train.py
| 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 | import os import gym import time import threading import random import numpy as np import tensorflow as tf from skimage.color import rgb2gray from skimage.transform import resize from tensorflow.compat.v1.train import AdamOptimizer from tensorflow.keras.layers import Conv2D, Flatten, Dense # 멀티쓰레딩을 위한 글로벌 변수 global episode, score_avg, score_max episode, score_avg, score_max = 0, 0, 0 num_episode = 8000000 # ActorCritic 인공신경망 class ActorCritic(tf.keras.Model): def __init__(self, action_size, state_size): super(ActorCritic, self).__init__() self.conv1 = Conv2D(32, (8, 8), strides=(4, 4), activation='relu', input_shape=state_size) self.conv2 = Conv2D(64, (4, 4), strides=(2, 2), activation='relu') self.conv3 = Conv2D(64, (3, 3), strides=(1, 1), activation='relu') self.flatten = Flatten() self.shared_fc = Dense(512, activation='relu') self.policy = Dense(action_size, activation='linear') self.value = Dense(1, activation='linear') def call(self, x): x = self.conv1(x) x = self.conv2(x) x = self.flatten(x) x = self.shared_fc(x) policy = self.policy(x) value = self.value(x) return policy, value # 브레이크아웃에서의 A3CAgent 클래스 (글로벌신경망) class A3CAgent(): def __init__(self, action_size, env_name): self.env_name = env_name # 상태와 행동의 크기 정의 self.state_size = (84, 84, 4) self.action_size = action_size # A3C 하이퍼파라미터 self.discount_factor = 0.99 self.no_op_steps = 30 self.lr = 1e-4 # 쓰레드의 갯수 self.threads = 16 # 글로벌 인공신경망 생성 self.global_model = ActorCritic(self.action_size, self.state_size) # 글로벌 인공신경망의 가중치 초기화 self.global_model.build(tf.TensorShape((None, *self.state_size))) # 인공신경망 업데이트하는 옵티마이저 함수 생성 self.optimizer = AdamOptimizer(self.lr, use_locking=True) # 텐서보드 설정 self.writer = tf.summary.create_file_writer('summary/breakout_a3c') # 학습된 글로벌신경망 모델을 저장할 경로 설정 self.model_path = os.path.join(os.getcwd(), 'save_model', 'model') # 쓰레드를 만들어 학습을 하는 함수 def train(self): # 쓰레드 수 만큼 Runner 클래스 생성 runners = [Runner(self.action_size, self.state_size, self.global_model, self.optimizer, self.discount_factor, self.env_name, self.writer) for i in range(self.threads)] # 각 쓰레드 시정 for i, runner in enumerate(runners): print("Start worker #{:d}".format(i)) runner.start() # 10분 (600초)에 한 번씩 모델을 저장 while True: self.global_model.save_weights(self.model_path, save_format="tf") time.sleep(60 * 10) # 액터러너 클래스 (쓰레드) class Runner(threading.Thread): global_episode = 0 def __init__(self, action_size, state_size, global_model, optimizer, discount_factor, env_name, writer): threading.Thread.__init__(self) # A3CAgent 클래스에서 넘겨준 하이준 파라미터 설정 self.action_size = action_size self.state_size = state_size self.global_model = global_model self.optimizer = optimizer self.discount_factor = discount_factor self.states, self.actions, self.rewards = [], [], [] # 환경, 로컬신경망, 텐서보드 생성 self.local_model = ActorCritic(action_size, state_size) self.env = gym.make(env_name) self.writer = writer # 학습 정보를 기록할 변수 self.avg_p_max = 0 self.avg_loss = 0 # k-타임스텝 값 설정 self.t_max = 20 self.t = 0 # 불필요한 행동을 줄여주기 위한 dictionary self.action_dict = {0:1, 1:2, 2:3, 3:3} # 텐서보드에 학습 정보를 기록 def draw_tensorboard(self, score, step, e): avg_p_max = self.avg_p_max / float(step) with self.writer.as_default(): tf.summary.scalar('Total Reward/Episode', score, step=e) tf.summary.scalar('Average Max Prob/Episode', avg_p_max, step=e) tf.summary.scalar('Duration/Episode', step, step=e) # 정책신경망의 출력을 받아 확률적으로 행동을 선택 def get_action(self, history): history = np.float32(history / 255.) policy = self.local_model(history)[0][0] policy = tf.nn.softmax(policy) action_index = np.random.choice(self.action_size, 1, p=policy.numpy())[0] return action_index, policy # 샘플을 저장 def append_sample(self, history, action, reward): self.states.append(history) act = np.zeros(self.action_size) act[action] = 1 self.actions.append(act) self.rewards.append(reward) # k-타임스텝의 prediction 계산 def discounted_prediction(self, rewards, done): discounted_prediction = np.zeros_like(rewards) running_add = 0 if not done: # value function last_state = np.float32(self.states[-1] / 255.) running_add = self.local_model(last_state)[-1][0].numpy() for t in reversed(range(0, len(rewards))): running_add = running_add * self.discount_factor + rewards[t] discounted_prediction[t] = running_add return discounted_prediction # 저장된 샘플들로 A3C의 오류함수를 계산 def compute_loss(self, done): discounted_prediction = self.discounted_prediction(self.rewards, done) discounted_prediction = tf.convert_to_tensor(discounted_prediction[:, None], dtype=tf.float32) states = np.zeros((len(self.states), 84, 84, 4)) for i in range(len(self.states)): states[i] = self.states[i] states = np.float32(states / 255.) policy, values = self.local_model(states) # 가치 신경망 업데이트 advantages = discounted_prediction - values critic_loss = 0.5 * tf.reduce_sum(tf.square(advantages)) # 정책 신경망 업데이트 action = tf.convert_to_tensor(self.actions, dtype=tf.float32) policy_prob = tf.nn.softmax(policy) action_prob = tf.reduce_sum(action * policy_prob, axis=1, keepdims=True) cross_entropy = - tf.math.log(action_prob + 1e-10) actor_loss = tf.reduce_sum(cross_entropy * tf.stop_gradient(advantages)) entropy = tf.reduce_sum(policy_prob * tf.math.log(policy_prob + 1e-10), axis=1) entropy = tf.reduce_sum(entropy) actor_loss += 0.01 * entropy total_loss = 0.5 * critic_loss + actor_loss return total_loss # 로컬신경망을 통해 그레이디언트를 계산하고, 글로벌 신경망을 계산된 그레이디언트로 업데이트 def train_model(self, done): global_params = self.global_model.trainable_variables local_params = self.local_model.trainable_variables with tf.GradientTape() as tape: total_loss = self.compute_loss(done) # 로컬신경망의 그레이디언트 계산 grads = tape.gradient(total_loss, local_params) # 안정적인 학습을 위한 그레이디언트 클리핑 grads, _ = tf.clip_by_global_norm(grads, 40.0) # 로컬신경망의 오류함수를 줄이는 방향으로 글로벌신경망을 업데이트 self.optimizer.apply_gradients(zip(grads, global_params)) # 로컬신경망의 가중치를 글로벌신경망의 가중치로 업데이트 self.local_model.set_weights(self.global_model.get_weights()) # 업데이트 후 저장된 샘플 초기화 self.states, self.actions, self.rewards = [], [], [] def run(self): # 액터러너끼리 공유해야하는 글로벌 변수 global episode, score_avg, score_max step = 0 while episode < num_episode: done = False dead = False score, start_life = 0, 5 observe = self.env.reset() # 랜덤으로 뽑힌 값 만큼의 프레임동안 움직이지 않음 for _ in range(random.randint(1, 30)): observe, _, _, _ = self.env.step(1) # 프레임을 전처리 한 후 4개의 상태를 쌓아서 입력값으로 사용. state = pre_processing(observe) history = np.stack([state, state, state, state], axis=2) history = np.reshape([history], (1, 84, 84, 4)) while not done: step += 1 self.t += 1 # 정책 확률에 따라 행동을 선택 action, policy = self.get_action(history) # 1: 정지, 2: 왼쪽, 3: 오른쪽 real_action = self.action_dict[action] # 죽었을 때 시작하기 위해 발사 행동을 함 if dead: action, real_action, dead = 0, 1, False # 선택한 행동으로 환경에서 한 타임스텝 진행 observe, reward, done, info = self.env.step(real_action) # 각 타임스텝마다 상태 전처리 next_state = pre_processing(observe) next_state = np.reshape([next_state], (1, 84, 84, 1)) next_history = np.append(next_state, history[:, :, :, :3], axis=3) # 정책확률의 최대값 self.avg_p_max += np.amax(policy.numpy()) if start_life > info['ale.lives']: dead = True start_life = info['ale.lives'] score += reward reward = np.clip(reward, -1., 1.) # 샘플을 저장 self.append_sample(history, action, reward) if dead: history = np.stack((next_state, next_state, next_state, next_state), axis=2) history = np.reshape([history], (1, 84, 84, 4)) else: history = next_history # 에피소드가 끝나거나 최대 타임스텝 수에 도달하면 학습을 진행 if self.t >= self.t_max or done: self.train_model(done) self.t = 0 if done: # 각 에피소드 당 학습 정보를 기록 episode += 1 score_max = score if score > score_max else score_max score_avg = 0.9 * score_avg + 0.1 * score if score_avg != 0 else score log = "episode: {:5d} | score : {:4.1f} | ".format(episode, score) log += "score max : {:4.1f} | ".format(score_max) log += "score avg : {:.3f}".format(score_avg) print(log) self.draw_tensorboard(score, step, episode) self.avg_p_max = 0 step = 0 # 학습속도를 높이기 위해 흑백화면으로 전처리 def pre_processing(observe): processed_observe = np.uint8( resize(rgb2gray(observe), (84, 84), mode='constant') * 255) return processed_observe if __name__ == "__main__": global_agent = A3CAgent(action_size=3, env_name="BreakoutDeterministic-v4") global_agent.train() | cs |
test.py
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 | import gym import time import random import numpy as np import tensorflow as tf from skimage.color import rgb2gray from skimage.transform import resize from tensorflow.keras.layers import Conv2D, Flatten, Dense # ActorCritic 인공신경망 class ActorCritic(tf.keras.Model): def __init__(self, action_size, state_size): super(ActorCritic, self).__init__() self.conv1 = Conv2D(32, (8, 8), strides=(4, 4), activation='relu', input_shape=state_size) self.conv2 = Conv2D(64, (4, 4), strides=(2, 2), activation='relu') self.conv3 = Conv2D(64, (3, 3), strides=(1, 1), activation='relu') self.flatten = Flatten() self.shared_fc = Dense(512, activation='relu') self.policy = Dense(action_size, activation='linear') self.value = Dense(1, activation='linear') def call(self, x): x = self.conv1(x) x = self.conv2(x) x = self.flatten(x) x = self.shared_fc(x) policy = self.policy(x) value = self.value(x) return policy, value # 브레이크아웃에서의 테스트를 위한 A3C 에이전트 클래스 class A3CTestAgent: def __init__(self, action_size, state_size, model_path): self.action_size = action_size self.model = ActorCritic(action_size, state_size) self.model.load_weights(model_path) def get_action(self, history): history = np.float32(history / 255.) policy = self.model(history)[0][0] policy = tf.nn.softmax(policy) action_index = np.random.choice(self.action_size, 1, p=policy.numpy())[0] return action_index, policy def pre_processing(observe): processed_observe = np.uint8( resize(rgb2gray(observe), (84, 84), mode='constant') * 255) return processed_observe if __name__ == "__main__": # 테스트를 위한 환경, 모델 생성 env = gym.make("BreakoutDeterministic-v4") state_size = (84, 84, 4) action_size = 3 model_path = './save_model/trained/model' render = True agent = A3CTestAgent(action_size, state_size, model_path) action_dict = {0:1, 1:2, 2:3, 3:3} num_episode = 10 for e in range(num_episode): done = False dead = False score, start_life = 0, 5 observe = env.reset() # 랜덤으로 뽑힌 값 만큼의 프레임동안 움직이지 않음 for _ in range(random.randint(1, 30)): observe, _, _, _ = env.step(1) # 프레임을 전처리 한 후 4개의 상태를 쌓아서 입력값으로 사용. state = pre_processing(observe) history = np.stack([state, state, state, state], axis=2) history = np.reshape([history], (1, 84, 84, 4)) while not done: if render: env.render() time.sleep(0.05) # 정책 확률에 따라 행동을 선택 action, policy = agent.get_action(history) # 1: 정지, 2: 왼쪽, 3: 오른쪽 real_action = action_dict[action] # 죽었을 때 시작하기 위해 발사 행동을 함 if dead: action, real_action, dead = 0, 1, False # 선택한 행동으로 환경에서 한 타임스텝 진행 observe, reward, done, info = env.step(real_action) # 각 타임스텝마다 상태 전처리 next_state = pre_processing(observe) next_state = np.reshape([next_state], (1, 84, 84, 1)) next_history = np.append(next_state, history[:, :, :, :3], axis=3) if start_life > info['ale.lives']: dead, start_life = True, info['ale.lives'] score += reward if dead: history = np.stack((next_state, next_state, next_state, next_state), axis=2) history = np.reshape([history], (1, 84, 84, 4)) else: history = next_history if done: # 각 에피소드 당 학습 정보를 기록 print("episode: {:3d} | score : {:4.1f}".format(e, score)) | cs |
강화학습 이론 및 실습(MD) 실습