Shape equals () - tensorflow

Question

I am reading the tests in the TensorFlow MNIST official model. Line 49 has:
self.assertEqual(loss.shape, ())
and selected lines leading up to it are:
BATCH_SIZE = 100
def dummy_input_fn():
image = tf.random_uniform([BATCH_SIZE, 784])
labels = tf.random_uniform([BATCH_SIZE, 1], maxval=9, dtype=tf.int32)
return image, labels
def make_estimator():
return tf.estimator.Estimator(
model_fn=mnist.model_fn, params={
'data_format': 'channels_last'
})
class Tests(tf.test.TestCase):
"""Run tests for MNIST model."""
def test_mnist(self):
classifier = make_estimator()
classifier.train(input_fn=dummy_input_fn, steps=2)
loss = eval_results['loss']
self.assertEqual(loss.shape, ())
but the TensorFlow documentation suggests that a shape is an array of numbers:
t = tf.constant([[[1, 1, 1], [2, 2, 2]], [[3, 3, 3], [4, 4, 4]]])
tf.shape(t) # [2, 2, 3]
These two statements that print the shape of the object don't help much:
print(loss.shape)
# prints `()`
print(tf.shape(loss))
# prints `Tensor("Shape:0", shape=(0,), dtype=int32)`
What is the meaning of a () shape?

Answer 1

Your loss is a NumPy object and not a TensorFlow object:
print(type(loss))
# prints <class 'numpy.float32'>
print(loss)
# prints 2.2745261
I assume that a shape of () in NumPy means a scalar, though I could not find a documentation for it. You can see the list of object attributes (fields and methods) with:
print(dir(loss))
# prints `['T', '__abs__', '__add__', '__and__',
# ... 'shape', 'size', 'sort', ... ]`