Question

我正在尝试从hdf5文件中排队数据。由于Tensorflow不支持hdf5，我创建了一个python函数，它从hdf5文件中读取示例，并在到达文件末尾时引发tf.errors.OutOfRangeError。然后我用tf.py_func包装这个python函数，并将其用作队列的入队操作。

这是我的代码：

import h5py
import tensorflow as tf
from tensorflow.python.framework import errors
import numpy as np

def read_from_hdf5(hdf5_file, batch_size):
    h5py_handle = h5py.File(hdf5_file)

    # Check shapes from the hdf5 file so that we can set the tensor shapes
    feature_shape = h5py_handle['features'].shape[1:]
    label_shape = h5py_handle['labels'].shape[1:]

    #generator that produces examples for training. It will be wrapped by tf.pyfunc to simulate a reader
    def example_generator(h5py_handle):
        for i in xrange(0, h5py_handle['features'].shape[0]-batch_size+1, batch_size):
            features = h5py_handle['features'][i:i+batch_size]
            labels = h5py_handle['labels'][i:i+batch_size]
            yield [features, labels]
        raise errors.OutOfRangeError(node_def=None, op=None, message='completed all examples in %s'%hdf5_file)

    [features_tensor, labels_tensor] = tf.py_func(
        example_generator(h5py_handle).next,
        [],
        [tf.float32, tf.float32],
        stateful=True)

    # Set the shape so that we can infer sizes etc in later layers.
    features_tensor.set_shape([batch_size, feature_shape[0], feature_shape[1], feature_shape[2]])
    labels_tensor.set_shape([batch_size, label_shape[0]])

    return features_tensor, labels_tensor


def load_data_from_filename_list(hdf5_files, batch_size, shuffle_seed=0):
    example_list = [read_from_hdf5(hdf5_file, batch_size) for hdf5_file in hdf5_files]
    min_after_dequeue = 10000
    capacity = min_after_dequeue + (len(example_list)+1) * batch_size #min_after_dequeue + (num_threads + a small safety margin) * batch_size
    features, labels = tf.train.shuffle_batch_join(example_list, batch_size, capacity=capacity, min_after_dequeue=min_after_dequeue, seed=shuffle_seed, enqueue_many=True)
    return features, labels, metadata

我预计tf.errors.OutOfRangeError将由QueueRunner处理，但是，我收到以下错误，程序崩溃。是否可以从py_func中进行这种读取，如果是这样，我做错了什么？如果没有，我应该采用什么方法呢？

Traceback (most recent call last):
  File "/users/anaconda2/lib/python2.7/site-packages/tensorflow/python/ops/script_ops.py", line 85, in __call__
    ret = func(*args)
  File "build/bdist.linux-x86_64/egg/tronn/datalayer.py", line 27, in example_generator
    raise errors.OutOfRangeError(node_def=None, op=None, message='completed all examples in %s'%hdf5_file)
tensorflow.python.framework.errors_impl.OutOfRangeError: completed all examples
W tensorflow/core/framework/op_kernel.cc:993] Internal: Failed to run py callback pyfunc_13: see error log.

Answer 1

看起来不支持py_func中的异常处理。

在py_func.cc

中考虑此代码

// Invokes the trampoline.
  PyObject* result = PyEval_CallObject(trampoline, args);
  Py_DECREF(args);
  if (result == nullptr) {
    if (PyErr_Occurred()) {
      // TODO(zhifengc): Consider pretty-print error using LOG(STDERR).
      PyErr_Print();
    }
    return errors::Internal("Failed to run py callback ", call->token,
                            ": see error log.");
  }

生成异常时会设置

PyErr_Occurred，因此会导致执行抛出Failed to run py callback。

py_func是一个奇怪的生物，因为它在你的Python客户端环境中运行。通常，当op（如reader op）失败时，从TF运行时传播的它会向Python客户端返回不正常状态，然后Python客户端将其转换为raise_exception_on_not_ok_status中的Python异常（在client.py:session.run中）。由于py_func正文在Python客户端中运行，因此需要修改TensorFlow以处理PyErr_Occurred以将错误状态插回到TensorFlow运行时。

Tensorflow：使用py_func的自定义数据读取器

1 个答案: