Onnx-tensorrt: [8] No importer registered for op: PRelu

[deimsdeutsch]

ONNX IR version: 0.0.3
Opset version: 8
Producer name:
Producer version:
Domain:
Model version: 0

Doc string:

Parsing model
While parsing node number 6 [PRelu -> "relu0"]:
ERROR: /opt/github4/onnx-tensorrt/ModelImporter.cpp:142 In function importNode:
[8] No importer registered for op: PRelu

[deimsdeutsch]

Can we integrate this into the solution:

https://github.com/Goingqs/TensorRT-Prelu

[Jerry3062]

I come cross this problem too.no PRelu layer，How Can I solve this problem?

[mibrahimy]

@deimsdeutsch we'll need to change the implementation, since how mxnet performs PReLU is different from Caffe. Parameters for mxnet PReLU will be different from caffe's implementation of PReLU.

[veepersgit]

Same problem while converting model from onnx( model from pytorch ) to trt.

[coolbei]

same problem. with op ATen. actually it's a torch.max call.

[mibrahimy]

Try changing the implementation of PReLU in onnx, which currently supports scalar values.

[deimsdeutsch]

Eagerly waiting for the solution....

[leiyu1980]

Eagerly waiting for the solution....

[YaminSAIC]

@coolbei Hi, I have the same problem with ATen, do you have a solution now?

[jamessmith90]

@ibrahimsherdil

3 months and still waiting for the solution....

[BinhBa]

So I find a workaround for the PReLU problem if you stuck with this. You have to create an intermediate model which replace the PReLU with the original implementation of the PReLU:

PReLU = ReLU(x) - prelu_weight * ReLU(-x)

In that:

• prelu_weight is the weight you get from the PReLU module.

It will be a lot of work. So stay strong!!!

[hadeson]

class PseudoPReLu(Module):
    def __init__(self, num_parameters=1, init=0.25, inplace=True):
        self.num_parameters = num_parameters
        super(PseudoPReLu, self).__init__()
        self.weight = Parameter(torch.Tensor(num_parameters).fill_(init))
        self.inplace = inplace

    def forward(self, input):
        res = F.relu(input, inplace=self.inplace)
        min_res = F.relu(-input, inplace=self.inplace)
        weight_broadcast = self.weight.reshape(1, self.weight.shape[0], 1, 1)       
        return res - weight_broadcast * min_res

    def extra_repr(self):
        return 'num_parameters={}'.format(self.num_parameters)

I take @LovExtreme advice and reimplement PReLU with ReLU. Parsing with no problem.

[josh-gleason]

Extending @hadeson's version to work for inputs other than 2D data tensors.

class PseudoPReLu(Module):
    def __init__(self, num_parameters=1, init=0.25, inplace=True):
        self.num_parameters = num_parameters
        super(PseudoPReLu, self).__init__()
        self.weight = Parameter(torch.Tensor(num_parameters).fill_(init))
        self.inplace = inplace

    def forward(self, input):
        res = F.relu(input, inplace=self.inplace)
        min_res = F.relu(-input, inplace=self.inplace)
        weight_shape = [-1 if idx == 1 else 1 for idx in range(len(x.data.detach().shape))]
        weight_broadcast = self.weight.reshape(*weight_shape)
        return res - weight_broadcast * min_res

    def extra_repr(self):
        return 'num_parameters={}'.format(self.num_parameters)

[deimsdeutsch]

Go to hell Nvidia.

[ashuezy]

Parsing model
While parsing node number 6 [PRelu -> "relu0"]:
ERROR: /opt/github/onnx-tensorrt-5.1/ModelImporter.cpp:143 In function importNode:
[8] No importer registered for op: PRelu

9 Months and no support...

[seovchinnikov]

I was putting hacks like this to replace unsupported operations since 2018...
It's nothing compared to conversion from TF (UFF) back to 2018...

[kevinch-nv]

PRelu has been supported since TensorRT 6.0. Closing this issue.