我正在使用卷积神经网络（tensorflow，alexnet）在游戏中驾驶汽车（极品飞车）。我只是一个想要使用机器学习的学生。我是初学者。
这是我的计划：
问题在于我没有找到一种方法来记录或检测我在玩游戏时所做的KeyPress。我希望python检测我按下哪些键，并将它们与游戏的帧图像一起存储在npy数组中。
我确实遇到了这段记录KeyPresses的代码，但只记录了字母键。我希望python也检测箭头键，空格键等。
import win32api as wapio
import time
keyList = ["\b"]
for char in "ABCDEFGHIJKLMNOPQRSTUVWXYZ 123456789,.'£$/\\":
keyList.append(char)
def key_check():
keys = []
for key in keyList:
if wapi.GetAsyncKeyState(ord(key)):
keys.append(key)
return keys
我还想要一种模拟KeyPresses的方法，这样我的模型实际上可以开车。
我确实拥有这段代码，完美无缺。我只想就这个问题提出建议。
import ctypes
import time
SendInput = ctypes.windll.user32.SendInput
W = 0x11
A = 0x1E
S = 0x1F
D = 0x20
# C struct redefinitions
PUL = ctypes.POINTER(ctypes.c_ulong)
class KeyBdInput(ctypes.Structure):
_fields_ = [("wVk", ctypes.c_ushort),
("wScan", ctypes.c_ushort),
("dwFlags", ctypes.c_ulong),
("time", ctypes.c_ulong),
("dwExtraInfo", PUL)]
class HardwareInput(ctypes.Structure):
_fields_ = [("uMsg", ctypes.c_ulong),
("wParamL", ctypes.c_short),
("wParamH", ctypes.c_ushort)]
class MouseInput(ctypes.Structure):
_fields_ = [("dx", ctypes.c_long),
("dy", ctypes.c_long),
("mouseData", ctypes.c_ulong),
("dwFlags", ctypes.c_ulong),
("time",ctypes.c_ulong),
("dwExtraInfo", PUL)]
class Input_I(ctypes.Union):
_fields_ = [("ki", KeyBdInput),
("mi", MouseInput),
("hi", HardwareInput)]
class Input(ctypes.Structure):
_fields_ = [("type", ctypes.c_ulong),
("ii", Input_I)]
# Actuals Functions
def PressKey(hexKeyCode):
extra = ctypes.c_ulong(0)
ii_ = Input_I()
ii_.ki = KeyBdInput( 0, hexKeyCode, 0x0008, 0, ctypes.pointer(extra) )
x = Input( ctypes.c_ulong(1), ii_ )
ctypes.windll.user32.SendInput(1, ctypes.pointer(x), ctypes.sizeof(x))
def ReleaseKey(hexKeyCode):
extra = ctypes.c_ulong(0)
ii_ = Input_I()
ii_.ki = KeyBdInput( 0, hexKeyCode, 0x0008 | 0x0002, 0,
ctypes.pointer(extra) )
x = Input( ctypes.c_ulong(1), ii_ )
ctypes.windll.user32.SendInput(1, ctypes.pointer(x), ctypes.sizeof(x))
if __name__ == '__main__':
while (True):
PressKey(0x11)
time.sleep(1)
ReleaseKey(0x11)
time.sleep(1)
下一步是实际收集游戏的帧（图像）。我使用cv2来捕捉我的屏幕区域，在那里我玩游戏。然后我：
将其转换为灰度
调整图片大小
将其保存为numpy文件（每捕获1000帧）
数据捕获完全正常，但是一旦我有大约50 000帧（或大约500 MB的数据），保存数据需要很长时间。有时，Python会在numpy文件保存时崩溃，导致我的所有训练数据丢失。
这是我捕获数据的代码：
import numpy as np
import cv2
import time
from grabscreen import grab_screen
from getkeys import key_check
import os
def keys_to_output(keys):
output = [0, 0, 0]
if 'A' in keys:
output[0] = 1
elif 'D' in keys:
output[2] = 1
else:
output[1] = 1
return output
file_name = 'training_data.npy'
if os.path.isfile(file_name):
print('File exists, loading previous data!')
training_data = list(np.load(file_name))
else:
print('File does not exist, starting fresh!')
training_data = []
def main():
for i in list(range(5))[::-1]:
print(i + 1)
time.sleep(1)
if os.path.isfile(file_name):
print('Existing Training Data:' + str(len(training_data)))
print('Capturing Data!')
else:
print('Capturing Data Freshly!')
while True:
screen = grab_screen(region=(40, 250, 860, 560))
screen = cv2.resize(screen, (120, 56))
screen = cv2.cvtColor(screen, cv2.COLOR_BGR2GRAY)
keys = key_check()
output = keys_to_output(keys)
training_data.append([screen, output])
if len(training_data) % 1000 == 0:
print('New Training Data: ' + str(len(training_data)))
print('Saving Data!')
np.save(file_name, training_data)
print('Data saved succesfully! You can quit now.')
print('Capturing data!')
main()
任何人都可以建议更好的方法来获取我的训练数据。另外，我听说过使用PyTables，但我不确定如何将它用于我的程序。
在此之后，我将我的数据洗牌，并平衡它：
import numpy as np
import pandas as pd
from collections import Counter
from random import shuffle
train_data = np.load('training_data.npy')
print('Training Data: ' + str(len(train_data)))
df = pd.DataFrame(train_data)
print(df.head())
print(Counter(df[1].apply(str)))
lefts = []
rights = []
forwards = []
shuffle(train_data)
for data in train_data:
img = data[0]
choice = data[1]
if choice == [1, 0, 0]:
lefts.append([img, choice])
elif choice == [0, 1, 0]:
forwards.append([img, choice])
elif choice == [0, 0, 1]:
rights.append([img, choice])
else:
print('no matches!!!')
forwards = forwards[:len(lefts)][:len(rights)]
lefts = lefts[:len(forwards)]
rights = rights[:len(forwards)]
final_data = forwards + lefts + rights
shuffle(final_data)
print('Final Balanced Data: ' + str(len(final_data)))
np.save('training_data_balanced.npy', final_data)
型号：
import tflearn
from tflearn.layers.conv import conv_2d, max_pool_2d
from tflearn.layers.core import input_data, dropout, fully_connected
from tflearn.layers.estimator import regression
from tflearn.layers.normalization import local_response_normalization
def alexnet(width, height, lr):
network = input_data(shape=[None, width, height, 1], name='input')
network = conv_2d(network, 96, 11, strides=4, activation='relu')
network = max_pool_2d(network, 3, strides=2)
network = local_response_normalization(network)
network = conv_2d(network, 256, 5, activation='relu')
network = max_pool_2d(network, 3, strides=2)
network = local_response_normalization(network)
network = conv_2d(network, 384, 3, activation='relu')
network = conv_2d(network, 384, 3, activation='relu')
network = conv_2d(network, 256, 3, activation='relu')
network = max_pool_2d(network, 3, strides=2)
network = local_response_normalization(network)
network = fully_connected(network, 4096, activation='tanh')
network = dropout(network, 0.5)
network = fully_connected(network, 4096, activation='tanh')
network = dropout(network, 0.5)
network = fully_connected(network, 3, activation='softmax')
network = regression(network, optimizer='momentum',
loss='categorical_crossentropy',
learning_rate=lr, name='targets')
model = tflearn.DNN(network, checkpoint_path='model_alexnet',
max_checkpoints=1, tensorboard_verbose=2, tensorboard_dir='log')
return model
训练模型：
import numpy as np
from alexnet import alexnet
WIDTH = 120
HEIGHT = 56
LR = 1e-3
EPOCHS = 15
MODEL_NAME = 'nfs-car-{}-{}-epochs.model'.format(LR, EPOCHS)
model = alexnet(WIDTH, HEIGHT, LR)
# for every epoch finished, save the model.
# therefore, if accuracy drops or loss increases, we can terminate the script.
# we will have a trained model with the best accuracy and previously saved epoch.
for i in range(EPOCHS):
train_data = np.load('training_data_balanced.npy')
train = train_data[:-10] # 80% of balanced data
test = train_data[-10:] # 20% of balanced data
X = np.array([i[0] for i in train]).reshape(-1, WIDTH, HEIGHT, 1)
Y = [i[1] for i in train]
test_x = np.array([i[0] for i in test]).reshape(-1, WIDTH, HEIGHT, 1)
test_y = [i[1] for i in test]
model.fit({'input': X}, {'targets': Y}, n_epoch=1,
validation_set=({'input': test_x}, {'targets': test_y}),
snapshot_step=500, show_metric=True, run_id=MODEL_NAME)
model.save(MODEL_NAME)
print('Saved epoch: ' + str(i + 1))
然后，当然，我测试模型：
import numpy as np
import cv2
import time
from grabscreen import grab_screen
from getkeys import key_check
from alexnet import alexnet
from directkeys import PressKey, ReleaseKey, W, A, D
WIDTH = 120
HEIGHT = 56
LR = 1e-3
EPOCHS = 15
MODEL_NAME = 'nfs-car-{}-{}-epochs.model'.format(LR, EPOCHS)
def straight():
PressKey(W)
ReleaseKey(A)
ReleaseKey(D)
def left():
PressKey(W)
PressKey(A)
ReleaseKey(D) # added
time.sleep(0.09)
ReleaseKey(A)
def right():
PressKey(W)
PressKey(D)
ReleaseKey(A) # added
time.sleep(0.09)
ReleaseKey(D)
model = alexnet(WIDTH, HEIGHT, LR)
model.load(MODEL_NAME)
def main():
for i in list(range(5))[::-1]:
print(i + 1)
time.sleep(1)
paused = False
while True:
if not paused:
screen = grab_screen(region=(40, 250, 860, 560))
screen = cv2.cvtColor(screen, cv2.COLOR_BGR2GRAY)
screen = cv2.resize(screen, (120, 56))
prediction = model.predict([screen.reshape(WIDTH, HEIGHT, 1)])[0]
print(prediction)
turn_thrush = .75
fwd_thrush = 0.70
if prediction[0] > turn_thrush:
left()
elif prediction[1] > fwd_thrush:
straight()
elif prediction[2] > turn_thrush:
right()
else:
straight()
keys = key_check()
if 'T' in keys:
if paused:
paused = False
print('Unpaused!')
time.sleep(1)
else:
print('Pausing!')
paused = True
ReleaseKey(A)
ReleaseKey(W)
ReleaseKey(D)
time.sleep(1)
main()
所以是的，这就是我的想法。然而，由于缺乏适当的信息，我无法完全应用这一概念。
答案 0 :(得分：0)
请查看此完整的实现，包括该项目的模拟器，模型训练，测试数据，源代码和摘要。