'TypeError: expected bytes, str found' when creating images dataset with Dataset API

Question

我想使用数据集 API 从我的图像创建一个 TensorFlow 数据集。这些图像以复杂的层次结构组织，但最后总是有两个目录 "False" 和 "Genuine"。这段代码是我写的

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

def enumerate_all_files(rootdir):
    for subdir, dir, files in os.walk(rootdir):
        for file in files:
            # return path to the file and its label
            # label is simply a 1 or 0 depending on whether an image is in the "Genuine" folder or not
            yield os.path.join(subdir, file), int(subdir.split(os.path.sep)[-1] == "Genuine")

def input_parser(img_path, label):
    # convert the label to one-hot encoding
    one_hot = tf.one_hot(label, 2)
    # read the img from file
    img_file = tf.read_file(img_path)
    img_decoded = tf.image.decode_png(img_file, channels=3)
    return img_decoded, one_hot

def get_dataset():
    generator = lambda: enumerate_all_files("/tmp/images/training/")
    dataset = Dataset.from_generator(generator, (tf.string, tf.int32)).shuffle(1000).batch(100)
    dataset = dataset.map(input_parser)
    return dataset

然而，当我运行它在我的终端中

tf.enable_eager_execution()
# all the code above
d = get_dataset()
for f in d.make_one_shot_iterator():
    print(f)

它因错误而崩溃

W tensorflow/core/framework/op_kernel.cc:1306] Unknown: SystemError: <weakref at 0x7ff8232f0620; to 'function' at 0x7ff8233c9048 (generator_py_func)> returned a result with an error set
TypeError: expected bytes, str found  

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
  File "lcnn.py", line 29, in <module>
    for f in d.make_one_shot_iterator():
  File "/opt/anaconda3/envs/tensorflow/lib/python3.6/site-packages/tensorflow/python/data/ops/iterator_ops.py", line 487, in __next__
    return self.next()
  File "/opt/anaconda3/envs/tensorflow/lib/python3.6/site-packages/tensorflow/python/data/ops/iterator_ops.py", line 518, in next
    return self._next_internal()
  File "/opt/anaconda3/envs/tensorflow/lib/python3.6/site-packages/tensorflow/python/data/ops/iterator_ops.py", line 508, in _next_internal
    output_shapes=self._flat_output_shapes)
  File "/opt/anaconda3/envs/tensorflow/lib/python3.6/site-packages/tensorflow/python/ops/gen_dataset_ops.py", line 1848, in iterator_get_next_sync
    "output_types", output_types, "output_shapes", output_shapes)
SystemError: <built-in function TFE_Py_FastPathExecute> returned a result with an error set

我在这里做错了什么？

编辑
我在不调用 map、shuffle 和 batch 的情况下尝试运行ning 代码，也没有注释掉 input_parser 方法，但错误仍然出现。

编辑 2
我将 Dataset.from_generator 更改为 Dataset.from_tensor_slices 以查看打开图片的代码是否有效。所以更改后的代码看起来像

def input_parser(img_path):
    # convert the label to one-hot encoding
    # one_hot = tf.one_hot(label, 2)
    # read the img from file
    img_file = tf.read_file(img_path)
    img_decoded = tf.image.decode_png(img_file, channels=3)
    return img_decoded

def get_dataset():
    dataset = Dataset.from_tensor_slices(["/tmp/images/training/1000010.png"]).map(input_parser).shuffle(1000).batch(100)
    return dataset

虽然这很好用

Answer 1

我不确定错误消息，但我尝试了您的代码，如果我在批处理和洗牌操作之前执行输入解析器映射，它对我有用：

def get_dataset():
    generator = lambda: enumerate_all_files("/tmp/images/training/")
    dataset = Dataset.from_generator(generator, (tf.string,tf.int32)).map(input_parser)
    dataset = dataset.shuffle(1000).batch(100)
    return dataset

Answer 2

我最终将代码从 Dataset.from_generator 更改为 Dataset.from_tensor_slices。最终代码：

import tensorflow as tf
from tensorflow.data import Dataset
import os
tf.enable_eager_execution()

def enumerate_all_files(rootdir):
    for subdir, dir, files in os.walk(rootdir):
        for file in files:
            # return path to the file and its label
            # label is simply a 1 or 0 depending on whether an image is in the "Genuine" folder or not
            yield os.path.join(subdir, file), int(subdir.split(os.path.sep)[-1] == "Genuine")

def input_parser(img_path, label):
    # convert the label to one-hot encoding
    one_hot = tf.one_hot(label, 2)
    # read the img from file
    img_file = tf.read_file(img_path)
    img_decoded = tf.image.decode_png(img_file, channels=3)
    return img_decoded, one_hot

def get_dataset():

    file_paths = []
    labels = []

    for i in enumerate_all_files("/media/kuba/Seagate Expansion Drive/MGR/Spektrogramy/FFT/training/"):
        file_paths.append(i[0])
        labels.append(i[1])
    dataset = Dataset.from_tensor_slices((file_paths, labels)).map(input_parser).shuffle(1000).batch(100)
    return dataset

d = get_dataset()
for f in d.make_one_shot_iterator():
    print(type(f))

'TypeError: expected bytes, str found' when creating images dataset with Dataset API

'TypeError: expected bytes, str found' when creating images dataset with Dataset API

python

tensorflow

tensorflow-datasets