def test_shuffle_tensor():
# Make numerics
numerical_tensor = np.random.uniform(size=(2, 5, 5, 2))
# Case 1: Known ndim, axis = -1
possible_numerical_outputs = [
numerical_tensor, numerical_tensor[..., ::-1]
]
# Make tensor and op
tensor = A.placeholder(shape=[None] * 4)
shuffled_tensor = A.shuffle_tensor(tensor, axis=-1)
# Evaluate
output = A.run(shuffled_tensor, {tensor: numerical_tensor})
# Check if output checks out
assert any([
np.allclose(numerical_output, output)
for numerical_output in possible_numerical_outputs
])
# Case 2: unknown ndim, axis = 0
possible_numerical_outputs = [
numerical_tensor, numerical_tensor[::-1, ...]
]
# Make tensor and op
tensor = A.placeholder()
shuffled_tensor = A.shuffle_tensor(tensor)
# Evaluate
output = A.run(shuffled_tensor, {tensor: numerical_tensor})
# Check if output checks out
assert any([
np.allclose(numerical_output, output)
for numerical_output in possible_numerical_outputs
])
# Check whether the gradients can be computed
tensor = A.placeholder()
shuffled_tensor = A.shuffle_tensor(tensor, differentiable=True)
grad_tensor = A.gradients(objective=A.reduce_(shuffled_tensor, 'mean'),
with_respect_to=tensor)
numerical_grad_tensor = A.run(grad_tensor, {tensor: numerical_tensor})
assert numerical_grad_tensor.values.shape == numerical_tensor.shape
# Check whether the lookup error is raised otherwise
with pytest.raises(LookupError):
tensor = A.placeholder()
shuffled_tensor = A.shuffle_tensor(tensor, differentiable=False)
grad_tensor = A.gradients(objective=A.reduce_(shuffled_tensor, 'mean'),
with_respect_to=tensor)
# Case 3: Unknown ndim, axis = -1
with pytest.raises(AssertionError):
# Make tensor and op
tensor = A.placeholder()
shuffled_tensor = A.shuffle_tensor(tensor, axis=-1)
# Case 4: Known dim, axis >= ndim
with pytest.raises(AssertionError):
# Make tensor and op
tensor = A.placeholder(shape=[None] * 4)
shuffled_tensor = A.shuffle_tensor(tensor, axis=4)
def test_tversky_distance():
_y = np.zeros(shape=(5, ))
_yt = np.zeros(shape=(5, ))
_y[0:2] = 1.
_yt[0:2] = 1.
y = A.placeholder()
yt = A.placeholder()
d = A.tversky_distance(y, yt, with_logits=False)
_d = A.run(d, {y: _y, yt: _yt})
assert _d == 0.
def test_normalize():
# Case 1: No mean and average known in advance
# (other test cases should be covered in tensorflow)
tensor = A.placeholder(shape=[None] * 2)
normalized_tensor = A.normalize(tensor)
rng = np.random.RandomState(42)
numerical_tensor = rng.uniform(size=(10, 10), low=-1., high=5.)
numerical_normalized_tensor = A.run(normalized_tensor,
{tensor: numerical_tensor})
assert np.allclose(np.mean(numerical_normalized_tensor), 0., atol=1e-7)
assert np.allclose(np.std(numerical_normalized_tensor), 1., atol=1e-3)
def test_feeder_runner():
# ---- BASIC TEST
# Define a random generator to mimic mnist
def random_generator_mnist(batch_size=10):
yield np.random.uniform(size=(batch_size, 28, 28, 1)).astype('float32'), \
np.random.uniform(size=(batch_size, 1, 1, 1)).astype('float32')
# Build preprocessing function
def preprocessor(inputs):
time.sleep(5)
return inputs
# Build feeder runner with multiple threads. Observe that the batch_size of the generator need
# not be that of the feeder_runner
feeder_runner = runners.FeederRunner(feeder=random_generator_mnist(),
preprocessor=preprocessor,
dtypes=['float32', 'float32'],
input_shape=[[None, 28, 28, 1],
[None, 1, 1, 1]],
num_threads=2,
batch_size=5)
images, labels = feeder_runner.dq()
# We need to make a session right now (in this thread) because tensorflow would set a different
# default session for every thread
# sess = A.Session.session
# Start generator threads
feeder_runner.start_runner()
# Dequeue
np_images, np_labels = A.run([images, labels])
# Stop generator threads
feeder_runner.stop_runner()
# Check shapes
assert np_images.shape == (5, 28, 28, 1)
assert np_labels.shape == (5, 1, 1, 1)
