def _run(self):
frame = TFFramework()
if self.args['list']:
print 'Supported Exemplars:'
for primop in frame.ExemplarRegistry:
print ' ',primop
return 0
elif self.args['primop'] is None:
self.subparser.error('A primop must be specified unless using --list')
primop_t = self.args['primop']
uas = UniformArgSampler()
primop_args = uas.sample(primop_t, n=self.args['n'], seed=self.args['seed'])
Exemplar = frame.ExemplarRegistry.lookup(primop_t)
self.features = Features()
for p_args in primop_args:
exemplar = Exemplar(p_args)
synthmodel = frame.SyntheticModel(exemplar)
frame.set_model(synthmodel)
profile = frame.get_timing(mode='inference', ops='native')
graph = frame.get_graph(mode='inference', scope='dynamic', ops='native')
id = graph.get_vertex_id_from_tf_name( exemplar.get_op_name() )
self.features.append( p_args, profile[id][0] )
if self.args['output'] is None:
self.args['output'] = str(primop_t)+'.features'
self.features.write(self.args['output'])
def test_argument_length_mismatch(self):
f = Features()
f.append([1,2,3], 0)
with self.assertRaises(AssertionError):
f.append([1,2,3,4], 0)
with self.assertRaises(AssertionError):
f.append([1,2], 0)
f.append([4,5,6], 0)
class SweepTool(AbstractTool):
TOOL_NAME='sweep'
TOOL_SUMMARY='Generating timing measurements suitable for training a performance estimator.'
CACHE_FORMAT=[] # FIXME: Features are not cached.
def add_subparser(self, argparser):
super(SweepTool, self).add_subparser(argparser)
self.subparser.add_argument('primop', nargs='?', type=str, help='Which primitive operation to sweep.')
self.subparser.add_argument('-n', type=int, default=1, help='Number of data points to collect.')
self.subparser.add_argument('-o', '--output', type=str, default=None, help='Output file with Estimator training features (default: <primop>.features).')
self.subparser.add_argument('--seed', type=int, default=None, help='Set the PRNG seed for this sweep.')
self.subparser.add_argument('-l','--list', default=False, action='store_true', help="Don't actually run a sweep, just print which primops are supported,")
# FIXME: Add argument to limit scope of sweep
return self.subparser
def _run(self):
frame = TFFramework()
if self.args['list']:
print 'Supported Exemplars:'
for primop in frame.ExemplarRegistry:
print ' ',primop
return 0
elif self.args['primop'] is None:
self.subparser.error('A primop must be specified unless using --list')
primop_t = self.args['primop']
uas = UniformArgSampler()
primop_args = uas.sample(primop_t, n=self.args['n'], seed=self.args['seed'])
Exemplar = frame.ExemplarRegistry.lookup(primop_t)
self.features = Features()
for p_args in primop_args:
exemplar = Exemplar(p_args)
synthmodel = frame.SyntheticModel(exemplar)
frame.set_model(synthmodel)
profile = frame.get_timing(mode='inference', ops='native')
graph = frame.get_graph(mode='inference', scope='dynamic', ops='native')
id = graph.get_vertex_id_from_tf_name( exemplar.get_op_name() )
self.features.append( p_args, profile[id][0] )
if self.args['output'] is None:
self.args['output'] = str(primop_t)+'.features'
self.features.write(self.args['output'])
# FIXME: Features objects cannot be cached in the normal way.
#def _load(self, filename):
# feats = Features()
# feats.read(filename)
# return feats
#def _save(self, data, filename): # pylint: disable=W0221
# data.write(filename)
def _output(self):
print 'Output'
for args,time in zip(self.features.op_arguments, self.features.measurements):
print args,':',time
def test_accessors(self): f = Features() f.append([1,2,3], 10) f.append([4,5,6], 11) op_args = f.op_arguments meas = f.measurements assert op_args.shape==(2,3) assert meas.shape==(2,) assert meas[0]==10 and meas[1]==11
def test_reproducible_parameters(self):
N = 1000
np.random.seed(13)
noise = 20 * np.random.randn(N)
line_x = np.random.randint(0, 1000, (N, 1))
m_true, b_true = 3, 20
line_y = m_true * line_x[:, 0] + b_true + noise
v_true = m_true * 100 + b_true
feats = Features().extend(line_x, line_y)
est = OLSEstimator()
est.fit(feats)
v = est.estimate([100])
assert np.abs(v - v_true) < 10., 'Incorrect estimation'
params = est.get_params()
assert np.abs(m_true - params[0]) < 1., 'Incorrect slope fit'
assert np.abs(b_true - params[1]) < 10., 'Incorrect intercept fit'
v = est.estimate([100])
assert np.abs(v - v_true) < 10., 'Incorrect estimation'
est2 = OLSEstimator()
est2.set_params(params)
v = est2.estimate([100])
assert np.abs(v - v_true) < 10., 'Incorrect estimation'
def _run(self):
if self.args['output'] is None:
self.args['output'] = str(self.args['primop']) + '-' + str(
self.args['estimator']) + '.est'
EstimatorClass = EstimatorRegistry.lookup(self.args['estimator'])
est = EstimatorClass()
f = Features()
for feats_file in self.args['features']:
f2 = Features()
f2.read(feats_file)
f.concatenate(f2)
est.fit(f)
self.data = est.get_params()
self._save(self.data, self.args['output'])
def test_io(self):
with in_temporary_directory() as d:
filename = os.path.join(os.path.abspath(d), 'zero-linear.est')
est = OLSEstimator()
feats = Features().append([0, 1, 2, 3], 0)
est.fit(feats)
est.write(filename)
# FIXME: flush?
est2 = OLSEstimator()
est2.read(filename)
assert len(est.get_params()) == len(est2.get_params())
for i, (lhs,
rhs) in enumerate(zip(est.get_params(),
est2.get_params())):
assert lhs == rhs, 'Parameter ' + str(
i) + ' does not match: "' + str(lhs) + '" vs. "' + str(
rhs) + '"'
def test_blind_predict(self):
est = OLSEstimator()
feats = Features().append([0], 0)
est.fit(feats)
v = est.estimate([0])
assert (v + 1), 'Estimate returned non-numeric value: ' + str(v)
def test_instantiate(self): _ = Features()
def test_io(self):
f = Features()
f.append([1,2,3], 10)
f.append([2,3,4], 11)
f.append([3,4,5], 12)
with in_temporary_directory():
f.write('test_file')
f2 = Features()
f2.read('test_file')
assert np.all(f.array==f2.array)
def test_invalid_shapes(self):
f = Features()
with self.assertRaises(AssertionError):
f.append(1, 0)
with self.assertRaises(AssertionError):
f.append([1,2,3],[4,5,6])
def test_add(self): f = Features() f.append([1,2,3], 4) f.append(np.array([1,2,3]), np.array(4)) assert len(f.measurements)==2 f2 = Features() f2.extend([[1,2,3],[4,5,6]], [7,8]) assert len(f2.measurements)==2 f2.concatenate(f) assert len(f2.measurements)==4