def _fuse_nnpack_convrelu(self, net, expected_result_num_ops,
expected_activation_arg=True):
self._add_nnpack(net)
transformer.FuseNNPACKConvRelu(net)
self.assertEquals(tu.numOps(net), expected_result_num_ops)
has_activation_arg = False
for arg in net.Proto().op[0].arg:
if tu.str_compare(arg.name, "activation"):
assert tu.str_compare(arg.s, "Relu")
has_activation_arg = True
if expected_activation_arg:
assert has_activation_arg
else:
assert not has_activation_arg
def test_transformer_FuseConvBNNoConvBiasDuplicatedName(self, size, input_channels, seed, order, epsilon):
workspace.ResetWorkspace()
net = core.Net("net")
c = input_channels
h = size
w = size
k = 3
net.Conv(["X", "w"], ["Y"], stride=1, pad=0, kernel=k, order=order)
net.SpatialBN(
["Y", "scale", "_bias0", "mean", "var"],
["Y2"],
is_test=True,
order=order,
epsilon=epsilon,
)
np.random.seed(seed)
if order == "NCHW":
tu.randBlobFloat32("X", 1, c, h, w)
tu.randBlobFloat32("w", c, c, k, k)
else:
tu.randBlobFloat32("X", 1, h, w, c)
tu.randBlobFloat32("w", c, k, k, c)
tu.randBlobsFloat32(["scale", "_bias0", "mean"], c)
# This is necessary because 1/sqrt(var) is used and if var is too small
# we get floating point artifacts that cause test failures
tu.randBlobFloat32("var", c, offset=0.5)
workspace.RunNetOnce(net)
preTransformOutput = workspace.FetchBlob("Y2").flatten()
workspace.FeedBlob("Y2", np.zeros((1, 1)))
transformer.FuseConvBN(net)
# Ensure fusion
assert tu.numOps(net) == 1
workspace.RunNetOnce(net)
postTransformOutput = workspace.FetchBlob("Y2").flatten()
print("pre")
print(preTransformOutput)
print("after")
print(postTransformOutput)
# Check that there is no numerical difference
assert np.allclose(
preTransformOutput,
postTransformOutput,
rtol=5e-02,
atol=1e-03
)
def test_transformer_FuseConv3DBN(
self, size, input_channels, kt, kh, kw, seed, epsilon
):
workspace.ResetWorkspace()
net = core.Net("net")
c = input_channels
t = size
h = size
w = size
net.Conv(
["X", "w", "b"],
["Y"],
kernels=[kt, kh, kw],
)
net.SpatialBN(
["Y", "scale", "bias", "mean", "var"],
["Y2"],
is_test=True,
epsilon=epsilon,
)
np.random.seed(seed)
tu.randBlobFloat32("X", 1, c, t, h, w)
tu.randBlobFloat32("w", c, c, kt, kh, kw)
tu.randBlobsFloat32(["b", "scale", "bias", "mean"], c)
# This is necessary because 1/sqrt(var) is used and if var is too small
# we get floating point artifacts that cause test failures
tu.randBlobFloat32("var", c, offset=0.5)
workspace.RunNetOnce(net)
preTransformOutput = workspace.FetchBlob("Y2").flatten()
workspace.FeedBlob("Y2", np.zeros((1, 1)))
transformer.FuseConvBN(net)
# Ensure fusion
assert tu.numOps(net) == 1
workspace.RunNetOnce(net)
postTransformOutput = workspace.FetchBlob("Y2").flatten()
# Check that there is no numerical difference
assert np.allclose(
preTransformOutput,
postTransformOutput,
rtol=1e-02,
atol=1e-04
)
