TypeError:使用带有估计器input_fn的数据集不支持可调用

Question

我正在尝试转换Iris教程(https://www.tensorflow.org/get_started/estimator)以从.png文件而不是.csv中读取训练数据.它可以使用,numpy_input_fn但不是我从它做的时候Dataset.我认为input_fn()是返回了错误的类型,但并不真正理解它应该是什么以及如何实现它.错误是:

  File "iris_minimal.py", line 27, in <module>
    model_fn().train(input_fn(), steps=1)
    ...
    raise TypeError('unsupported callable') from ex
TypeError: unsupported callable

TensorFlow版本是1.3.完整代码:

import tensorflow as tf
from tensorflow.contrib.data import Dataset, Iterator

NUM_CLASSES = 3

def model_fn():
    feature_columns = [tf.feature_column.numeric_column("x", shape=[4])]
    return tf.estimator.DNNClassifier([10, 20, 10], feature_columns, "tmp/iris_model", NUM_CLASSES)

def input_parser(img_path, label):
    one_hot = tf.one_hot(label, NUM_CLASSES)
    file_contents = tf.read_file(img_path)
    image_decoded = tf.image.decode_png(file_contents, channels=1)
    image_decoded = tf.image.resize_images(image_decoded, [2, 2])
    image_decoded = tf.reshape(image_decoded, [4])
    return image_decoded, one_hot      

def input_fn():
    filenames = tf.constant(['images/image_1.png', 'images/image_2.png'])
    labels = tf.constant([0,1])
    data = Dataset.from_tensor_slices((filenames, labels))
    data = data.map(input_parser)
    iterator = data.make_one_shot_iterator()
    features, labels = iterator.get_next()
    return features, labels

model_fn().train(input_fn(), steps=1)

Answer 1

我注意到你的代码片段中有几个错误:

• train方法接受输入函数,所以它应该是input_fn,而不是input_fn().
• 这些特征被期望作为字典,例如{'x': features}.
• DNNClassifier使用SparseSoftmaxCrossEntropyWithLogits损失函数.稀疏意味着它假定序数类表示,而不是单热表示,因此您的转换是不必要的(此问题解释了tf中交叉熵损失之间的差异).

请尝试以下代码:

import tensorflow as tf
from tensorflow.contrib.data import Dataset

NUM_CLASSES = 3

def model_fn():
    feature_columns = [tf.feature_column.numeric_column("x", shape=[4], dtype=tf.float32)]
    return tf.estimator.DNNClassifier([10, 20, 10], feature_columns, "tmp/iris_model", NUM_CLASSES)

def input_parser(img_path, label):
    file_contents = tf.read_file(img_path)
    image_decoded = tf.image.decode_png(file_contents, channels=1)
    image_decoded = tf.image.resize_images(image_decoded, [2, 2])
    image_decoded = tf.reshape(image_decoded, [4])
    label = tf.reshape(label, [1])
    return image_decoded, label

def input_fn():
    filenames = tf.constant(['input1.jpg', 'input2.jpg'])
    labels = tf.constant([0,1], dtype=tf.int32)
    data = Dataset.from_tensor_slices((filenames, labels))
    data = data.map(input_parser)
    data = data.batch(1)
    iterator = data.make_one_shot_iterator()
    features, labels = iterator.get_next()
    return {'x': features}, labels

model_fn().train(input_fn, steps=1)