def test_lasagne_model(num_classes):
bounds = (0, 255)
channels = num_classes
def mean_brightness_net(images):
logits = GlobalPoolLayer(images)
return logits
images_var = T.tensor4('images', dtype='float32')
images = InputLayer((None, channels, 5, 5), images_var)
logits = mean_brightness_net(images)
model = LasagneModel(images, logits, bounds=bounds)
test_images = np.random.rand(2, channels, 5, 5).astype(np.float32)
test_label = 7
assert model.forward(test_images).shape \
== (2, num_classes)
test_logits = model.forward_one(test_images[0])
assert test_logits.shape == (num_classes, )
test_gradient = model.gradient_one(test_images[0], test_label)
assert test_gradient.shape == test_images[0].shape
np.testing.assert_almost_equal(
model.forward_and_gradient_one(test_images[0], test_label)[0],
test_logits)
np.testing.assert_almost_equal(
model.forward_and_gradient_one(test_images[0], test_label)[1],
test_gradient)
assert model.num_classes() == num_classes
def test_lasagne_gradient(num_classes):
bounds = (0, 255)
channels = num_classes
def mean_brightness_net(images):
logits = GlobalPoolLayer(images)
return logits
images_var = T.tensor4('images', dtype='float32')
images = InputLayer((None, channels, 5, 5), images_var)
logits = mean_brightness_net(images)
preprocessing = (np.arange(num_classes)[:, None, None],
np.random.uniform(size=(channels, 5, 5)) + 1)
model = LasagneModel(images,
logits,
preprocessing=preprocessing,
bounds=bounds)
# theano and lasagne calculate the cross-entropy from the probbilities
# rather than combining softmax and cross-entropy calculation; they
# therefore have lower numerical accuracy
epsilon = 1e-3
np.random.seed(23)
test_image = np.random.rand(channels, 5, 5).astype(np.float32)
test_label = 7
_, g1 = model.forward_and_gradient_one(test_image, test_label)
l1 = model._loss_fn(test_image[None] - epsilon / 2 * g1, [test_label])
l2 = model._loss_fn(test_image[None] + epsilon / 2 * g1, [test_label])
assert 1e5 * (l2 - l1) > 1
# make sure that gradient is numerically correct
np.testing.assert_array_almost_equal(1e5 * (l2 - l1),
1e5 * epsilon * np.linalg.norm(g1)**2,
decimal=1)