def test_elemwise(kind):
x1 = mge.tensor(np.random.normal(size=(3, 3)).astype("float32"))
x1_scale = np.float32(np.random.rand() + 1)
x1 = fake_quant(x1, x1_scale)
x1.qparams.update(create_qparams(QuantMode.SYMMERTIC, "qint8", x1_scale))
x1_int8 = quant(x1, x1_scale)
x2 = mge.tensor(np.random.normal(size=(3, 3)).astype("float32"))
x2_scale = np.float32(np.random.rand() + 1)
x2 = fake_quant(x2, x2_scale)
x2.qparams.update(create_qparams(QuantMode.SYMMERTIC, "qint8", x2_scale))
x2_int8 = quant(x2, x2_scale)
output_scale = np.float32(np.random.rand() + 1)
output_dtype = dtype.qint8(output_scale)
quantized_kind = "q" + kind
if kind in ("abs", "sin"):
desired_out = fake_quant(_elwise(x1, mode=kind), output_scale)
actual_out = (_elemwise_multi_type(
x1_int8, mode=quantized_kind, dtype=output_dtype).numpy() *
output_scale)
else:
desired_out = fake_quant(_elwise(x1, x2, mode=kind), output_scale)
actual_out = (_elemwise_multi_type(
x1_int8, x2_int8, mode=quantized_kind, dtype=output_dtype).numpy()
* output_scale)
np.testing.assert_allclose(actual_out, desired_out.numpy())
def test_passive_observer():
qparams = create_qparams(QuantMode.SYMMERTIC, "qint8", mge.tensor(1.0))
m = PassiveObserver("qint8")
m.set_qparams(qparams)
assert m.orig_scale == 1.0
assert m.scale.numpy() == 1.0
assert m.get_qparams().dtype_meta == qparams.dtype_meta
assert m.get_qparams().scale == qparams.scale
assert m.get_qparams() == qparams
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_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_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())
