When I converted nn.LocalResponseNorm(size=5, alpha=1e-4, beta=0.75, k=2.) to onnx using torch.onnx.export, lots of layers are generated, and consume time more in inference. The code and onnx model are shown below.
class AlexNetADDCStandard_LRN(nn.Module):
def __init__(self, num_classes: int = 1000, dropout: float = 0.5):
#def __init__(self, num_classes=1000):
super(AlexNetADDCStandardLRN, self).__init__()
self.features = nn.Sequential(
#nn.Conv2d( 3, 96, kernel_size=11, stride=4, padding=2),
nn.Conv2d( 3, 96, kernel_size=11, stride=4, padding=0),
nn.ReLU(inplace=True),
nn.LocalResponseNorm(size=5, alpha=1e-4, beta=0.75, k=2.), # add LRN
#nn.MaxPool2d(kernel_size=3, stride=2),
nn.Conv2d(96, 256, kernel_size=5, padding=2),
...
However, onnx suppotrs a LRN operation layer independently, and I found that below model is possble. Can torch.onnx.export generate just one LRN layer ??
