【学习笔记】Hands-on ML with sklearn&tensorflow [TF] [2]placeholder nodes实现mini-batch

为了实现mini-batch，需要一种节点，在每次迭代使用一个新的batch，可以用placeholder node实现这个功能。

>>>A = placeholder(tf.float32, shape=(None, 3))
>>>B = A + 5

#这里shape=(None, 3)限制了向A节点feed data的维数

使用eval()的feed_dict参数传入A的值并计算相应的B.eval()：

>>>with tf.Session() as less:
...    B_val_1 = B.eval(feed_dict={A: [[1, 2, 3]]})
...    B_val_2 = B.eval(feed_fict={A: [[4, 5, 6], [7, 8, 9]]})

>>>print(B_val_1)
[[6. 7. 8.]]
>>>print(B_val_2)
[[ 9. 10. 11.]
 [12. 13. 14.]]

mini-batch的实现：

X = tf.placeholder(tf.float32, shape = (None, n + 1)), name = 'X')
y = tf.placeholder(tf.float32, shape = (None, 1)), name = 'y')

batch_size = 100 #设置batch大小
n_batches = int(np.ceil(m/batch_size))

def fetch_batch(epoch, batch_index, batch_size):
    [...] #读入数据
    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(train_op, feed_dict = {X: X_batch, y: y_batch})
    
    best_theta = theta.eval()