我正在使用具有tensorlow后端的keras来创建Deep Q学习代理，以在openai健身房玩atari游戏。但是当我训练模型时，我的GPU利用率保持在8％到10％之间。我是新手，无法弄清楚如何提高我的GPU利用率。您能否提供一些改善建议？ 这是代码：
import gym
import random
import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
import matplotlib.animation as anim
import time
from keras.models import Sequential
from keras.layers import Conv2D, Dense, Flatten, Lambda
from keras.optimizers import RMSprop
from keras import backend as k
from skimage.color import rgb2gray
from skimage.transform import resize
from collections import deque
class DQNAgent() :
def __init__(self, n_actions):
self.learning_rate = 0.00025
self.epsilon = 1.0
self.epsilon_min = 0.1
self.epsilon_decay = 0.0001
self.gamma = 0.99
self.n_actions = n_actions
self.batch_size = 32
self.model = self.create_model()
self.memory = deque(maxlen=100000)
def create_model(self) :
model = Sequential()
model.add(Lambda(lambda x : x/255.0, input_shape=(84, 84, 4)))
model.add(Conv2D(filters=16, kernel_size=(8,8), strides=(4,4), activation='relu'))
model.add(Conv2D(filters=32, kernel_size=(4,4), strides=(2,2), activation='relu'))
model.add(Flatten())
model.add(Dense(units=256, activation='relu'))
model.add(Dense(units=self.n_actions))
model.compile(optimizer=RMSprop(learning_rate=self.learning_rate, rho=0.95, epsilon=0.01), loss=huber_loss)
return model
def act(self, state) :
if random.random() <= self.epsilon :
return random.randint(0, self.n_actions - 1)
return(np.argmax(self.model.predict(state)[0]))
def remember(self, state, action, reward, next_state, dead) :
self.memory.append((state, action, reward, next_state, dead))
def replay(self) :
mini_batch = random.sample(self.memory, self.batch_size)
state = np.zeros((self.batch_size, 84, 84, 4))
next_state = np.zeros_like(state)
target = np.zeros((self.batch_size,))
action, reward, dead = [], [], []
for idx, val in enumerate(mini_batch) :
state[idx] = val[0]
action.append(val[1])
reward.append(val[2])
next_state[idx] = val[3]
dead.append(val[4])
future_q = self.model.predict(next_state, batch_size=self.batch_size)
for i in range(self.batch_size) :
if dead[i] :
target[i] = -1
else :
target[i] = reward[i] + self.gamma*np.amax(future_q[i])
action_one_hot = get_one_hot(action, self.n_actions)
target_one_hot = action_one_hot * target[:, None]
loss = self.model.fit(state, target_one_hot, batch_size=self.batch_size, epochs=1, verbose=0).history['loss'][0]
return loss
def preprocess(self, image) :
return np.uint8(resize(rgb2gray(image), output_shape=(84, 84), mode='constant') * 255)
def save_model(self) :
self.model.save_weights('model.json')
def load_model(self) :
self.model.load_weights('model.json')
def get_one_hot(arr, num) :
return np.eye(num)[np.array(arr).reshape(-1)]
def huber_loss(y, q_value):
error = k.abs(y - q_value)
quadratic_part = k.clip(error, 0.0, 1.0)
linear_part = error - quadratic_part
loss = k.mean(0.5 * k.square(quadratic_part) + linear_part)
return loss
def train(resume=False) :
env = gym.make('BreakoutDeterministic-v4')
agent = DQNAgent(env.action_space.n)
for i in range(1000) :
state = env.reset()
if resume :
agent.load_model()
# Do no operation for 30 iterations
for _ in range(30) :
state, _, _, _ = env.step(1)
state = agent.preprocess(state)
state = np.stack((state, state, state, state), axis = 2)
state = np.reshape(state, (1, 84, 84, 4))
done, dead = False, False
score, loss, lives = 0, 0, 5
while not done :
env.render()
# Select action based on the state
action = agent.act(state)
if len(agent.memory) > 5000 and agent.epsilon > agent.epsilon_min :
agent.epsilon -= agent.epsilon_decay
# Take a step in the environment
next_state, reward, done, info = env.step(action)
score += reward
if lives > info['ale.lives'] :
dead = True
lives = info['ale.lives']
next_state = agent.preprocess(next_state)
next_state = np.reshape(next_state, (1, 84, 84, 1))
next_state = np.append(next_state, state[:,:,:,:3], axis = 3)
# Store into memory
agent.remember(state, action, reward, next_state, dead)
# if enough memory size start training
if len(agent.memory) > 5000 :
loss += agent.replay()
if dead :
dead = False
else :
state = next_state
if done :
print("Episode : {0}, score : {1}, loss : {2}, memory size : {3}".format(i, score, loss, len(agent.memory)))
env.close()
agent.save_model()
def test() :
env = gym.make('BreakoutDeterministic-v4')
agent = DQNAgent(env.action_space.n)
agent.load_model()
for i in range(100) :
state = env.reset()
for _ in range(30) :
state, _, _, _ = env.step(0)
state = agent.preprocess(state)
state = np.stack((state, state, state, state), axis = 2)
state = np.reshape(state, (1, 84, 84, 4))
done, dead = False, False
score, lives = 0, 5
while not done :
env.render()
action = agent.act(state)
next_state, reward, done, info = env.step(action)
score += reward
if lives > info['ale.lives'] :
dead = True
lives = info['ale.lives']
next_state = agent.preprocess(next_state)
next_state = np.reshape(next_state, (1, 84, 84, 1))
next_state = np.append(next_state, state[:,:,:,:3], axis = 3)
if dead :
dead = False
else :
state = next_state
if done :
print("Episode : {0}, score : {1}".format(i, score))
if __name__ == "__main__":
train(False)
#test()
一旦将model.fit称为gpu，使用率就会急剧下降。所以我想问题出在那儿？
我尝试增加批量大小，但是这使我的gpu使用率仅为9％到11％。
我正在使用笔记本电脑：
Nvidia GTX 1050ti
8 GB内存
2.20 GHz的i7-8750H处理器