Boolean expression in Tensorflow CIFAR-10 tutorial

Question

I'm reading the codes about a TF official example about cifar10 on https://github.com/tensorflow/models/blob/master/official/resnet/cifar10_main.py and I have some questions:

In the function input_fn, what does

num_images = is_training and _NUM_IMAGES['train'] or _NUM_IMAGES['validation']

... mean? How can we get the right size of data while training and validating through this function?

In the function main, there's a similar one

input_function = FLAGS.use_synthetic_data and get_synth_input_fn() or input_fn

Again, I don't know how it works.

score 1 · Accepted Answer · answered Mar 22 '18 at 13:31

num_images = is_training and _NUM_IMAGES['train'] or _NUM_IMAGES['validation']

is equivalent to

if is_training:
    num_images = _NUM_IMAGES['train']
else:
    num_images = _NUM_IMAGES['validation']

In the same vein:

input_function = FLAGS.use_synthetic_data and get_synth_input_fn() or input_fn

is equivalent to:

if FLAGS.use_synthetic_data:
   input_function = get_synth_input_fn()
else:
    input_function = input_fn()

While my given more verbose variants may be more readable, the original tensorflow version is more compact.

The and operator short circuits, e.g in

(A and B)

B is only evaluated if A is true.

This means that in:

A and B or C

If A is true, then B is evaluated and or never gets to evaluate C, so the result is B. If A is false, then B is never evaluated and the result is C.

For more information study the docs:

https://docs.python.org/2/library/stdtypes.html#boolean-operations-and-or-not

@KM/ if you like my answer, it would be nice if you validated it by accepting it :) I would appreciate that a lot. If something is missing: feel free to point it out. — MFreidank, Mar 26 '18 at 17:22

score 0 · Answer 2 · answered Mar 22 '18 at 13:29

This is a clever application of boolean operations with integers (or actually any objects) in python (see this question for more details). Example:

>>> True and 10 or 20
Out[11]: 10
>>> False and 10 or 20
Out[12]: 20

>>> a = False and (lambda x: x) or (lambda x: 2*x)
>>> a(1)
Out[14]: 2

So the result of num_images is an integer (first or second one depending on is_training), the result of input_function is a function (again, the first or the second one depending on the flag use_synthetic_data).

score 0 · Answer 3 · answered Mar 22 '18 at 13:31

You are the victim of the bad code style of TensorFlow in this case. The tutorial is written by using a particular Python anti-trick in which you use and to select the final object of two objects that evaluate to True in a Boolean context, and you use or to select the final object in the case when the first object evaluates to False.

You can see it more easily with some simpler examples:

In [9]: 3 and "hello"
Out[9]: 'hello'

In [10]: 3 and False or "foobar"
Out[10]: 'foobar'

So these lines are selecting the necessary function or data size by chaining together these two anti-tricks.

This makes for needless obfuscated and unreadable code, and the minute someone defends it as OK just because the language supports it or "it's Pythonic" you can mostly just stop trusting their advice.

Boolean expression in Tensorflow CIFAR-10 tutorial

3 Answers3