def testWhile(self):
with NetBuilder(_use_control_ops=True) as nb:
ops.Copy(ops.Const(0), "i")
ops.Copy(ops.Const(1), "one")
ops.Copy(ops.Const(2), "two")
ops.Copy(ops.Const(2.0), "x")
ops.Copy(ops.Const(3.0), "y")
ops.Copy(ops.Const(2.0), "z")
# raises x to the power of 4 and y to the power of 2
# and z to the power of 3
with ops.WhileNet():
with ops.Condition():
ops.Add(["i", "one"], "i")
ops.LE(["i", "two"])
ops.Pow("x", "x", exponent=2.0)
with ops.IfNet(ops.LT(["i", "two"])):
ops.Pow("y", "y", exponent=2.0)
with ops.Else():
ops.Pow("z", "z", exponent=3.0)
ops.Add(["x", "y"], "x_plus_y")
ops.Add(["x_plus_y", "z"], "s")
assert len(nb.get()) == 1, "Expected a single net produced"
net = nb.get()[0]
net.AddGradientOperators(["s"])
workspace.RunNetOnce(net)
# (x^4)' = 4x^3
self.assertAlmostEqual(workspace.FetchBlob("x_grad"), 32)
self.assertAlmostEqual(workspace.FetchBlob("x"), 16)
# (y^2)' = 2y
self.assertAlmostEqual(workspace.FetchBlob("y_grad"), 6)
self.assertAlmostEqual(workspace.FetchBlob("y"), 9)
# (z^3)' = 3z^2
self.assertAlmostEqual(workspace.FetchBlob("z_grad"), 12)
self.assertAlmostEqual(workspace.FetchBlob("z"), 8)
def test_if_net(self):
with NetBuilder() as nb:
x0 = ops.Const(0)
x1 = ops.Const(1)
x2 = ops.Const(2)
y0 = ops.Const(0)
y1 = ops.Const(1)
y2 = ops.Const(2)
# basic logic
first_res = ops.Const(0)
with ops.IfNet(ops.Const(True)):
ops.Const(1, blob_out=first_res)
with ops.Else():
ops.Const(2, blob_out=first_res)
second_res = ops.Const(0)
with ops.IfNet(ops.Const(False)):
ops.Const(1, blob_out=second_res)
with ops.Else():
ops.Const(2, blob_out=second_res)
# nested and sequential ifs,
# empty then/else,
# passing outer blobs into branches,
# writing into outer blobs, incl. into input blob
# using local blobs
with ops.IfNet(ops.LT([x0, x1])):
local_blob = ops.Const(900)
ops.Add([ops.Const(100), local_blob], [y0])
gt = ops.GT([x1, x2])
with ops.IfNet(gt):
# empty then
pass
with ops.Else():
ops.Add([y1, local_blob], [local_blob])
ops.Add([ops.Const(100), y1], [y1])
with ops.IfNet(ops.EQ([local_blob, ops.Const(901)])):
ops.Const(7, blob_out=y2)
ops.Add([y1, y2], [y2])
with ops.Else():
# empty else
pass
plan = Plan('if_net_test')
plan.AddStep(to_execution_step(nb))
ws = workspace.C.Workspace()
ws.run(plan)
first_res_value = ws.blobs[str(first_res)].fetch()
second_res_value = ws.blobs[str(second_res)].fetch()
y0_value = ws.blobs[str(y0)].fetch()
y1_value = ws.blobs[str(y1)].fetch()
y2_value = ws.blobs[str(y2)].fetch()
self.assertEquals(first_res_value, 1)
self.assertEquals(second_res_value, 2)
self.assertEquals(y0_value, 1000)
self.assertEquals(y1_value, 101)
self.assertEquals(y2_value, 108)
self.assertTrue(str(local_blob) not in ws.blobs)
def testIf(self):
W_a_values = [2.0, 1.5]
B_a_values = [0.5]
W_b_values = [7.0, 3.5]
B_b_values = [1.5]
with NetBuilder(_use_control_ops=True) as init_nb:
W_a = ops.UniformFill([], "W_a", shape=[1, 2], min=-1., max=1.)
B_a = ops.ConstantFill([], "B_a", shape=[1], value=0.0)
W_b = ops.UniformFill([], "W_b", shape=[1, 2], min=-1., max=1.)
B_b = ops.ConstantFill([], "B_b", shape=[1], value=0.0)
W_gt_a = ops.GivenTensorFill(
[], "W_gt_a", shape=[1, 2], values=W_a_values)
B_gt_a = ops.GivenTensorFill([], "B_gt_a", shape=[1], values=B_a_values)
W_gt_b = ops.GivenTensorFill(
[], "W_gt_b", shape=[1, 2], values=W_b_values)
B_gt_b = ops.GivenTensorFill([], "B_gt_b", shape=[1], values=B_b_values)
params = [W_gt_a, B_gt_a, W_a, B_a, W_gt_b, B_gt_b, W_b, B_b]
with NetBuilder(_use_control_ops=True, initial_scope=params) as train_nb:
Y_pred = ops.ConstantFill([], "Y_pred", shape=[1], value=0.0)
Y_noise = ops.ConstantFill([], "Y_noise", shape=[1], value=0.0)
switch = ops.UniformFill(
[], "switch", shape=[1], min=-1., max=1., run_once=0)
zero = ops.ConstantFill([], "zero", shape=[1], value=0.0)
X = ops.GaussianFill(
[], "X", shape=[4096, 2], mean=0.0, std=1.0, run_once=0)
noise = ops.GaussianFill(
[], "noise", shape=[4096, 1], mean=0.0, std=1.0, run_once=0)
with ops.IfNet(ops.LT([switch, zero])):
Y_gt = ops.FC([X, W_gt_a, B_gt_a], "Y_gt")
ops.Add([Y_gt, noise], Y_noise)
ops.FC([X, W_a, B_a], Y_pred)
with ops.Else():
Y_gt = ops.FC([X, W_gt_b, B_gt_b], "Y_gt")
ops.Add([Y_gt, noise], Y_noise)
ops.FC([X, W_b, B_b], Y_pred)
dist = ops.SquaredL2Distance([Y_noise, Y_pred], "dist")
loss = dist.AveragedLoss([], ["loss"])
assert len(init_nb.get()) == 1, "Expected a single init net produced"
assert len(train_nb.get()) == 1, "Expected a single train net produced"
train_net = train_nb.get()[0]
gradient_map = train_net.AddGradientOperators([loss])
init_net = init_nb.get()[0]
ITER = init_net.ConstantFill(
[], "ITER", shape=[1], value=0, dtype=core.DataType.INT64)
train_net.Iter(ITER, ITER)
LR = train_net.LearningRate(ITER, "LR", base_lr=-0.1,
policy="step", stepsize=20, gamma=0.9)
ONE = init_net.ConstantFill([], "ONE", shape=[1], value=1.)
train_net.WeightedSum([W_a, ONE, gradient_map[W_a], LR], W_a)
train_net.WeightedSum([B_a, ONE, gradient_map[B_a], LR], B_a)
train_net.WeightedSum([W_b, ONE, gradient_map[W_b], LR], W_b)
train_net.WeightedSum([B_b, ONE, gradient_map[B_b], LR], B_b)
workspace.RunNetOnce(init_net)
workspace.CreateNet(train_net)
# print("Before training, W_a is: {}".format(workspace.FetchBlob("W_a")))
# print("Before training, B_a is: {}".format(workspace.FetchBlob("B_a")))
# print("Before training, W_b is: {}".format(workspace.FetchBlob("W_b")))
# print("Before training, B_b is: {}".format(workspace.FetchBlob("B_b")))
for _epoch in range(1000):
workspace.RunNet(train_net.Proto().name)
# print("After training, W_a is: {}".format(workspace.FetchBlob("W_a")))
# print("After training, B_a is: {}".format(workspace.FetchBlob("B_a")))
# print("After training, W_b is: {}".format(workspace.FetchBlob("W_b")))
# print("After training, B_b is: {}".format(workspace.FetchBlob("B_b")))
# print("Ground truth W_a is: {}".format(workspace.FetchBlob("W_gt_a")))
# print("Ground truth B_a is: {}".format(workspace.FetchBlob("B_gt_a")))
# print("Ground truth W_b is: {}".format(workspace.FetchBlob("W_gt_b")))
# print("Ground truth B_b is: {}".format(workspace.FetchBlob("B_gt_b")))
values_map = {
"W_a": W_a_values,
"B_a": B_a_values,
"W_b": W_b_values,
"B_b": B_b_values,
}
train_eps = 0.01
for blob_name, values in values_map.items():
trained_values = workspace.FetchBlob(blob_name)
if trained_values.ndim == 2:
self.assertEqual(trained_values.shape[0], 1)
trained_values = trained_values[0][:]
else:
self.assertEqual(trained_values.ndim, 1)
self.assertEqual(trained_values.size, len(values))
for idx in range(len(trained_values)):
self.assertTrue(abs(trained_values[idx] - values[idx]) < train_eps)
# In[1]:
from caffe2.python import workspace
from caffe2.python.core import Plan, to_execution_step, Net
from caffe2.python.net_builder import ops, NetBuilder
# In[2]:
with NetBuilder() as nb:
ops.Const(0.0, blob_out="zero")
ops.Const(1.0, blob_out="one")
ops.Const(0.5, blob_out="x")
ops.Const(0.0, blob_out="y")
with ops.IfNet(ops.GT(["x", "zero"])):
ops.Copy("one", "y")
with ops.Else():
ops.Copy("zero", "y")
# Note the usage of NetBuilder's ops.IfNet and ops.Else calls: ops.IfNet accepts a blob reference or blob name as an input, it expects an input blob to have a scalar value convertible to bool, also note that optional ops.Else is at the same level as ops.IfNet and immediately follows corresponding ops.IfNet. Let's execute resulting net (execution step) and check values of blobs.
# In[3]:
plan = Plan('if_net_test')
plan.AddStep(to_execution_step(nb))
ws = workspace.C.Workspace()
ws.run(plan)
print('x = ', ws.blobs["x"].fetch())
print('y = ', ws.blobs["y"].fetch())
# Before going further, it's important to understand the semantics of execution blocks ('then' and 'else' branches in the example above), i.e. handling of reads and writes into global (defined outside of the block) and local (defined inside the block) blobs.
