writer.add_scalar("learner/alpha_loss", alpha_loss.detach().item(), t)
                writer.add_scalar("learner/alpha", alpha, t)
                writer.add_scalar("learner/entropy", entropy.detach().mean().item(), t)

        # CPC update handing
        if args.cpc and e > args.cpc_batch * 2 and e % args.cpc_update_freq  == 0:
            for _ in range(args.cpc_update_freq):
                data, indexes, min_len = replay_buffer.sample_traj(args.cpc_batch)
                cpc_optimizer.zero_grad()
                c_hidden = global_cpc.init_hidden(len(data), args.c_dim)
                acc, loss, latents = global_cpc(data, c_hidden)

                # replay_buffer.update_latent(indexes, min_len, latents.detach())
                loss.backward()
                # add gradient clipping
                nn.utils.clip_grad_norm_(global_cpc.parameters(), max_norm=20, norm_type=2)
                cpc_optimizer.step()
                writer.add_scalar("learner/cpc_acc", acc, t)
                writer.add_scalar("learner/cpc_loss", loss.detach().item(), t)

        # CPC latent update
        # if args.cpc and e > args.cpc_batch and e % 500 == 0 and e != last_updated:
        #     replay_buffer.create_latents(e=e)
        #     last_updated = e

        # deliver the model
        if e % (args.n_process * 2) == 0 and e >= args.update_after and e != last_deliver:
            temp = copy.deepcopy(global_ac).cpu()
            shared_ac_state_dict = copy.deepcopy(temp.state_dict())
            for i in range(args.opp_num, args.n_process + args.opp_num):  # n is test model queue
                model_q[i].put(shared_ac_state_dict, )
            last_deliver = e

        # evaluation model
        if t > 0 and t % args.test_every == 0 and e >= args.update_after and  test_t != t:
            temp = copy.deepcopy(global_ac).cpu()
            test_model = copy.deepcopy(temp.state_dict())
            send_obj = (test_model, t)
            # in case the large model state dict will make the queue full to stuck the training process
            evaluation_queue.append(send_obj)
            if any([model_q[i].empty() for i in range(args.opp_num)]):
                temp = evaluation_queue.pop(0)
                for i in range(args.opp_num):
                    model_q[i].put(temp, )
            test_t = t

        # save the model
        if e % args.save_freq == 0 and e >= args.update_after and e != last_saved:
            torch.save(global_ac.state_dict(), os.path.join(experiment_dir, args.model_para))
            if args.cpc:
                torch.save(global_cpc.state_dict(), os.path.join(experiment_dir, args.cpc_para))
            # state_dict_trans(global_ac.state_dict(), os.path.join(experiment_dir, args.numpy_para))
            # torch.save((e, t, list(scores), list(wins)), os.path.join(args.save_dir, args.exp_name, "model_data_{}".format(e)))
            print("Saving model at episode:{}".format(e))
            last_saved = e
    print("after the training loop, waiting for the end of evaluation")
    # consume all the queue to make sure all child processes can be closed correctly
    while not buffer_q.empty():
        buffer_q.get()
    while len(evaluation_queue) > 0:
        if any([model_q[i].empty() for i in range(args.opp_num)]):
            temp = evaluation_queue.pop(0)
            for i in range(args.opp_num):
                model_q[i].put(temp, )
    for p in processes:
        p.join()