【tensorflow】‘BatchDataset‘ object has no attribute ‘make_one_shot_iterator‘

过程

尝试在tf.sess中使用dataset。
首先直接用的以下方法制作dataset：

    dataset = tf.data.Dataset.from_tensor_slices((imgs, lables))
    dataset = dataset.map(read_img_func).batch(batch_size)

运行时用的：

for img, label in input_data:
	res = sess.run(self._out, feed_dict={self._input: img})

报错：

RuntimeError: `tf.data.Dataset` only supports Python-style iteration in eager mode or within tf.function.

随后换成以下代码：

iterators = input_data.make_one_shot_iterator()
img, label = iterators.get_next()

iterators = input_data.make_initializable_iterator()
img, label = iterators.get_next()

报错：

'BatchDataset' object has no attribute 'make_one_shot_iterator' 
'BatchDataset' object has no attribute 'make_initializable_iterator'

看了github上的源码dataset_op，似乎应该替换成tf.compat.v1.data.make_one_shot_iterator(dataset):

iterator = tf.compat.v1.data.make_initializable_iterator(input_data)
# iterators = input_data.make_one_shot_iterator()
img, label = iterator.get_next()
res_list = []
while True:
    try:
        import pdb
        pdb.set_trace()
        res = sess.run(self._out, feed_dict={self._input: img})
    except tf.errors.OutOfRangeError:
        print("iterator done")
        break

报错：

TypeError: The value of a feed cannot be a tf.Tensor object. Acceptable feed values include Python scalars, strings, lists, numpy ndarrays, or TensorHandles. For reference, the tensor object was Tensor("IteratorGetNext:0", shape=(None, 224, 224, 3), dtype=float32) which was passed to the argument `feed_dict` with key Tensor("Placeholder_137:0", shape=(None, 224, 224, 3), dtype=float32).

看样子在feed_dict里面传入的值加一个.eval(session=sess)就可以了。

res = sess.run(self._out, feed_dict={self._input: img.eval(session=sess)})

然后又报错：

GetNext() failed because the iterator has not been initialized. Ensure that you have run the initializer operation for this iterator before getting the next element.

加个这个就行：sess.run(iterator.initializer)

结论

with tf.device(‘/gpu:0’):
        iterator = tf.compat.v1.data.make_initializable_iterator(input_data)
        # iterators = input_data.make_one_shot_iterator()
        sess.run(iterator.initializer)
        img, label = iterator.get_next()
        res_list = []
        while True:
            try:
                res, _input = sess.run([self._out,self._input], feed_dict={self._input: img.eval(session=sess)})
                res_list.append(res)
            except tf.errors.OutOfRangeError:
                print("iterator done")
                break
        res = np.concatenate(res_list,axis=0)