如何将 Tensorflow 数据集 API 与训练和验证集一起使用

Posted 2023-03-12

【中文标题】如何将 Tensorflow 数据集 API 与训练和验证集一起使用

【英文标题】：How to use Tensorflow dataset API with training and validation sets

【发布时间】：2018-05-01 14:11:44

【问题描述】：

手头的简单任务：运行 N 个 epoch 的训练，在每个 epoch 之后计算准确的验证准确度。时期大小可以等于完整的训练集或一些预定义的迭代次数。在验证期间，每个验证集输入都必须被评估一次。

将 one_shot_iterators、可初始化迭代器和/或用于该任务的句柄混合在一起的最佳方法是什么？

这是我认为它应该如何工作的脚手架：

def build_training_dataset():
    pass

def build_validation_dataset():
    pass

def construct_train_op(dataset):
    pass

def magic(iterator):
    pass

USE_CUSTOM_EPOCH_SIZE = True
CUSTOM_EPOCH_SIZE = 60
MAX_EPOCHS = 100


training_dataset = build_training_dataset()
validation_dataset = build_validation_dataset()


# Magic goes here to build a nice one-instance dataset
dataset = magic(training_dataset, validation_dataset)

train_op = construct_train_op(dataset)

# Run N epochs in which the training dataset is traversed, followed by the
# validation dataset.
with tf.Session() as sess:
    for epoch in MAX_EPOCHS:

        # train
        if USE_CUSTOM_EPOCH_SIZE:
            for _ in range(CUSTOM_EPOCH_SIZE):
                sess.run(train_op)
        else:
            while True:
                # I guess smth like this:
                try:
                    sess.run(train_op)
                except tf.errors.OutOfRangeError:
                    break # we are done with the epoch

        # validation
        validation_predictions = []
        while True:
            try:
                np.append(validation_predictions, sess.run(train_op)) # but for validation this time
            except tf.errors.OutOfRangeError:
                print('epoch %d finished with accuracy: %f' % (epoch validation_predictions.mean()))
                break 

【参考方案1】：

由于解决方案比我预期的要复杂得多，因此它分为 2 个和平：

0) 两个示例共享的辅助代码：

USE_CUSTOM_EPOCH_SIZE = True
CUSTOM_EPOCH_SIZE = 60
MAX_EPOCHS = 100

TRAIN_SIZE = 500
VALIDATION_SIZE = 145
BATCH_SIZE = 64


def construct_train_op(batch):
    return batch


def build_train_dataset():
    return tf.data.Dataset.range(TRAIN_SIZE) \
        .map(lambda x: x + tf.random_uniform([], -10, 10, tf.int64)) \
        .batch(BATCH_SIZE)

def build_test_dataset():
    return tf.data.Dataset.range(VALIDATION_SIZE) \
        .batch(BATCH_SIZE)

1) 对于等于训练数据集大小的 epoch：

# datasets construction
training_dataset = build_train_dataset()
validation_dataset = build_test_dataset()

# handle constructions. Handle allows us to feed data from different dataset by providing a parameter in feed_dict
handle = tf.placeholder(tf.string, shape=[])
iterator = tf.data.Iterator.from_string_handle(handle, training_dataset.output_types, training_dataset.output_shapes)
next_element = iterator.get_next()

train_op = construct_train_op(next_element)

training_iterator = training_dataset.make_initializable_iterator()
validation_iterator = validation_dataset.make_initializable_iterator()

with tf.Session() as sess:
    training_handle = sess.run(training_iterator.string_handle())
    validation_handle = sess.run(validation_iterator.string_handle())

    for epoch in range(MAX_EPOCHS):
        #train
        sess.run(training_iterator.initializer)
        total_in_train = 0
        while True:
            try:
                train_output = sess.run(train_op, feed_dict=handle: training_handle)
                total_in_train += len(train_output)
            except tf.errors.OutOfRangeError:
                assert total_in_train == TRAIN_SIZE
                break # we are done with the epoch

        # validation
        validation_predictions = []
        sess.run(validation_iterator.initializer)
        while True:
            try:
                pred = sess.run(train_op, feed_dict=handle: validation_handle)
                validation_predictions = np.append(validation_predictions, pred)
            except tf.errors.OutOfRangeError:
                assert len(validation_predictions) == VALIDATION_SIZE
                print('Epoch %d finished with accuracy: %f' % (epoch, np.mean(validation_predictions)))
                break

2) 对于自定义 epoch 大小：

# datasets construction
training_dataset = build_train_dataset().repeat() # CHANGE 1
validation_dataset = build_test_dataset()

# handle constructions. Handle allows us to feed data from different dataset by providing a parameter in feed_dict
handle = tf.placeholder(tf.string, shape=[])
iterator = tf.data.Iterator.from_string_handle(handle, training_dataset.output_types, training_dataset.output_shapes)
next_element = iterator.get_next()


train_op = construct_train_op(next_element)

training_iterator = training_dataset.make_one_shot_iterator() # CHANGE 2
validation_iterator = validation_dataset.make_initializable_iterator()

with tf.Session() as sess:
    training_handle = sess.run(training_iterator.string_handle())
    validation_handle = sess.run(validation_iterator.string_handle())

    for epoch in range(MAX_EPOCHS):
        #train
        # CHANGE 3: no initiazation, not try/catch
        for _ in range(CUSTOM_EPOCH_SIZE): 
            train_output = sess.run(train_op, feed_dict=handle: training_handle)


        # validation
        validation_predictions = []
        sess.run(validation_iterator.initializer)
        while True:
            try:
                pred = sess.run(train_op, feed_dict=handle: validation_handle)
                validation_predictions = np.append(validation_predictions, pred)
            except tf.errors.OutOfRangeError:
                assert len(validation_predictions) == VALIDATION_SIZE
                print('Epoch %d finished with accuracy: %f' % (epoch, np.mean(validation_predictions)))
                break

【讨论】：

您确定需要在每个 epoch 重新初始化训练迭代器（方法 1）吗？