Access one image of dataset obtained with tfds.load()-CodePudding

I download my dataset like this: dataset = tfds.load('cifar10', split='train', shuffle_files=True). After applying the batch() method I have that dataset is:

Attributes of object: {'image': <tf.Tensor 'IteratorGetNext:0' shape=(128, 32, 32, 3) dtype=uint8>, 'label': <tf.Tensor 'IteratorGetNext:1' shape=(128,) dtype=int64>}

If I use next(iter(train_dataset)) or for elem in train_dataset.take(1): print (elem.numpy()) I need eager_execution.

If I try dataset[image], I get that the object is not subscriptable.

If I try with:

iterator=dataset.make_one_shot_iterator()
  next_val = iterator.get_next()
  print(f'my_image: {next_val}')
  print(f'my_image: {next_val['image']}')

I get:

    print(f'my_image: {next_val['image']}')
                                 ^
SyntaxError: invalid syntax

Even if I enable eager execution and do batch = iter(dataset).get_next() I get

Attributes of object: {'image': <tf.Tensor: id=156, shape=(128, 32, 32, 3), dtype=uint8, numpy=
array([[[[ 53,  53,  60],
         [ 63,  62,  66],
         [ 74,  72,  75],
         ...,
         [187, 194, 232],

but I can't access the single image.

EDIT: I'm using tensorflow 1.15.5

CodePudding user response：

You can access the batched array with:

dataset.x

where it will have shape (num_of_images, rows, cols, channel).

If you want a single image, you can treat it as an array (because it is):

first_image = dataset.x[0]
last_image = dataset.x[-1]

pyplot.plt.axis("off")
pyplot.imshow(first_image)

Edit: Additionally, you can get the corresponding array of labels with:

dataset.y

CodePudding user response：

I found the solution (simpler than I thought):

for element in dataset:
  single_element = element
  break

image = single_element['image']
label = single_element['label']