编译模型的时候出现error: failed to legalize operation 'MGB.ConvBias'

[BobbyXue666]

dump代码如下 @jit.trace(symbolic=True, capture_as_const=True) def fun(data, *, net): pred = net(data) return pred

fun(data, net=model) fun.dump("test.mge", arg_names=["data"],optimize_for_inference=False, enable_fuse_conv_bias_nonlinearity=False, enable_fuse_conv_bias_with_z=False)

[chenqy4933]

麻烦提供一下MegEngine的版本，以及最小复现，就目前的信息很难定位问题。

[BobbyXue666]

MegEngine 1.12.1+cu114 megfile 2.0.0.post2 最小复现如下：

class TestModel(M.Module):

    def __init__(self,):

        super().__init__()
        self.conv1 = M.Conv2d(2,16,kernel_size=3, stride=2, padding=1, bias=False)
        self.bn1 = M.BatchNorm2d(16)
        self.relu1 = M.ReLU()
        self.conv2 = M.Conv2d(16,4,kernel_size=3, stride=2, padding=1, bias=False)
        self.bn2 = M.BatchNorm2d(4)
        self.relu2 = M.ReLU()
        self.conv3 = M.Conv2d(4,1,kernel_size = 5, stride=1, padding=0, bias=False )
        self.ac3 = M.Sigmoid()

    def forward(self, x):

        out = self.relu1(self.bn1(self.conv1(x)))
        out = self.relu2(self.bn2(self.conv2(out)))
        out = self.ac3(self.conv3(out))

        return out

model = TestModel()
model.eval()
data = tensor(np.random.random([1, 2, 20, 20]).astype(np.float32))
print(model(data))

@jit.trace(symbolic=True, capture_as_const=True)
def fun(data, *, net):
    pred = net(data)
    return pred

fun(data, net=model)
fun.dump("test.mge", arg_names=["data"],optimize_for_inference=False, enable_fuse_conv_bias_nonlinearity=False, enable_fuse_conv_bias_with_z=False)

[chenqy4933]

用你上面的代码和对应的MegEngine的版本，我这边测试没有问题的。

[BobbyXue666]

这是我的json文件

{
    "dump_dir":"./batch_dump/",
    "models":[
        {
            "model_name":"test",
            "model_path":"./example/test.mge",
            "input_shape_str":"data=(1,2,20,20)",
            "enable_nchw44":true
        }
    ],
    "cv":{
        "transpose":["ui8"],
        "roicopy":["ui8"],
        "rotate":["ui8"],
        "flip":["ui8"],
        "resize_linear":["ui8"],
        "warp_affine_replicate_linear":["ui8"],
        "rgb2bgr":["ui8"],
        "yuv2bgr_nv21":["ui8"],
        "rgb2yuv":["ui8"]
    }
}

这是生成模型的命令 ./script/ppl_gen.sh ./bin/mgb-to-tinynn ./example/demo.json test --arm64

[BobbyXue666]

将bias去掉可以编过，想知道是为什么 self.conv1 = M.Conv2d(2,16,kernel_size=3, stride=2, padding=1)

[haolongzhangm]

@BobbyXue666 optimize_for_inference=False 这个写成 true 试试