1

我已将数据保存到 tfrecord 文件中。它有 1000 个样本和 2 个特征（一个是输入，另一个是输出）。输入是形状 [1,20] 和输出 [1,10]。它们都是从扁平的 numpy 数组创建的。我正在尝试从它们创建批次，以便我可以使用它们来训练我的网络，但我无法弄清楚如何。

这是我训练网络的代码

learning_rate = 0.01
epochs = 2
batch_size = 200 #total 5 batches
dataSize = 1000

dataset = rd.getData()

x = tf.placeholder(shape=(None,20), dtype=tf.float32)
y = tf.placeholder(shape=(None,10), dtype=tf.float32)

w1 = tf.Variable(tf.random_normal([20, 20], stddev=0.03))
w2 = tf.Variable(tf.random_normal([20, 20], stddev=0.03))
w3 = tf.Variable(tf.random_normal([20, 20], stddev=0.03))
w4 = tf.Variable(tf.random_normal([20, 20], stddev=0.03))
w5 = tf.Variable(tf.random_normal([20, 10], stddev=0.03))

b1 = tf.Variable(tf.random_normal([20]))
b2 = tf.Variable(tf.random_normal([20]))
b3 = tf.Variable(tf.random_normal([20]))
b4 = tf.Variable(tf.random_normal([20]))
b5 = tf.Variable(tf.random_normal([10]))

out1 = tf.add(tf.matmul(x, w1), b1)
out1 = tf.tanh(out1)

out2 = tf.add(tf.matmul(out1, w2), b2)
out2 = tf.tanh(out2)

out3 = tf.add(tf.matmul(out2, w3), b3)
out3 = tf.tanh(out3)

out4 = tf.add(tf.matmul(out3, w4), b4)
out4 = tf.tanh(out4)

out5 = tf.add(tf.matmul(out4, w5), b5)
finalOut = tf.tanh(out5)

cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=finalOut))


optimiser = tf.train.RMSPropOptimizer(learning_rate=learning_rate).minimize(cost)

# finally setup the initialisation operator
init_op = tf.global_variables_initializer()


with tf.Session() as sess:
   # initialise the variables
  sess.run(init_op)
  total_batch = int(dataSize / batch_size)
  for epoch in range(epochs):

    iterator = dataset.make_one_shot_iterator()
    avg_cost = 0

    for i in range(total_batch):

      #create batch
      batch_y = []
      batch_x = []
      for counter in range(0,batch_size):
        uv, z = iterator.get_next()
        batch_x.append(uv)
        batch_y.append(z)

      _, c = sess.run([optimiser, cost], 
                     feed_dict={x: batch_x, y: batch_y})
      avg_cost += c / total_batch
      print("Epoch:", (epoch + 1), "cost =", "{:.3f}".format(avg_cost))

这是我从中获取数据的文件。

def decode(serialized_example):

  features = tf.parse_single_example(
  serialized_example,
  features={'uv': tf.FixedLenFeature([1,20], tf.float32),
            'z': tf.FixedLenFeature([1,10], tf.float32)}) 

  return features['uv'], features['z']


def getData():

  filename = ["train.tfrecords"]
  dataset = tf.data.TFRecordDataset(filename).map(decode)
  return dataset

错误：

Traceback (most recent call last):
 File "network.py", line 102, in <module>
feed_dict={x: batch_x, y: batch_y})
  File "C:\Users\User\AppData\Local\Programs\Python\Python36\lib\site-packages\tensorflow\python\client\session.py", line 889, in run
run_metadata_ptr)
  File "C:\Users\User\AppData\Local\Programs\Python\Python36\lib\site-packages\tensorflow\python\client\session.py", line 1089, in _run
np_val = np.asarray(subfeed_val, dtype=subfeed_dtype)
  File "C:\Users\User\AppData\Roaming\Python\Python36\site-packages\numpy\core\numeric.py", line 531, in asarray
return array(a, dtype, copy=False, order=order)
ValueError: setting an array element with a sequence.

在看了其他问题之后，我在想也许我的批次应该是 ndarray 什么的？但我不知道如何将我的数据集变成那种形式。我什至无法弄清楚如何在没有迭代器的情况下使用我的数据。任何指导都会很棒！谢谢

4

1 回答 1

2

请尝试关注，看看是否有帮助。

  1. tf.parse_single_example不接收批次维度。因此，

    features = tf.parse_single_example(
    serialized_example,
    features={'uv': tf.FixedLenFeature([20], tf.float32),
        'z': tf.FixedLenFeature([10], tf.float32)})

  2. 从TensorFlow Guide on Dataset API 的 Simple Batching 部分，你会发现它print(sess.run(next_element))运行了 3 次，但next_element只声明了一次。dataset.make_one_shot_iterator()同样，在您的代码中，无需iterator.get_next()在 for 循环下运行。数据集声明可以放在开头或内部getData()，以便于理解。

  3. 可以使用以下方法形成数据批处理：

    # read file
dataset = tf.data.TFRecordDataset(filename)
# parse each instance
dataset = dataset.map(your_parser_fun, num_parallel_calls=num_threads)
# preprocessing, e.g. scale to range [0, 1]
dataset = dataset.map(some_preprocessing_fun)
# shuffle
dataset = dataset.shuffle(buffer_size)
# form batch and epoch
dataset = dataset.batch(batch_size)
dataset = dataset.repeat(num_epoch)
iterator = dataset.make_one_shot_iterator()
# get a batch
x_batch, y_batch = self.iterator.get_next()

# do calculations
...

  4. 检查处理多个纪元部分以查看使用 for 循环设置纪元的示例。

于 2017-12-23T05:08:17.260 回答