def __call__(self, in_obj, keep=None, **kwargs):
if self.mask is None:
in_axes = in_obj.axes.sample_axes()
self.mask = ng.persistent_tensor(axes=in_axes).named('mask')
self.mask = ng.less_equal(ng.uniform(self.mask, low=0.0, high=1.0),
keep)
return ng.multiply(self.mask, in_obj) * (1. / keep)
def __call__(self, in_obj):
if Layer.inference_mode:
return self.keep * in_obj
else:
if self.mask is None:
in_axes = in_obj.axes.sample_axes()
self.mask = ng.persistent_tensor(axes=in_axes).named('mask')
self.mask = ng.uniform(self.mask, low=0.0, high=1.0) <= self.keep
return self.mask * in_obj
def test_uniform_range_posneg(input_tensor):
"""TODO."""
ng_a = ng.uniform(input_tensor, low=-0.5, high=0.5)
with executor(ng_a) as ex:
result = ex()
print(result)
assert np.all(result < 0.5)
assert np.all(result >= -0.5)
assert not np.all(result >= 0.0)
def __call__(self, in_obj, **kwargs):
if Layer.inference_mode:
return self.keep * in_obj
else:
if self.mask is None:
in_axes = in_obj.axes.sample_axes()
channel_axes = ng.make_axes([in_axes.channel_axis()])
self.mask = ng.persistent_tensor(
axes=channel_axes).named('channel_mask')
self.mask = ng.uniform(self.mask, low=0.0, high=1.0) <= self.keep
return self.mask * in_obj
def test_uniform_range_pos(transformer_factory, input_tensor):
"""TODO."""
ng_a = ng.uniform(input_tensor, low=0.0, high=0.5)
with executor(ng_a) as ex:
result = ex()
print(result)
assert np.all(result < 0.5)
assert np.all(result >= 0.0)
assert not np.all(result == 0.0)
def test_rng_repetition():
"""
Tests rng ops, to make sure they run every execution and not just initialization
"""
axes = ng.make_axes([ng.make_axis(2), ng.make_axis(2)])
x = ng.variable(initial_value=np.array([[1, 2], [3, 4]]), axes=axes)
y = ng.uniform(x)
mysum = ng.sum(y)
trans = ng.transformers.make_transformer()
rand_comp = trans.computation(mysum)
val1 = rand_comp().copy()
val2 = rand_comp().copy()
assert val1 != val2
trans.close()
def test_uniform_range_posneg(transformer_factory):
"""TODO."""
M = ng.make_axis(5, name='M')
N = ng.make_axis(8, name='N')
ng_a = ng.persistent_tensor([M, N], initial_value=10.0)
ng_a = ng.uniform(ng_a, low=-0.5, high=0.5)
result = executor(ng_a)()
print(result)
assert np.all(result < 0.5)
assert np.all(result >= -0.5)
assert not np.all(result >= 0.0)
# build network graph
generated = generator(z)
D1 = discriminator(image)
D2 = discriminator(generated)
weight_clip_value = None # no weight clipping
gp_scale = args.gp_scale # gradient penalty coefficient
loss_d = D1 - D2
loss_g = D2
# calculate gradient for all losses
x = ng.variable(initial_value=0.5, axes=[])
epsilon = ng.uniform(x)
interpolated = epsilon * image + (1 - epsilon) * generated
D3 = discriminator(interpolated)
with name_scope(name="GradientPenalty"):
gradient = ng.deriv(ng.sum(D3, out_axes=[]), interpolated)
grad_norm = ng.L2_norm(gradient)
gradient_penalty = ng.square(grad_norm - 1)
# add gradient penalty
# TODO
# when gp_scale is set to 0 the behavior is not as expected
# loss_d = loss_d + 0 * gp is not loss_d = loss_d + 0
# we can get rid of if statement once this is fixed
# https://github.com/NervanaSystems/private-ngraph/issues/2145
if gp_scale:
def train_outputs(self, in_obj):
in_axes = in_obj.axes.sample_axes()
self.mask = self.mask or ng.persistent_tensor(
axes=in_axes).named('mask')
self.mask = ng.uniform(self.mask, low=0.0, high=1.0) <= self.keep
return self.mask * in_obj
