TensorFlow中获取指定迭代MSE值的问题:原因及解决方法
我在使用TensorFlow做梯度下降训练时,想要在特定的epoch和batch组合时打印MSE的实际数值,但直接打印mse得到的是Tensor对象(示例输出:Epoch 0 Batch_Index 0 MSE: Tensor("mse_2:0", shape=(), dtype=float32))。我理解这是因为MSE依赖tf.placeholder节点,但我已经在sess.run(training_op, feed_dict={X: X_batch, y: y_batch})里传入了数据,本以为此时可以获取MSE的值,结果用mse.eval()的时候报错:
InvalidArgumentError: You must feed a value for placeholder tensor 'X_2' with dtype float and shape [?,9]...
为什么会出现这种情况?该怎么修改代码才能输出指定迭代时的MSE值?
附相关代码片段:
import numpy as np from sklearn.datasets import fetch_california_housing housing = fetch_california_housing() m, n = housing.data.shape housing_data_plus_bias = np.c_[np.ones((m, 1)), housing.data] # ADD COLUMN OF 1s for BIAS! from sklearn.preprocessing import StandardScaler scaler = StandardScaler() scaled_housing_data = scaler.fit_transform(housing.data) scaled_housing_data_plus_bias = np.c_[np.ones((m, 1)), scaled_housing_data] X = tf.placeholder(tf.float32, shape=(None, n + 1), name="X") y = tf.placeholder(tf.float32, shape=(None, 1), name="y") theta = tf.Variable(tf.random_uniform([n + 1, 1], -1.0, 1.0, seed=42), name="theta") y_pred = tf.matmul(X, theta, name="predictions") error = y_pred - y mse = tf.reduce_mean(tf.square(error), name="mse") optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate) training_op = optimizer.minimize(mse) init = tf.global_variables_initializer() n_epochs = 100 batch_size = 100 n_batches = int(np.ceil(m / batch_size)) learning_rate = 0.01 def fetch_batch(epoch, batch_index, batch_size): np.random.seed(epoch * n_batches + batch_index) # not shown in the book indices = np.random.randint(m, size=batch_size) # not shown X_batch = scaled_housing_data_plus_bias[indices] # not shown y_batch = housing.target.reshape(-1, 1)[indices] # not shown return X_batch, y_batch with tf.Session() as sess: sess.run(init) for epoch in range(n_epochs): for batch_index in range(n_batches): X_batch, y_batch = fetch_batch(epoch, batch_index, batch_size) sess.run(training_op, feed_dict={X: X_batch, y: y_batch}) if (epoch % 50 == 0 and batch_index % 100 == 0): print("Epoch", epoch, "Batch_Index", batch_index, "MSE:", mse) best_theta = theta.eval()
原因分析
其实这是TensorFlow惰性执行机制导致的,给你拆解下:
- 你定义的
mse只是计算图里的一个节点,它本身不存储任何数值,每次要获取它的结果,都必须给它依赖的所有占位符(也就是X和y)喂入数据。 - 你之前调用
sess.run(training_op)的时候,确实传入了feed_dict,但这只是让TensorFlow执行了training_op对应的计算(也就是更新参数theta),并没有同时计算mse的值,更不会把这次喂的数据缓存下来供后续使用。mse.eval()本质上等价于sess.run(mse),但这时候你没传feed_dict,TensorFlow找不到X和y的数据,自然就报错了。
解决方法
这里有两种常用的方式,你可以根据需求选:
方法一:同一次sess.run()中同时执行训练和计算MSE(推荐)
这种方式效率更高,因为不需要重复喂数据:
修改循环内的代码,把训练和MSE计算合并到一次sess.run()调用中:
with tf.Session() as sess: sess.run(init) for epoch in range(n_epochs): for batch_index in range(n_batches): X_batch, y_batch = fetch_batch(epoch, batch_index, batch_size) if (epoch % 50 == 0 and batch_index % 100 == 0): # 同时执行训练操作和计算MSE,返回对应的结果 _, current_mse = sess.run([training_op, mse], feed_dict={X: X_batch, y: y_batch}) print("Epoch", epoch, "Batch_Index", batch_index, "MSE:", current_mse) else: # 只执行训练操作 sess.run(training_op, feed_dict={X: X_batch, y: y_batch}) best_theta = theta.eval()
sess.run()可以接收一个节点列表,一次性执行多个计算,返回对应顺序的结果。这里_用来接收training_op的返回值(它没有实际意义,因为训练操作只是更新参数),current_mse就是当前batch的MSE实际数值。
方法二:单独计算MSE并传入feed_dict
如果调试时想单独查看MSE,也可以在需要打印的时候单独调用sess.run(mse),但必须再次传入feed_dict:
with tf.Session() as sess: sess.run(init) for epoch in range(n_epochs): for batch_index in range(n_batches): X_batch, y_batch = fetch_batch(epoch, batch_index, batch_size) sess.run(training_op, feed_dict={X: X_batch, y: y_batch}) if (epoch % 50 == 0 and batch_index % 100 == 0): # 单独计算MSE,记得传入feed_dict current_mse = sess.run(mse, feed_dict={X: X_batch, y: y_batch}) # 或者用 mse.eval(feed_dict={X: X_batch, y: y_batch}) 效果一样 print("Epoch", epoch, "Batch_Index", batch_index, "MSE:", current_mse) best_theta = theta.eval()
这种方式会多一次MSE的计算过程,效率比方法一低一点,但逻辑更直观,适合调试阶段使用。
内容的提问来源于stack exchange,提问作者sebtac

