Error Code 4:Internal Error (/encoder/layers.10/Conv: number of kernel weights does not match tensor dimensions)

[Liupei1101]

onnx convert to trt model failed .
commend ./trtexec --onnx=/RAID5/user/liupei/ProPainter/E2FGVI/release_model/E2FGVI-HQ-CVPR22_LIUPEI.onnx --saveEngine=/RAID5/user/liupei/ProPainter/E2FGVI/release_model/E2FGVI-HQ-CVPR22_LIUPEI.trt --explicitBatch

/encoder/layers.10/Conv /encoder/layers.10/Conv:kernel weights has count 1474560 but 737280 was expected /encoder/layers.10/Conv: count of 1474560 weights in kernel, but kernel dimensions (3,3) with 320 input channels, 512 output channels and 2 groups were specified. Expected Weights count is 320 * 3*3 * 512 / 2 = 737280

[Liupei1101]

[zerollzeng]

1. does it work with onnx runtime?
2. I guess you are feeding incompatible input shapes.

[Liupei1101]

1. does it work with onnx runtime?
2. I guess you are feeding incompatible input shapes.

/encoder/layers.10/Conv /encoder/layers.10/Conv:kernel weights has count 1474560 but 737280 was expected /encoder/layers.10/Conv: count of 1474560 weights in kernel, but kernel dimensions (3,3) with 320 input channels, 512 output channels and 2 groups were specified. Expected Weights count is 320 * 3*3 * 512 / 2 = 737280

when static_axes onnx convert to tensorRT model, it works well, when convert dynamic_axes onnx model to tensorRT, the error oappear again. i do not know why?

[Liupei1101]

onnx convert to trt model failed . commend ./trtexec --onnx=/RAID5/user/liupei/ProPainter/E2FGVI/release_model/E2FGVI-HQ-CVPR22_LIUPEI.onnx --saveEngine=/RAID5/user/liupei/ProPainter/E2FGVI/release_model/E2FGVI-HQ-CVPR22_LIUPEI.trt --explicitBatch

/encoder/layers.10/Conv /encoder/layers.10/Conv:kernel weights has count 1474560 but 737280 was expected /encoder/layers.10/Conv: count of 1474560 weights in kernel, but kernel dimensions (3,3) with 320 input channels, 512 output channels and 2 groups were specified. Expected Weights count is 320 * 3*3 * 512 / 2 = 737280

if pth convert to onnx model without dynamic_axes, trtexec convert to tensorrt model works well, when with dynamic_axes , trtexec convert to tensorrt errot appear， i do not know why???

[zerollzeng]

Could you please share the onnx here? Maybe a bug.

[zerollzeng]

I quickly create a minimal reproduce but cannot repro:

$ ls
conv.onnx  make_conv.py
(base) zeroz @ ipp2-2325 2024/04/13-02:31:32 ~/scratch.zeroz_sw/github_bug/3720
$ cat make_conv.py
import torch

class TinyModel(torch.nn.Module):

    def __init__(self):
        super(TinyModel, self).__init__()
        self.conv = torch.nn.Conv2d(640, 512, (3,3), stride=1, padding=1, groups=2)

    def forward(self, x):
        x = self.conv(x)
        return x

dummy_input = torch.randn(1, 640, 14, 14, device="cpu")
model = TinyModel()

input_names = [ "input" ]
output_names = [ "output" ]

# dynamic_axes={'input' : {0 : 'bs'}, 'output' : {0 : 'bs'}}
# export(..., dynamic_axes = dynamic_axes)
torch.onnx.export(model, dummy_input, "conv.onnx", opset_version=13, verbose=True, input_names=input_names, output_names=output_names)

[Liupei1101]

i find the error,because

when group>1, reshape node -1 can not know the shape, when i convert to

it solved. thanks

[Liupei1101]

I quickly create a minimal reproduce but cannot repro:

$ ls
conv.onnx  make_conv.py
(base) zeroz @ ipp2-2325 2024/04/13-02:31:32 ~/scratch.zeroz_sw/github_bug/3720
$ cat make_conv.py
import torch

class TinyModel(torch.nn.Module):

    def __init__(self):
        super(TinyModel, self).__init__()
        self.conv = torch.nn.Conv2d(640, 512, (3,3), stride=1, padding=1, groups=2)

    def forward(self, x):
        x = self.conv(x)
        return x

dummy_input = torch.randn(1, 640, 14, 14, device="cpu")
model = TinyModel()

input_names = [ "input" ]
output_names = [ "output" ]

# dynamic_axes={'input' : {0 : 'bs'}, 'output' : {0 : 'bs'}}
# export(..., dynamic_axes = dynamic_axes)
torch.onnx.export(model, dummy_input, "conv.onnx", opset_version=13, verbose=True, input_names=input_names, output_names=output_names)

thanks,i solved.

[zerollzeng]

Ok, I'm closing this.