# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import argparse import cv2 import gym import os import threading import torch import parl import numpy as np from tqdm import tqdm from parl.utils import tensorboard, logger from parl.algorithms import DQN, DDQN from agent import AtariAgent from atari_wrapper import FireResetEnv, FrameStack, LimitLength, MapState from model import AtariModel from replay_memory import ReplayMemory, Experience from utils import get_player MEMORY_SIZE = int(1e6) MEMORY_WARMUP_SIZE = MEMORY_SIZE // 20 IMAGE_SIZE = (84, 84) CONTEXT_LEN = 4 FRAME_SKIP = 4 UPDATE_FREQ = 4 GAMMA = 0.99 def run_train_episode(env, agent, rpm): total_reward = 0 all_cost = [] state = env.reset() steps = 0 while True: steps += 1 context = rpm.recent_state() context.append(state) context = np.stack(context, axis=0) action = agent.sample(context) next_state, reward, isOver, _ = env.step(action) rpm.append(Experience(state, action, reward, isOver)) if rpm.size() > MEMORY_WARMUP_SIZE: if steps % UPDATE_FREQ == 0: batch_all_state, batch_action, batch_reward, batch_isOver = rpm.sample_batch( args.batch_size) batch_state = batch_all_state[:, :CONTEXT_LEN, :, :] batch_next_state = batch_all_state[:, 1:, :, :] cost = agent.learn(batch_state, batch_action, batch_reward, batch_next_state, batch_isOver) all_cost.append(cost) total_reward += reward state = next_state if isOver: mean_loss = np.mean(all_cost) if all_cost else None return total_reward, steps, mean_loss def run_evaluate_episode(env, agent): state = env.reset() total_reward = 0 while True: pred_Q = agent.predict(state) action = pred_Q.max(1)[1].item() state, reward, isOver, _ = env.step(action) total_reward += reward if isOver: return total_reward def get_fixed_states(rpm, batch_size): states = [] for _ in range(3): batch_all_state = rpm.sample_batch(batch_size)[0] batch_state = batch_all_state[:, :CONTEXT_LEN, :, :] states.append(batch_state) fixed_states = np.concatenate(states, axis=0) return fixed_states def evaluate_fixed_Q(agent, states): with torch.no_grad(): max_pred_Q = agent.alg.model(states).max(1)[0].mean() return max_pred_Q.item() def get_grad_norm(model): total_norm = 0 for p in model.parameters(): if p.grad is not None: param_norm = p.grad.data.norm(2) total_norm += param_norm.item()**2 total_norm = total_norm**(1. / 2) return total_norm def main(): env = get_player( args.rom, image_size=IMAGE_SIZE, train=True, frame_skip=FRAME_SKIP) test_env = get_player( args.rom, image_size=IMAGE_SIZE, frame_skip=FRAME_SKIP, context_len=CONTEXT_LEN) rpm = ReplayMemory(MEMORY_SIZE, IMAGE_SIZE, CONTEXT_LEN) act_dim = env.action_space.n device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') model = AtariModel(CONTEXT_LEN, act_dim, args.algo) if args.algo in ['DQN', 'Dueling']: algorithm = DQN(model, gamma=GAMMA, lr=args.lr) elif args.algo is 'Double': algorithm = DDQN(model, gamma=GAMMA, lr=args.lr) agent = AtariAgent(algorithm, act_dim=act_dim) with tqdm( total=MEMORY_WARMUP_SIZE, desc='[Replay Memory Warm Up]') as pbar: while rpm.size() < MEMORY_WARMUP_SIZE: total_reward, steps, _ = run_train_episode(env, agent, rpm) pbar.update(steps) # Get fixed states to check value function. fixed_states = get_fixed_states(rpm, args.batch_size) fixed_states = torch.tensor(fixed_states, dtype=torch.float, device=device) # train test_flag = 0 total_steps = 0 with tqdm(total=args.train_total_steps, desc='[Training Model]') as pbar: while total_steps < args.train_total_steps: total_reward, steps, loss = run_train_episode(env, agent, rpm) total_steps += steps pbar.update(steps) if total_steps // args.test_every_steps >= test_flag: while total_steps // args.test_every_steps >= test_flag: test_flag += 1 eval_rewards = [] for _ in range(3): eval_rewards.append(run_evaluate_episode(test_env, agent)) tensorboard.add_scalar('dqn/eval', np.mean(eval_rewards), total_steps) tensorboard.add_scalar('dqn/score', total_reward, total_steps) tensorboard.add_scalar('dqn/loss', loss, total_steps) tensorboard.add_scalar('dqn/exploration', agent.exploration, total_steps) tensorboard.add_scalar('dqn/Q value', evaluate_fixed_Q(agent, fixed_states), total_steps) tensorboard.add_scalar('dqn/grad_norm', get_grad_norm(agent.alg.model), total_steps) if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('--rom', default='rom_files/breakout.bin') parser.add_argument( '--batch_size', type=int, default=32, help='batch size for training') parser.add_argument('--lr', default=3e-4, help='learning_rate') parser.add_argument('--algo', default='DQN', help='DQN/Double/Dueling DQN') parser.add_argument( '--train_total_steps', type=int, default=int(1e7), help='maximum environmental steps of games') parser.add_argument( '--test_every_steps', type=int, default=int(1e5), help='the step interval between two consecutive evaluations') args = parser.parse_args() rom_name = args.rom.split('/')[-1].split('.')[0] logger.set_dir(os.path.join('./train_log', rom_name)) main()