C3w3-lab2: python and tuple of tensors, how..?

Course Q&A Machine Learning Engineering for Production(MLOps) Machine Learning Modeling Pipelines in Production
1

Am trying & failing to fully understand how the following code works.

def format_image(image, label):

This simple function is defined as taking two arguments (and running it, shows that it receives tensorflow.python.framework.ops.Tensor tensors).

But the following line seems to pass train_examples, a tf.python.data.ops.dataset_ops.PrefetchDataset to format_image() (via the map() function) like so:

train_batches = train_examples.shuffle(num_examples // 4).map(format_image).batch(BATCH_SIZE).prefetch(1)

When I do:

for element in train_examples.as_numpy_iterator():
	print(type(element))
	break

I get <class 'tuple'>

I (very) tentatively concluded that the map() line is passing a tuple of tensors to the function (somehow …wrapped inside a tf.data.Dataset … ?), and not as two separate tensors.

But if this is the case why does Python allow this? The function is expecting two arguments, not a single tuple.

So, if I do (while commenting out the tf.image.resize() line):

format_image('confused', 1)

I get (as expected) no errors.

But if I do:

my_tuple = ('confused', 1)
format_image(my_tuple)

I get (as expected):

TypeError: format_image() missing 1 required positional argument: 'label'

So, why does train_examples.map(format_image) not also throw this TypeError ?

(Without being able to get such info from the documentation or even from print statements, I think these powerful libraries of Tensorflow are less accessible to newcomers, … I mean, beyond copy-pasting this stuff and seeing that it just works, without really knowing how, will probably cause problems somewhere down the line, imo).

2

Hello Shahin,

In the beginning I also had many issues understanding the datastructures used in Tensorflow. It helps to take some time going through the Tensorflow documentation and playing with the notebooks. A great place to start for Datasets is the tf.data guide:

tf.data: Build TensorFlow input pipelines | TensorFlow Core

The following paragraph from the guide may help explaining your question above:
The tf.dataAPI introduces a tf.data.Datasetabstraction that represents a sequence of elements, in which each element consists of one or more components. For example, in an image pipeline, an element might be a single training example, with a pair of tensor components representing the image and its label.

There is another helpful link for understanding TFRecord files and how to link to the underlying raw data inside of it:

Good luck and happy learning!

Maarten

3 Likes