def __init__(self, ):
super().__init__()
self.data = np.ones((2, 3, 224, 224)).astype(np.float32)
self.data1 = np.random.random((1, 3, 1, 1)).astype(np.float32)
self.add1 = M.Elemwise("add")
self.add2 = M.Elemwise("add")
self.add3 = M.Elemwise("add")
scale = mge.tensor((16.0 / 128.0))
self.quant_stub = QuantStub()
self.quant_stub.act_fake_quant = FakeQuantize(
_builtin_quant_dtypes["qint8"])
self.quant_stub.act_fake_quant.set_qparams(
create_qparams(
dtype_meta=_builtin_quant_dtypes["qint8"],
scale=scale,
zero_point=None,
))
self.quant_stub1 = QuantStub()
self.quant_stub1.act_fake_quant = FakeQuantize(
_builtin_quant_dtypes["qint8"])
self.quant_stub1.act_fake_quant.set_qparams(
create_qparams(
dtype_meta=_builtin_quant_dtypes["qint8"],
scale=scale,
zero_point=None,
))
def __init__(self, inp, oup, stride, expand_ratio):
super(InvertedResidual, self).__init__()
self.stride = stride
assert stride in [1, 2]
hidden_dim = int(round(inp * expand_ratio))
self.use_res_connect = self.stride == 1 and inp == oup
layers = []
if expand_ratio != 1:
# pw
layers.append(
M.ConvBnRelu2d(inp, hidden_dim, kernel_size=1, bias=False))
layers.extend([
# dw
M.ConvBnRelu2d(
hidden_dim,
hidden_dim,
kernel_size=3,
padding=1,
stride=stride,
groups=hidden_dim,
bias=False,
),
# pw-linear
M.ConvBn2d(hidden_dim, oup, kernel_size=1, bias=False),
])
self.conv = M.Sequential(*layers)
self.add = M.Elemwise("ADD")
def __init__(
self,
in_channels,
channels,
stride=1,
groups=1,
base_width=64,
dilation=1,
norm=M.BatchNorm2d,
):
assert norm is M.BatchNorm2d, "Quant mode only support BatchNorm2d currently."
super(Bottleneck, self).__init__()
width = int(channels * (base_width / 64.0)) * groups
self.conv_bn_relu1 = M.ConvBnRelu2d(in_channels, width, 1, 1, bias=False)
self.conv_bn_relu2 = M.ConvBnRelu2d(
width,
width,
3,
stride,
padding=dilation,
groups=groups,
dilation=dilation,
bias=False,
)
self.conv_bn3 = M.ConvBn2d(width, channels * self.expansion, 1, 1, bias=False)
self.downsample = (
M.Identity()
if in_channels == channels * self.expansion and stride == 1
else M.ConvBn2d(
in_channels, channels * self.expansion, 1, stride, bias=False
)
)
self.add = M.Elemwise("FUSE_ADD_RELU")
def __init__(
self,
in_channels,
channels,
stride=1,
groups=1,
base_width=64,
dilation=1,
norm=M.BatchNorm2d,
):
assert norm is M.BatchNorm2d, "Quant mode only support BatchNorm2d currently."
super(BasicBlock, self).__init__()
if groups != 1 or base_width != 64:
raise ValueError("BasicBlock only supports groups=1 and base_width=64")
if dilation > 1:
raise NotImplementedError("Dilation > 1 not supported in BasicBlock")
self.conv_bn_relu1 = M.ConvBnRelu2d(
in_channels, channels, 3, stride, padding=dilation, bias=False
)
self.conv_bn2 = M.ConvBn2d(channels, channels, 3, 1, padding=1, bias=False)
self.downsample = (
M.Identity()
if in_channels == channels and stride == 1
else M.ConvBn2d(in_channels, channels, 1, stride, bias=False)
)
self.add = M.Elemwise("FUSE_ADD_RELU")
def __init__(self):
super().__init__()
self.modules = [
M.Elemwise("ADD"),
M.Elemwise("ADD"),
OrderedDict([("a", M.Elemwise("ADD")), ("b", M.Elemwise("ADD"))]),
M.Elemwise("RELU"),
M.Elemwise("RELU"),
]
def test_elemwise(kind):
normal_net = Float.Elemwise(kind)
normal_net.eval()
qat_from_float = QAT.Elemwise.from_float_module(normal_net)
qat_from_float.eval()
disable_observer(qat_from_float)
disable_fake_quant(qat_from_float)
qat_net = QAT.Elemwise(kind)
qat_net.eval()
disable_observer(qat_net)
propagate_qconfig(qat_net, min_max_fakequant_qconfig)
init_qat_net(qat_net)
q_net = Q.Elemwise.from_qat_module(qat_net)
q_net.eval()
x1_scale = np.float32(np.random.rand() + 1)
x1 = mge.tensor(np.random.normal(size=(3, 3)).astype("float32"))
x1 = fake_quant_act(x1, x1_scale)
x1.qparams.scale = x1_scale
x2_scale = np.float32(np.random.rand() + 1)
x2 = mge.tensor(np.random.normal(size=(3, 3)).astype("float32"))
x2 = fake_quant_act(x2, x2_scale)
x2.qparams.scale = x2_scale
x1_int8 = quant(x1, x1_scale)
x2_int8 = quant(x2, x2_scale)
# test correctness of `Float`, `QAT` and `Quantized`
if kind in ("add", "mul", "fuse_add_relu"):
normal = normal_net(x1, x2)
qat_without_fakequant = qat_from_float(x1, x2)
fake_quant_normal = fake_quant_act(normal_net(x1, x2), act_scale)
qat = qat_net(x1, x2)
q = q_net(x1_int8, x2_int8).numpy() * act_scale
else:
normal = normal_net(x1)
qat_without_fakequant = qat_from_float(x1)
fake_quant_normal = fake_quant_act(normal_net(x1), act_scale)
qat = qat_net(x1)
q = q_net(x1_int8).numpy() * act_scale
np.testing.assert_allclose(qat_without_fakequant, normal)
np.testing.assert_allclose(qat, fake_quant_normal)
np.testing.assert_allclose(q, fake_quant_normal.numpy())
def test_elemwise(kind):
normal_net = Float.Elemwise(kind)
normal_net.eval()
qat_net = QAT.Elemwise(kind)
qat_net.eval()
disable_observer(qat_net)
propagate_qconfig(qat_net, min_max_fakequant_qconfig)
init_qat_net(qat_net)
q_net = Q.Elemwise.from_qat_module(qat_net)
q_net.eval()
x1_scale = np.float32(np.random.rand() + 1)
x1 = mge.tensor(np.random.normal(size=(3, 3)).astype("float32"))
x1 = fake_quant(x1, x1_scale)
x1.q_dict["scale"] = x1_scale
x2_scale = np.float32(np.random.rand() + 1)
x2 = mge.tensor(np.random.normal(size=(3, 3)).astype("float32"))
x2 = fake_quant(x2, x2_scale)
x2.q_dict["scale"] = x2_scale
x1_int8 = quant(x1, x1_scale)
x2_int8 = quant(x2, x2_scale)
if kind in ("ADD", "MUL", "FUSE_ADD_RELU"):
normal_out = fake_quant(normal_net(x1, x2), act_scale)
qat_out = qat_net(x1, x2)
q_out = q_net(x1_int8, x2_int8).numpy() * act_scale
else:
normal_out = fake_quant(normal_net(x1), act_scale)
qat_out = qat_net(x1)
q_out = q_net(x1_int8).numpy() * act_scale
np.testing.assert_allclose(qat_out, normal_out)
np.testing.assert_allclose(q_out, normal_out.numpy())
def __init__(self, ):
super().__init__()
self.conv0 = MyConvBnRelu2d(3, 3, 3, 1, 1)
self.conv1 = M.ConvBn2d(3, 3, 1, 1, 0)
self.conv2 = M.ConvBn2d(3, 3, 1, 1, 0)
self.add = M.Elemwise("FUSE_ADD_RELU")
def __init__(self):
super().__init__()
self.a = M.Elemwise("ADD")
self.b = self.a
