# 输入：100*100的灰度图片，前面的None是batch size，这里都为1 
x = tf.placeholder(tf.float32, shape=[None, 100, 100, 1]) 
# 输出：一个浮点数，就是按压时间，单位s 
y_ = tf.placeholder(tf.float32, shape=[None, 1]) 
 
# 第一层卷积 12个feature map 
W_conv1 = weight_variable([5, 5, 1, 12], 0.1) 
b_conv1 = bias_variable([12], 0.1) 
# 卷积后为96*96*12 
 
h_conv1 = tf.nn.relu(conv2d(x, W_conv1) + b_conv1) 
h_pool1 = max_pool_4x4(h_conv1) 
# 池化后为24*24*12 
 
# 第二层卷积 24个feature map 
W_conv2 = weight_variable([5, 5, 12, 24], 0.1) 
b_conv2 = bias_variable([24], 0.1) 
# 卷积后为20*20*24 
 
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2) 
h_pool2 = max_pool_4x4(h_conv2) 
# 池化后为5*5*24 
 
# 全连接层5*5*24 --> 32 
W_fc1 = weight_variable([5 * 5 * 24, 32], 0.1) 
b_fc1 = bias_variable([32], 0.1) 
h_pool2_flat = tf.reshape(h_pool2, [-1, 5 * 5 * 24]) 
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1) 
 
# drapout，play时为1训练时为0.6 
keep_prob = tf.placeholder(tf.float32) 
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob) 
# 学习率 
learn_rate = tf.placeholder(tf.float32) 
 
# 32 --> 1 
W_fc2 = weight_variable([32, 1], 0.1) 
b_fc2 = bias_variable([1], 0.1) 
y_fc2 = tf.matmul(h_fc1_drop, W_fc2) + b_fc2 
 
# 因输出直接是时间值，而不是分类概率，所以用平方损失 
cross_entropy = tf.reduce_mean(tf.square(y_fc2 - y_)) 
train_step = tf.train.AdamOptimizer(learn_rate).minimize(cross_entropy) 

# 获取屏幕截图并转换为模型的输入 
def get_screen_shot(): 
  # 使用adb命令截图并获取图片，这里如果把后缀改成jpg会导致TensorFlow读不出来 
  os.system('adb shell screencap -p /sdcard/jump_temp.png') 
  os.system('adb pull /sdcard/jump_temp.png .') 
  # 使用PIL处理图片，并转为jpg 
  im = Image.open(r"./jump_temp.png") 
  w, h = im.size 
  # 将图片压缩，并截取中间部分，截取后为100*100 
  im = im.resize((108, 192), Image.ANTIALIAS) 
  region = (4, 50, 104, 150) 
  im = im.crop(region) 
  # 转换为jpg 
  bg = Image.new("RGB", im.size, (255, 255, 255)) 
  bg.paste(im, im) 
  bg.save(r"./jump_temp.jpg") 
 
  img_data = tf.image.decode_jpeg(tf.gfile.FastGFile('./jump_temp.jpg', 'rb').read()) 
  # 使用TensorFlow转为只有1通道的灰度图 
  img_data_gray = tf.image.rgb_to_grayscale(img_data) 
  x_in = np.asarray(img_data_gray.eval(), dtype='float32') 
 
  # [0,255]转为[0,1]浮点 
  for i in range(len(x_in)): 
    for j in range(len(x_in[i])): 
      x_in[i][j][0] /= 255 
 
  # 因为输入shape有batch维度，所以还要套一层 
  return [x_in] 

while True: 
 
  ………… 
 
  # 每训练100个保存一次 
  if train_count % 100 == 0: 
    saver_init.save(sess, "./save/mode.mod") 
 
  …………   
  
  sess.run(train_step, feed_dict={x: x_in, y_: y_out, keep_prob: 0.6, learn_rate: 0.00005})