def pipeline(self, default=steps.standard, config=None):
if self.options['--scalar']:
resources = scalar_pipeline.resources
else:
resources = standard_pipeline.resources
all_steps = self.options['pipeline']
pos = [p for p in all_steps if not isinstance(p, Not)]
neg = {p.value for p in all_steps if isinstance(p, Not)}
if not pos:
pos = default
final = [p for p in pos if p not in neg]
pdef = PipelineDefinition(resources=resources,
steps={p._name: p
for p in final})
opts = self.options['opts']
if opts and steps.opt not in final:
raise Exception('Optimizations can only be applied if the'
' opt step is in the pipeline')
elif opts:
pdef = pdef.configure({'opt.opts': Merge(opts)})
if callable(config):
pdef = config(pdef)
elif config:
pdef = pdef.configure(config)
return pdef.make()
def op_pipeline():
return PipelineDefinition(
addp=OpStep.partial(op=lambda p, x: p + x, param=1),
mulp=OpStep.partial(op=lambda p, x: p * x, param=2),
neg=OpStep.partial(op=lambda p, x: -x),
square=OpStep.partial(op=lambda p, x: x * x),
)
def op_pipeline():
return PipelineDefinition(
resources=dict(param=2),
steps=dict(
addp=OpStep.partial(op=lambda p, x: p + x, param=1),
mulp=OpResourceStep.partial(op=lambda p, x: p * x),
neg=OpStep.partial(op=lambda p, x: -x),
square=OpStep.partial(op=lambda p, x: x * x),
))
def grad_wrap(graph, argspec):
mg = GradOperation(graph,
wrt=['*'],
dout_parameter=True,
always_return_tuple=True)
sig = mg.make_signature(argspec)
g = mg.generate_graph(sig)
return {'graph': g}
grad_pipeline = PipelineDefinition(
resources=standard_resources,
parse=steps.step_parse,
resolve=steps.step_resolve,
infer=steps.step_infer,
specialize=steps.step_specialize,
opt=steps.step_debug_opt,
validate=steps.step_validate,
export=steps.step_debug_export,
)
def test_GradTester():
def f(x, y):
return x / y
def df(x, y, dz):
return dz / y, -dz * x / (y * y)
arglist = (
(7.3, 4.2),
mg = GradOperation(graph,
wrt=['*'],
dout_parameter=True,
always_return_tuple=True)
sig = mg.make_signature(argspec)
g = mg.generate_graph(sig)
return {'graph': g}
grad_pipeline = PipelineDefinition(
resources=standard_resources,
parse=steps.step_parse,
resolve=steps.step_resolve,
infer=steps.step_infer,
specialize=steps.step_specialize,
opt=steps.step_debug_opt,
validate=steps.step_validate,
export=steps.step_debug_export,
).configure({
'resources.operation_whitelist': grad_whitelist,
'resources.validate_abstract': grad_validate_abstract
})
def test_GradTester():
def f(x, y):
return x / y
def df(x, y, dz):
return dz / y, -dz * x / (y * y)