def test(signal_re, labels_re):
# 创建一个默认图,在该图中执行以下操作(多数操作和train中一样)
with tf.Graph().as_default() as g:
x = tf.placeholder(tf.float32, [
test_num_examples, deap_forward.IMAGE_SIZE_X,
deap_forward.IMAGE_SIZE_Y, deap_forward.NUM_CHANNELS
])
y_ = tf.placeholder(tf.float32, [None, deap_forward.OUTPUT_NODE]) # ??
y = deap_forward.forward(x, False, None) # ??
ema = tf.train.ExponentialMovingAverage(
deap_backward.MOVING_AVERAGE_DECAY)
ema_restore = ema.variables_to_restore()
saver = tf.train.Saver(ema_restore)
# correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1)) # 判断预测值和实际值是否相同
# correct_prediction = (abs(y-y_) < 0.5) # 预测结果与真实结果相差是否小于0.5
# accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32)) # 求平均值得到准确率
while True:
with tf.Session() as sess:
ckpt = tf.train.get_checkpoint_state(
deap_backward.MODEL_SAVE_PATH)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
# 根据读入的模型名字切分出该模型是属于迭代了多少次保存的
global_step = ckpt.model_checkpoint_path.split(
'/')[-1].split('-')[-1]
reshaped_x = np.reshape(
signal_re,
(test_num_examples, deap_forward.IMAGE_SIZE_X,
deap_forward.IMAGE_SIZE_Y, deap_forward.NUM_CHANNELS))
# 计算出测试集上准确率
pred_labels = sess.run(y, feed_dict={x: reshaped_x})
y_ = labels_re
pred_correct = 0
for i in range(test_num_examples):
if pred_labels[i] >= 0.5:
pred_labels[i] = 1
else:
pred_labels[i] = 0
if pred_labels[i] == y_[i]:
pred_correct += 1
Recognition_rate = pred_correct / test_num_examples
# pred_labels = np.asarray(reshaped_x)
# print(pred_labels) # 测试用
print("----------我是可爱的分割线----------") # debug用
# print(labels_re) # 测试用
print("After %s training step(s), test accuracy = %g" %
(global_step, Recognition_rate))
else:
print('No checkpoint file found')
return
time.sleep(TEST_INTERVAL_SECS) # 每隔TEST_INTERVAL_SECS秒寻找一次是否有最新模型
def backward(signal_re, labels_re):
# x,y_是定义的占位符,需要指定参数的类型,维度(要和网络的输入与输出维度一致),类似于函数的形参,运行时必须输入的参数
x = tf.placeholder(tf.float32, [
BATCH_SIZE, deap_forward.IMAGE_SIZE_X, deap_forward.IMAGE_SIZE_Y,
deap_forward.NUM_CHANNELS
])
y_ = tf.placeholder(tf.float32, [None, deap_forward.OUTPUT_NODE])
# 调用前向传播网络得到维度为8的tensor
y = deap_forward.forward(x, True, REGULARIZER)
# 声明一个全局计数器,并输出化为0 并说明是不参与训练过程的
global_step = tf.Variable(0, trainable=False)
# 先是对网络最后一层的输出y做softmax,通常是求取输出属于某一类的概率,其实就是num_classes的大小的向量
# 再将此向量和实际标签值做交叉熵,需要说明的是该函数的返回是一个向量
# ce = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=y, labels=tf.argmax(y_, 1)) # 不会用 暂放
# ce = tf.square(y - y_) # 有问题
# 再对得到的向量求均值就得到loss
cem = tf.reduce_mean(tf.square(y - y_))
loss = cem + tf.add_n(tf.get_collection('losses')) # 添加正则化中的losses
# 实现指数级的减小学习率,可以让模型在训练的前期快速接近较优解，又可以保证模型在训练后期不会有太大波动
# 计算公式:decayed_learning_rate = learning_rate*decay_rate^(global_step/decay_steps)
learning_rate = tf.train.exponential_decay(
LEARNING_RATE_BASE,
global_step,
40 / BATCH_SIZE, # 每几轮改变一次
LEARNING_RATE_DECAY,
staircase=True) # True(global_step/decay_steps)取整,阶梯 False 平滑
# 传入学习率,构造一个实现梯度下降算法的优化器,再通过使用minimize更新存储要训练的变量列表来减少loss
train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(
loss, global_step=global_step)
# 实现滑动平均模型,参数MOVING_AVERAGE_DECAY用于控制模型更新速度.训练过程中会对每一个变量维护一个影子变量,这个影子变量的初始值
# 就是相应变量的初始值,每次变量更新时,影子变量就会随之更新
ema = tf.train.ExponentialMovingAverage(MOVING_AVERAGE_DECAY, global_step)
ema_op = ema.apply(tf.trainable_variables())
with tf.control_dependencies([train_step, ema_op
]): # 将train_step和eam_op两个训练操作绑定到train_op上
train_op = tf.no_op(name='train')
saver = tf.train.Saver() # 实例化一个保存和恢复变量的saver
# 创建一个会话,并通过python中的上下文管理器来管理这个会话
with tf.Session() as sess:
init_op = tf.global_variables_initializer() # 初始化计算图中的变量
sess.run(init_op)
ckpt = tf.train.get_checkpoint_state(
MODEL_SAVE_PATH) # 通过checkpoint文件定位到最新保存的模型
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path) # 加载最新的模型
for i in range(STEPS):
start = (i * BATCH_SIZE) % BATCH_SIZE_ALL
end = start + BATCH_SIZE
xs = signal_re
ys = labels_re[start:end] # 读取一个batch的数据
reshaped_xs = np.reshape( # 将输入数据xs转换成与网络输入相同形状的矩阵
xs[start:end], # 读取一个batch的数据
(BATCH_SIZE, deap_forward.IMAGE_SIZE_X,
deap_forward.IMAGE_SIZE_Y, deap_forward.NUM_CHANNELS))
# 喂入训练图像&标签, 开始训练
# print(reshaped_xs) # debug
# print(ys) # debug
# print("-------------我是分割线-------------") # debug用
_, loss_value, step = sess.run([train_op, loss, global_step],
feed_dict={
x: reshaped_xs,
y_: ys
})
# print("-------------我是分割线-------------") # debug用
if i % 10 == 0: # 每迭代10次打印loss信息,20000次保存最新的模型
print(
"After %d training step(s), loss on training batch is %g."
% (step, loss_value))
if i % 20000 == 0 and i != 0:
saver.save(sess,
os.path.join(MODEL_SAVE_PATH, MODEL_NAME),
global_step=global_step)
print("After %d training step(s), Model has been saved." %
step)
