Question

我尝试设置批量大小并运行Autoencoder程序，因为没有足够的内存来使用完整批次。所以我尝试使用tf.train.batch。但由于函数的参数是张量，我试图将np数组转换为tf.convert_to_tensor的张量。但内存超过2GB，无法改变为张量。我如何用小批量训练？下面是我的代码。

N_img=47000000
batch_size=100
X_train = np.zeros(shape=(N_img, Freq_LEN, LOOK_LEN, 1), dtype='float32')
x = tf.placeholder(tf.float32, [None, FRM_LEN/2,FRM_LEN/2,1])  #FRM_LEN=256
y = tf.placeholder(tf.float32, [None, FRM_LEN/2,FRM_LEN/2,1])
X_train=tf.convert_to_tensor(X_train)
X_train_batch= tf.train.batch(X_train,batch_size=batch_size)

print("Start training..")

for step in range(n_iters):
    sess.run(optm, feed_dict={x: X_train_batch, y: X_train_batch, keepprob: 0.7})
    if step % 100 == 0:
       print(step,sess.run(cost, feed_dict={x: X_train_batch, y: X_train_batch, keepprob: 1}))

print("finish training")

Answer 1

尝试创建一个不需要张量参数的自定义generate_batch函数（以避免tf.convert_to_tensor操作），例如：

import numpy as np
batch_size = 100

X_train = np.zeros(shape=(N_img, Freq_LEN, LOOK_LEN, 1), dtype='float32')
y_train = np.zeros(shape=(N_img, Freq_LEN, LOOK_LEN, 1), dtype='float32')

data_index = 0

def generate_batch(batch_size):
    global data_index
    batch = np.ndarray(shape=(batch_size, Freq_LEN, LOOK_LEN, 1), dtype=np.float32)  #the same shapes as train data
    labels = np.ndarray(shape=(batch_size, Freq_LEN, LOOK_LEN, 1), dtype=np.float32)
    for i in range(batch_size):
        batch[i] = X_train[data_index]
        labels[i] = y_train[data_index]
        data_index = (data_index + 1) % len(X_train)
    return batch, labels

for step in range(n_iters):
    X_train_batch, X_train_batch = generate_batch(batch_size)
    sess.run(optm, feed_dict={x: X_train_batch, y: X_train_batch, keepprob: 0.7})
    if step % 100 == 0:
        print(step,sess.run(cost, feed_dict={x: X_train_batch, y: X_train_batch, keepprob: 1}))

Answer 2

强制您的批处理在CPU上进行：

....
with tf.device('/cpu:0'):
    x = tf.placeholder(tf.float32, [None, FRM_LEN/2,FRM_LEN/2,1])  #FRM_LEN=256
    y = tf.placeholder(tf.float32, [None, FRM_LEN/2,FRM_LEN/2,1])
    X_train=tf.convert_to_tensor(X_train)
    X_train_batch= tf.train.batch(X_train,batch_size=batch_size)
print("Start training..")
....

在python tensorflow中划分批处理

2 个答案: