def _build_module(self, input_layer):
# This is almost exactly the same as Dueling Network but we predict the future measurements for each action
multistep_measurements_size = self.measurements_size[0] * self.num_predicted_steps_ahead
# actions expectation tower (expectation stream) - E
with name_scope("expectation_stream"):
expectation_stream = neon.Sequential([
neon.Affine(nout=256, activation=neon.Rectlin(),
weight_init=self.weights_init, bias_init=self.biases_init),
neon.Affine(nout=multistep_measurements_size,
weight_init=self.weights_init, bias_init=self.biases_init)
])(input_layer)
# action fine differences tower (action stream) - A
with name_scope("action_stream"):
action_stream_unnormalized = neon.Sequential([
neon.Affine(nout=256, activation=neon.Rectlin(),
weight_init=self.weights_init, bias_init=self.biases_init),
neon.Affine(nout=self.num_actions * multistep_measurements_size,
weight_init=self.weights_init, bias_init=self.biases_init),
neon.Reshape((self.num_actions, multistep_measurements_size))
])(input_layer)
action_stream = action_stream_unnormalized - ng.mean(action_stream_unnormalized)
repeated_expectation_stream = ng.slice_along_axis(expectation_stream, expectation_stream.axes[0], 0)
repeated_expectation_stream = ng.expand_dims(repeated_expectation_stream, output_axis, 0)
# merge to future measurements predictions
self.output = repeated_expectation_stream + action_stream
def test_nested_namescope():
"""
Ops under multiple namescopes should have nested names. Namescope names don't nest
so that they can be reused in different nestings.
"""
with name_scope("scope1") as scope1:
assert scope1.name == "scope1"
val1 = ScopedNameableValue("val1")
assert val1.name == "scope1/val1"
with name_scope("scope2") as scope2:
assert scope2.name == "scope2"
val2 = ScopedNameableValue("val2")
assert val2.name == "scope1/scope2/val2"
def scope_ops(name=None, mode=None, subgraph=None, metadata=None):
"""
All ops created within the context manager will be added to a subgraph
Arguments:
name (str): variable scope to use for all created ops
mode (str): mode (e.g. "inference", "training") to annotate on all created ops
subgraph (SubGraph): subgraph instance to add ops to. If not provided, one will be created
metadata (dict): a dictionary of metadata to add to all created ops
Yields:
instance of SubGraph
"""
if subgraph is None:
subgraph = SubGraph()
if metadata is None:
metadata = dict()
if mode is not None:
metadata["mode"] = mode
with name_scope(name=name, reuse_scope=True):
with ng.Op.all_ops() as ops:
with ng.metadata(**metadata):
yield (subgraph)
subgraph.ops.extend(ops)
def test_scope_reuse():
"""
Namescopes are only reused if explicitly requested. Otherwise, a uniquely named will
be created.
"""
with name_scope("scope") as scope1:
val1 = ScopedNameableValue("val1")
with name_scope("scope", reuse_scope=True) as scope2:
val2 = ScopedNameableValue("val2")
with name_scope("scope", reuse_scope=False) as scope3:
val3 = ScopedNameableValue("val3")
assert scope1 is scope2
assert scope1 is not scope3
assert val1.name == "scope/val1"
assert val2.name == "scope/val2"
assert val3.name != "scope/val3"
def __call__(self, prev_input_placeholder=None):
with name_scope(self.get_name()):
# create the input axes
axes = []
if len(self.input_size) == 2:
axis_names = ['H', 'W']
else:
axis_names = ['C', 'H', 'W']
for axis_size, axis_name in zip(self.input_size, axis_names):
axes.append(ng.make_axis(axis_size, name=axis_name))
batch_axis_full = ng.make_axis(self.batch_size, name='N')
input_axes = ng.make_axes(axes)
if prev_input_placeholder is None:
self.input = ng.placeholder(input_axes + [batch_axis_full])
else:
self.input = prev_input_placeholder
self._build_module()
return self.input, self.output(self.input)
def __call__(self, input_layer):
"""
Wrapper for building the module graph including scoping and loss creation
:param input_layer: the input to the graph
:return: the output of the last layer and the target placeholder
"""
with name_scope(self.get_name()):
self._build_module(input_layer)
self.output = force_list(self.output)
self.target = force_list(self.target)
self.input = force_list(self.input)
self.loss_type = force_list(self.loss_type)
self.loss = force_list(self.loss)
self.regularizations = force_list(self.regularizations)
if self.is_local:
self.set_loss()
if self.is_local:
return self.output, self.target, self.input
else:
return self.output, self.input
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:
loss_d = loss_d + gp_scale * gradient_penalty
mean_cost_d = ng.mean(loss_d, out_axes=[])
mean_cost_g = ng.mean(loss_g, out_axes=[])
def __call__(self, input_layer):
with name_scope(self.get_name()):
self.input = input_layer
self._build_module()
return self.input, self.output(self.input)
def mean_squared_error(targets, outputs, weights=1.0, scope=""):
with name_scope(scope):
# TODO: reduce mean over the action axis
loss = ng.squared_L2(targets - outputs)
weighted_loss = loss * weights
return weighted_loss
