def test_quant_stub():
normal_net = Float.QuantStub()
normal_net.eval()
qat_from_float = QAT.QuantStub.from_float_module(normal_net)
qat_from_float.eval()
disable_observer(qat_from_float)
disable_fake_quant(qat_from_float)
qat_net = QAT.QuantStub()
qat_net.eval()
disable_observer(qat_net)
propagate_qconfig(qat_net, min_max_fakequant_qconfig)
init_qat_net(qat_net)
q_net = Q.QuantStub.from_qat_module(qat_net)
q_net.eval()
x = mge.tensor(np.random.normal(size=(3, 3)).astype("float32"))
normal = normal_net(x)
qat_without_fakequant = qat_from_float(x)
fake_quant_normal = fake_quant_act(normal_net(x), act_scale)
qat = qat_net(x)
q = q_net(x).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_linear():
normal_net = Float.Linear(3, 3, bias=True)
normal_net.eval()
qat_net = QAT.Linear(3, 3, bias=True)
qat_net.eval()
disable_observer(qat_net)
propagate_qconfig(qat_net, min_max_fakequant_qconfig)
init_qat_net(qat_net)
x = mge.tensor(np.random.normal(size=(3, 3)).astype("float32"))
x = fake_quant(x, inp_scale)
x.q_dict["scale"] = inp_scale
x_int8 = quant(x, inp_scale)
weight = np.random.normal(size=(3, 3)).astype("float32")
bias = np.random.normal(size=(3,)).astype("float32")
normal_net.weight.set_value(fake_quant(weight, weight_scale))
normal_net.bias.set_value(fake_quant(bias, inp_scale * weight_scale))
qat_net.weight.set_value(weight)
qat_net.bias.set_value(bias)
q_net = Q.Linear.from_qat_module(qat_net)
q_net.eval()
normal_out = fake_quant(normal_net(x), act_scale)
qat_out = qat_net(x)
q_out = q_net(x_int8).numpy() * act_scale
np.testing.assert_allclose(qat_out, normal_out)
np.testing.assert_allclose(q_out, normal_out.numpy())
def init_qat_net():
net = QATNet()
propagate_qconfig(net, min_max_fakequant_qconfig)
min_val = np.random.randint(-127, 0, size=(3,))
max_val = np.random.randint(1, 127, size=(3,))
net.quant.act_observer.min_val.set_value(min_val[0])
net.quant.act_observer.max_val.set_value(max_val[0])
net.linear.weight_observer.min_val.set_value(min_val[1])
net.linear.weight_observer.max_val.set_value(max_val[1])
net.linear.act_observer.min_val.set_value(min_val[2])
net.linear.act_observer.max_val.set_value(max_val[2])
return net
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_conv(module):
normal_net = getattr(Float, module)(3, 3, 3, 1, 1, 1, bias=True)
normal_net.eval()
qat_net = getattr(QAT, module)(3, 3, 3, 1, 1, 1, bias=True)
qat_net.eval()
disable_observer(qat_net)
propagate_qconfig(qat_net, min_max_fakequant_qconfig)
init_qat_net(qat_net)
x = mge.tensor(np.random.normal(size=(1, 3, 3, 3)).astype("float32"))
inp_scale = gen_inp_scale()
x = fake_quant_act(x, inp_scale)
x.qparams.update(create_qparams(QuantMode.SYMMERTIC, "qint8", inp_scale))
x_int8 = quant(x, inp_scale)
weight = np.random.normal(size=(3, 3, 3, 3)).astype("float32")
bias = np.random.normal(size=(1, 3, 1, 1)).astype("float32")
if module in ("ConvBn2d", "ConvBnRelu2d"):
normal_net.conv.weight[...] = fake_quant_weight(weight, weight_scale)
normal_net.conv.bias[...] = fake_quant_bias(bias,
inp_scale * weight_scale)
qat_net.conv.weight[...] = Parameter(weight)
qat_net.conv.bias[...] = Parameter(bias)
else:
normal_net.weight[...] = fake_quant_weight(weight, weight_scale)
normal_net.bias[...] = fake_quant_bias(bias, inp_scale * weight_scale)
qat_net.weight[...] = Parameter(weight)
qat_net.bias[...] = Parameter(bias)
qat_from_float = getattr(QAT, module).from_float_module(normal_net)
qat_from_float.eval()
disable_observer(qat_from_float)
disable_fake_quant(qat_from_float)
q_net = getattr(Q, module).from_qat_module(qat_net)
q_net.eval()
normal = normal_net(x)
qat_without_fakequant = qat_from_float(x)
fake_quant_normal = fake_quant_act(normal_net(x), act_scale)
qat = qat_net(x)
q = q_net(x_int8).numpy() * act_scale
np.testing.assert_allclose(qat_without_fakequant, normal, atol=1e-5)
np.testing.assert_allclose(qat, fake_quant_normal, atol=act_scale)
np.testing.assert_allclose(q, fake_quant_normal.numpy(), atol=act_scale)
def test_conv(module):
normal_net = getattr(Float, module)(3, 3, 3, 1, 1, 1, bias=True)
normal_net.eval()
qat_net = getattr(QAT, module)(3, 3, 3, 1, 1, 1, bias=True)
qat_net.eval()
disable_observer(qat_net)
propagate_qconfig(qat_net, min_max_fakequant_qconfig)
init_qat_net(qat_net)
x = mge.tensor(np.random.normal(size=(1, 3, 3, 3)).astype("float32"))
x = fake_quant(x, inp_scale)
x.q_dict["scale"] = inp_scale
x_int8 = quant(x, inp_scale)
weight = np.random.normal(size=(3, 3, 3, 3)).astype("float32")
bias = np.random.normal(size=(1, 3, 1, 1)).astype("float32")
if module in ("ConvBn2d", "ConvBnRelu2d"):
normal_net.conv.weight.set_value(fake_quant(weight, weight_scale))
normal_net.conv.bias.set_value(
fake_quant(bias, inp_scale * weight_scale))
qat_net.conv.weight.set_value(weight)
qat_net.conv.bias.set_value(bias)
else:
normal_net.weight.set_value(fake_quant(weight, weight_scale))
normal_net.bias.set_value(fake_quant(bias, inp_scale * weight_scale))
qat_net.weight.set_value(weight)
qat_net.bias.set_value(bias)
qat_from_float = getattr(QAT, module).from_float_module(normal_net)
qat_from_float.eval()
disable_observer(qat_from_float)
disable_fake_quant(qat_from_float)
q_net = getattr(Q, module).from_qat_module(qat_net)
q_net.eval()
normal = normal_net(x)
qat_without_fakequant = qat_from_float(x)
fake_quant_normal = fake_quant(normal_net(x), act_scale)
qat = qat_net(x)
q = q_net(x_int8).numpy() * act_scale
np.testing.assert_allclose(qat_without_fakequant, normal, atol=1e-6)
np.testing.assert_allclose(qat, fake_quant_normal)
np.testing.assert_allclose(q, fake_quant_normal.numpy())
def test_propagate_qconfig():
net = QATNet()
propagate_qconfig(net, min_max_fakequant_qconfig)
assert all([
net.quant.weight_observer is None,
net.quant.weight_fake_quant is None,
isinstance(net.quant.act_observer, MinMaxObserver),
isinstance(net.quant.act_fake_quant, FakeQuantize),
isinstance(net.linear.weight_observer, MinMaxObserver),
isinstance(net.linear.weight_fake_quant, FakeQuantize),
isinstance(net.linear.act_observer, MinMaxObserver),
isinstance(net.linear.act_fake_quant, FakeQuantize),
net.dequant.weight_observer is None,
net.dequant.weight_fake_quant is None,
net.dequant.act_observer is None,
net.dequant.act_observer is None,
])
def test_linear():
normal_net = Float.Linear(3, 3, bias=True)
normal_net.eval()
qat_net = QAT.Linear(3, 3, bias=True)
qat_net.eval()
disable_observer(qat_net)
propagate_qconfig(qat_net, min_max_fakequant_qconfig)
init_qat_net(qat_net)
x = mge.tensor(np.random.normal(size=(3, 3)).astype("float32"))
inp_scale = gen_inp_scale()
x = fake_quant_act(x, inp_scale)
x.qparams.update(create_qparams(QuantMode.SYMMERTIC, "qint8", inp_scale))
x_int8 = quant(x, inp_scale)
weight = np.random.normal(size=(3, 3)).astype("float32")
bias = np.random.normal(size=(3, )).astype("float32")
normal_net.weight[...] = fake_quant_weight(weight, weight_scale)
normal_net.bias[...] = fake_quant_bias(bias, inp_scale * weight_scale)
qat_net.weight[...] = Parameter(weight)
qat_net.bias[...] = Parameter(bias)
qat_from_float = QAT.Linear.from_float_module(normal_net)
qat_from_float.eval()
disable_fake_quant(qat_from_float)
disable_observer(qat_from_float)
q_net = Q.Linear.from_qat_module(qat_net)
q_net.eval()
normal = normal_net(x)
qat_without_fakequant = qat_from_float(x)
fake_quant_normal = fake_quant_act(normal_net(x), act_scale)
qat = qat_net(x)
q = q_net(x_int8).numpy() * act_scale
np.testing.assert_allclose(qat_without_fakequant, normal)
np.testing.assert_allclose(qat, fake_quant_normal.numpy())
np.testing.assert_allclose(q, fake_quant_normal.numpy())
def test_dequant_stub():
normal_net = Float.DequantStub()
normal_net.eval()
qat_net = QAT.DequantStub()
qat_net.eval()
disable_observer(qat_net)
propagate_qconfig(qat_net, min_max_fakequant_qconfig)
init_qat_net(qat_net)
q_net = Q.DequantStub.from_qat_module(qat_net)
q_net.eval()
x = mge.tensor(np.random.normal(size=(3, 3)).astype("float32"))
x = fake_quant(x, inp_scale)
x.q_dict["scale"] = inp_scale
normal_out = normal_net(x)
qat_out = qat_net(x)
q_out = q_net(quant(x, inp_scale)).numpy()
np.testing.assert_allclose(qat_out, normal_out)
np.testing.assert_allclose(q_out, normal_out.numpy())
def test_concat():
normal_net = Float.Concat()
normal_net.eval()
qat_net = QAT.Concat()
qat_net.eval()
disable_observer(qat_net)
propagate_qconfig(qat_net, min_max_fakequant_qconfig)
init_qat_net(qat_net)
inps = []
inps_int8 = []
for i in range(3):
inp_scale = gen_inp_scale()
inps.append(mge.tensor(
np.random.normal(size=(3, 3)).astype("float32")))
inps[i] = fake_quant_act(inps[i], inp_scale)
inps[i].qparams.update(
create_qparams(QuantMode.SYMMERTIC, "qint8", inp_scale))
inps_int8.append(quant(inps[i], inp_scale))
qat_from_float = QAT.Concat.from_float_module(normal_net)
qat_from_float.eval()
disable_fake_quant(qat_from_float)
disable_observer(qat_from_float)
q_net = Q.Concat.from_qat_module(qat_net)
q_net.eval()
normal = normal_net(inps)
qat_without_fakequant = qat_from_float(inps)
fake_quant_normal = fake_quant_act(normal_net(inps), act_scale)
qat = qat_net(inps)
q = q_net(inps_int8).numpy() * act_scale
np.testing.assert_allclose(qat_without_fakequant, normal)
np.testing.assert_allclose(qat, fake_quant_normal.numpy())
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())
