def test_dpu_export_onnx_quant_conv_bias():
FEATURES = 7
IN_SIZE = (1, IN_CH, FEATURES, FEATURES)
KERNEL_SIZE = 3
class Model(torch.nn.Module):
def __init__(self):
super().__init__()
self.conv1 = QuantConv2d(
out_channels=OUT_CH,
in_channels=IN_CH,
kernel_size=KERNEL_SIZE,
bias=True,
weight_quant=Int8WeightPerTensorFixedPoint,
bias_quant=Int8BiasPerTensorFixedPointInternalScaling,
input_quant=Int8ActPerTensorFixedPoint,
output_quant=Int8ActPerTensorFixedPoint,
return_quant_tensor=False)
self.conv1.weight.data.uniform_(-0.01, 0.01)
def forward(self, x):
return self.conv1(x)
inp = gen_linspaced_data(reduce(mul, IN_SIZE), -1, 1).reshape(IN_SIZE)
model = Model()
model(torch.from_numpy(inp)) # accumulate scale factors
model.eval()
PyXIRManager.export(model, input_shape=IN_SIZE)
def test_standard_onnx_quant_linear_bias_export():
IN_SIZE = (IN_CH, IN_CH)
class Model(torch.nn.Module):
def __init__(self):
super().__init__()
self.linear = QuantLinear(
in_features=IN_CH,
out_features=OUT_CH,
bias=True,
weight_quant=Int8WeightPerTensorFixedPoint,
bias_quant=Int8BiasPerTensorFixedPointInternalScaling,
input_quant=Int8ActPerTensorFixedPoint,
output_quant=Int8ActPerTensorFixedPoint,
return_quant_tensor=False)
self.linear.weight.data.uniform_(-0.01, 0.01)
def forward(self, x):
return self.linear(x)
inp = gen_linspaced_data(reduce(mul, IN_SIZE)).reshape(IN_SIZE)
model = Model()
model(torch.from_numpy(inp)) # accumulate scale factors
model.eval()
PyXIRManager.export(model, input_shape=IN_SIZE)