EfficientNet convert to tensorrt model failed.

[FroyoZzz]

Hi,all.
I am trying to convert a trained EfficientNet-b2 model into tensorrt format.When I use a tensor of shape (1, 3, 512, 512) ,it succeed and everything worked well.But when I try a tensor of shape (1, 3, 1024, 1024),it failed in converting onnx to trt. I got this error:

[E] [TRT] Parameter check failed at: Network.cpp::nvinfer1::Network::addPoolingNd::500, condition: allDimsGtEq(windowSize, 1) && volume(windowSize) < MAX_KERNEL_DIMS_PRODUCT
While parsing node number 10 [GlobalAveragePool]:
ERROR: builtin_op_importers.cpp:1175 In function importGlobalAveragePool:
[8] Assertion failed: layer_ptr

Such a high resolution is not supported yet?
I am using the Windows 10; Tensorrt 6.0.1; CUDA 10.1;and pytorch 1.2 version.
Thank you!

[githubgsq]

@FroyoZzz I met the similar error. Have you solved it please?

[basaltzhang]

I think this value is coded in TensorRT. Unfortunately, this part is not in open source directories.
One tricky method is double avg_pooling

        # Squeeze and Excitation
        if self.has_se:
            if self.training:
                x_squeezed = F.adaptive_avg_pool2d(x, 1)
            else:
                a = F.adaptive_avg_pool2d(x, 2)
                x_squeezed = F.adaptive_avg_pool2d(a, 1)

[DataXujing]

I met the similar error. Have you solved it please?

[ttyio]

Hello @FroyoZzz @githubgsq @basaltzhang @daquexian, Have you solved the issue?The error what you seen is a plausibility check for kernel size in TensorRT. We have plan to relax the check in future versions.
If you use TensorRT 7.*, the onnx operator GlobalAveragePool is implemented using IReduceLayer, so you will no longer hit this issue. Could you take a try? thanks!

[kevinch-nv]

Closing due to inactivity - if you still have issues with the latest TRT version feel free to open a new issue.