将队列分割为训练集和测试集

Question

我设置了一个流水线，从文件名队列开始，如以下伪代码所示：

filename_queue = tf.train.string_input_producer(["file0.pd", "file1.pd"])

指向包含多个序列化的tf.train.Example图像的TFRecords。按照TensorFlow指南，以下是读取一个示例的函数：

def read_my_file_format(filename_queue):
  reader = tf.SomeReader()
  key, record_string = reader.read(filename_queue)
  example, label = tf.some_decoder(record_string)
  processed_example = some_processing(example)
  return processed_example, label

这是用于批处理队列的：

def input_pipeline(filenames, batch_size):
  filename_queue = tf.train.string_input_producer(filenames)
  example, label = read_my_file_format(filename_queue)

  example_batch, label_batch = tf.train.shuffle_batch(
      [example, label], batch_size=batch_size, capacity=100,
      min_after_dequeue=10)
  return example_batch, label_batch

我正在寻找一种将数据随机分为训练集和测试集的方法。我不想将训练集和测试集保存到不同的文件中，而是希望独立于读取文件的方式将图像随机分配到训练或测试集中。理想情况下，我希望将输入流水线分成一个训练队列和一个测试队列。

以下是我通常在numpy中执行巨大数据集拆分时所做的内容。

import numpy as np
from numpy.random import choice
from numpy.random import RandomState

queue = range(10)
weights = (.8,.2) # create 2 partitions with this weights

def sampler(partition, seed=0):
    rng = RandomState(seed)
    return lambda x: rng.choice(np.arange(len(weights)), p=weights) == partition

def split(queue, weights):
    # filter the queue for each partition
    return [filter(sampler(partition), queue) for partition in range(len(weights)) ]

(train, test) = split(queue, weights)               


print(list(train)) # [0, 1, 2, 3, 4, 5, 6, 9]
print(list(test))  # [7, 8]

- Manuel Schmidt

---

你会多次处理每个文件吗？如果是，将它们随机分成训练集和测试集可能会有问题。 - Mad Wombat

是的，因为我从每张图片中取样一个随机补丁， - Manuel Schmidt

在这种情况下，您需要想办法记住将文件的哪个部分指定为测试，哪个部分指定为训练。最好是从一开始就这样做，甚至在读取文件之前就这样做。浏览文件名列表，读取元数据（如果您的图像大小/形状不同），并随机决定哪些位将用于测试，哪些位将用于训练集。将此信息与文件名一起存储在字典或其他东西中，然后将该字典馈送到队列中，而不仅仅是文件名。 - Mad Wombat

使用sklearn.model_selection.train_test_split()函数。 - igrinis

根据我的经验，将数据在运行时切分成训练集和测试集总是一个糟糕的想法。 - P-Gn

Answer 1

建议使用Tensorflow数据集API（map(), interleave(), filter()）：

import tensorflow as tf
import numpy as np

def _parse_function(example_proto):
    """ Parse TFRecord data """
    features = {"image": tf.FixedLenFeature((), tf.string, default_value=""),
              "label": tf.FixedLenFeature((), tf.int64, default_value=0)}
    parsed_features = tf.parse_single_example(example_proto, features)
    return parsed_features

def split_train_test(parsed_features, train_rate=0.8, seed=11):
    """ Randomly classify samples into training or testing split """
    # Snippet by Igor Gadelha Pereira (https://dev59.com/V1gR5IYBdhLWcg3wBZW3#49825457)
    parsed_features['is_train'] = tf.gather(tf.random_uniform([1], seed=seed) < train_rate, 0)
    return parsed_features

def filter_per_split(parsed_features, train=True):
    """ Filter samples depending on their split """
    return parsed_features['is_train'] if train else ~parsed_features['is_train']

def select_features(parsed_features, keys=["image", "label"]):
    """ Return array of features selected by key """
    selected_features = [parsed_features[key] for key in keys]
    return selected_features

weights = (.8,.2)
num_files = 3
file_block_length = 1
files = ["/tmp/file{}.tfrecords".format(i) for i in range(num_files)]
# ... where file{i}.tfrecords contains:
# [{"label": i, "image": "class_{}/img_{}.png".format(i, k)} for k in range(10)]

# Create TFRecord file list list:
files = tf.data.Dataset.from_tensor_slices(files)
# Interleave all records:
dataset = files.interleave(lambda x: tf.data.TFRecordDataset(x),
                           cycle_length=num_files, block_length=file_block_length)
# ^ dataset containing:
# [rec0@file0, rec0@file1, rec0@file2, rec1@file0, rec1@file1, rec1@file2, ...]

# Parse TFRecord samples:
dataset = dataset.map(_parse_function)

# Randomly classify samples between training or testing:
dataset = dataset.map(lambda x: split_train_test(x, train_rate=weights[0]))

# Split into 2 datasets accordingly:
dataset_train = dataset.filter(lambda x: filter_per_split(x, train=True))
dataset_test = dataset.filter(lambda x: filter_per_split(x, train=False))

# Opt. remove "is_train" key, keeping only the original features:
dataset_train = dataset_train.map(select_features)
dataset_test = dataset_test.map(select_features)

# Use:
iterator_train = dataset_train.make_one_shot_iterator()
iterator_test = dataset_test.make_one_shot_iterator()
with tf.Session() as sess:
    for it, name in zip([iterator_train, iterator_test], ["Training", "Testing"]):
        x = it.get_next()
        count = 0
        print("{} Split:".format(name))
        try:
            while True:
                print(sess.run(x))
                count += 1
        except:
            print("- End of Split ({} / {}".format(count, num_files * 10))

输出：

Training Split:
(b'class_0/img_0.png', 0)
(b'class_1/img_0.png', 1)
(b'class_2/img_0.png', 2)
(b'class_0/img_1.png', 0)
(b'class_1/img_1.png', 1)
(b'class_1/img_2.png', 1)
(b'class_2/img_2.png', 2)
(b'class_0/img_3.png', 0)
(b'class_1/img_3.png', 1)
(b'class_2/img_3.png', 2)
(b'class_1/img_4.png', 1)
(b'class_2/img_4.png', 2)
(b'class_0/img_5.png', 0)
(b'class_1/img_5.png', 1)
(b'class_2/img_5.png', 2)
(b'class_0/img_6.png', 0)
(b'class_1/img_6.png', 1)
(b'class_2/img_6.png', 2)
(b'class_0/img_7.png', 0)
(b'class_1/img_7.png', 1)
(b'class_2/img_7.png', 2)
(b'class_0/img_8.png', 0)
(b'class_1/img_8.png', 1)
(b'class_2/img_8.png', 2)
(b'class_0/img_9.png', 0)
(b'class_1/img_9.png', 1)
(b'class_2/img_9.png', 2)
- End of Split (27 / 30
Testing Split:
(b'class_2/img_1.png', 2)
(b'class_0/img_2.png', 0)
(b'class_0/img_4.png', 0)
- End of Split (3 / 30

- benjaminplanche