def __init__(self):
super().__init__()
self.sub1 = LinearReluModel()
self.sub2 = QuantWrapper(TwoLayerLinearModel())
self.fc3 = QuantWrapper(torch.nn.Linear(5, 5).to(dtype=torch.float))
self.fc3.qconfig = default_qconfig
self.sub2.qconfig = default_qconfig
def __init__(self):
super(AnnotatedNestedModel, self).__init__()
self.sub1 = LinearReluModel()
self.sub2 = TwoLayerLinearModel()
self.fc3 = QuantWrapper(torch.nn.Linear(5, 5).to(dtype=torch.float))
self.fc3.qconfig = default_qconfig
self.sub2.fc1 = QuantWrapper(self.sub2.fc1)
self.sub2.fc1.qconfig = default_per_channel_qconfig
def __init__(self, qengine):
super().__init__()
self.sub1 = LinearReluModel()
self.sub2 = TwoLayerLinearModel()
self.fc3 = QuantWrapper(torch.nn.Linear(5, 5).to(dtype=torch.float))
self.fc3.qconfig = default_qconfig
self.sub2.fc1 = QuantWrapper(self.sub2.fc1)
if qengine == 'fbgemm':
self.sub2.fc1.qconfig = default_per_channel_qconfig
else:
self.sub2.fc1.qconfig = default_qconfig
def __init__(self):
super(SkipQuantModel, self).__init__()
self.qconfig = default_qconfig
self.sub = QuantWrapper(InnerModule())
self.fc = torch.nn.Linear(5, 5).to(dtype=torch.float)
# don't quantize this fc
self.fc.qconfig = None
def __init__(self, qengine):
super().__init__()
self.qconfig = torch.quantization.get_default_qconfig(qengine)
self.sub = QuantWrapper(InnerModule())
self.fc = torch.nn.Linear(5, 5).to(dtype=torch.float)
# don't quantize this fc
self.fc.qconfig = None
def __init__(self, with_relu=False):
super().__init__()
layers = [nn.Conv2d(3, 32, 3, bias=True), nn.BatchNorm2d(32)]
if with_relu:
layers.append(nn.ReLU())
self.conv = nn.Sequential(*layers)
self.quant_wrap = QuantWrapper(self.conv)
self.with_relu = with_relu
def __init__(self, with_relu=False):
super().__init__()
layers = [nn.Linear(16, 32)]
if with_relu:
layers.append(nn.ReLU())
self.fc = nn.Sequential(*layers)
self.quant_wrap = QuantWrapper(self.fc)
self.with_relu = with_relu
def __init__(self):
super().__init__()
self.sub1 = LinearReluModel()
self.sub2 = TwoLayerLinearModel()
self.fc3 = QuantWrapper(torch.nn.Linear(5, 5).to(dtype=torch.float))
self.fc3.qconfig = default_qconfig
self.sub2.qconfig = default_qconfig
custom_options = {
'dtype': torch.quint8,
'qscheme': torch.per_tensor_affine
}
custom_qconfig = QConfig(activation=default_observer.with_args(**custom_options),
weight=default_weight_observer)
self.sub2.fc1.qconfig = custom_qconfig
self.sub2.fc1 = QuantWrapper(self.sub2.fc1)
self.sub2.fc2 = QuantWrapper(self.sub2.fc2)
def test_linear_chain(self):
class LinearChain(nn.Module):
def __init__(self):
super(LinearChain, self).__init__()
self.linear1 = nn.Linear(3, 4)
self.linear2 = nn.Linear(4, 5)
self.linear3 = nn.Linear(5, 6)
def forward(self, x):
x = self.linear1(x)
x = self.linear2(x)
x = self.linear3(x)
return x
float_model = QuantWrapper(LinearChain())
img_data = [(torch.rand(10, 3, dtype=torch.float), torch.randint(0, 1, (2,), dtype=torch.long))
for _ in range(50)]
self.correct_artificial_bias_float(float_model, img_data)
self.correct_artificial_bias_quantize(float_model, img_data)
def test_conv_chain(self):
class ConvChain(nn.Module):
def __init__(self):
super(ConvChain, self).__init__()
self.conv2d1 = nn.Conv2d(3, 4, 5, 5)
self.conv2d2 = nn.Conv2d(4, 5, 5, 5)
self.conv2d3 = nn.Conv2d(5, 6, 5, 5)
def forward(self, x):
x = self.conv2d1(x)
x = self.conv2d2(x)
x = self.conv2d3(x)
return x
float_model = QuantWrapper(ConvChain())
img_data = [(torch.rand(10, 3, 125, 125, dtype=torch.float), torch.randint(0, 1, (2,), dtype=torch.long))
for _ in range(50)]
self.correct_artificial_bias_float(float_model, img_data)
self.correct_artificial_bias_quantize(float_model, img_data)
def test_resnet_base(self, qconfig):
r"""Test quantization for bottleneck topology used in resnet/resnext
and add coverage for conversion of average pool and float functional
"""
model = ResNetBase().float().eval()
model = QuantWrapper(model)
model.qconfig = qconfig
fuse_list = ['module.conv1', 'module.bn1', 'module.relu1']
fuse_modules(model, fuse_list, inplace=True)
model = prepare(model)
self.checkObservers(model)
test_only_eval_fn(model, self.img_data)
model = convert(model)
def checkQuantized(model):
self.assertEqual(type(model.module.conv1), nn._intrinsic.quantized.ConvReLU2d)
self.assertEqual(type(model.module.myop), nn.quantized.QFunctional)
self.assertEqual(type(model.module.avgpool), nn.AdaptiveAvgPool2d)
test_only_eval_fn(model, self.img_data)
checkQuantized(model)
def __init__(self, add_stub=False):
super().__init__()
self.hswish = QuantWrapper(nn.Hardswish())
def __init__(self):
super().__init__()
layers = [nn.ConvTranspose2d(3, 32, 3, bias=True)]
self.conv = nn.Sequential(*layers)
self.quant_wrap = QuantWrapper(self.conv)
def __init__(self):
super().__init__()
self.qconfig = default_qconfig
self.fc1 = QuantWrapper(torch.nn.Linear(5, 5).to(dtype=torch.float))
def __init__(self, qengine='fbgemm'):
super().__init__()
self.qconfig = torch.quantization.get_default_qconfig(qengine)
self.fc1 = QuantWrapper(torch.nn.Linear(5, 5).to(dtype=torch.float))
def __init__(self):
super().__init__()
self.fc1 = torch.nn.Linear(5, 8).to(dtype=torch.float)
self.fc2 = QuantWrapper(torch.nn.Linear(8, 5).to(dtype=torch.float))
self.fc2.qconfig = torch.quantization.get_default_qconfig("fbgemm")
def __init__(self):
super().__init__()
self.pool = QuantWrapper(nn.AdaptiveAvgPool2d((1, 1)))
def __init__(self):
super().__init__()
self.pool = QuantWrapper(nn.AvgPool2d(kernel_size=2))
def __init__(self):
super(AnnotatedSingleLayerLinearModel, self).__init__()
self.qconfig = default_qconfig
self.fc1 = QuantWrapper(torch.nn.Linear(5, 5).to(dtype=torch.float))
def __init__(self):
super(AnnotatedTwoLayerLinearModel, self).__init__()
self.fc1 = torch.nn.Linear(5, 8).to(dtype=torch.float)
self.fc2 = QuantWrapper(torch.nn.Linear(8, 5).to(dtype=torch.float))
self.fc2.qconfig = default_qconfig
def __init__(self):
super().__init__()
self.relu = QuantWrapper(nn.ReLU())
def __init__(self):
super().__init__()
self.relu = QuantWrapper(nn.ReLU())
self.quant = QuantStub()
self.dequant = DeQuantStub()
def __init__(self):
super().__init__()
self.leaky_relu = QuantWrapper(nn.LeakyReLU())