F.conv onnx export better support

[lucasjinreal]

Pls test this simple model export:

class MG(nn.Module):

    def __init__(self):
        super().__init__()

    def forward(self, x, b):
        preds = F.conv2d(x, b,
                             stride=1)
        return preds


torch_model = MG()
x = torch.randn([1, 4, 24, 24])
b = torch.randn([8, 4, 3, 3])
torch_out = torch_model(x, b)

# Export the model
torch.onnx.export(torch_model,               # model being run
                  (x, b),
                  "a.onnx",
                  export_params=True,        # store the trained parameter weights inside the model file
                  opset_version=12,          # the ONNX version to export the model to
                  do_constant_folding=True,
                  verbose=True)
print('Done!')

This is a dead simple model, but Pytorch can not export it make it convertable to trt.

When I convert to trt, it got:

❯ onnx2trt a.onnx 
----------------------------------------------------------------
Input filename:   a.onnx
ONNX IR version:  0.0.6
Opset version:    12
Producer name:    pytorch
Producer version: 1.7
Domain:           
Model version:    0
Doc string:       
----------------------------------------------------------------
Parsing model
While parsing node number 0 [Conv -> "2"]:
ERROR: /home/onnx-tensorrt/builtin_op_importers.cpp:512 In function importConv:
[8] Assertion failed: ctx->network()->hasExplicitPrecision() && "TensorRT only supports multi-input conv for explicit precision QAT networks!"

I don't sure is pytorch side problem or onnx-tensorrt side problem, But I can not convert any model which contains self-defined F.conv op for example, SOLOv2.

Please help me if anyone knows how to solve it.

[lucasjinreal]

I guess this issue is too decent that no one really got this step? @kevinch-nv Please have a look, pytorch code is included.

[lucasjinreal]

@mk-nvidia Pls have a look

pytorch/pytorch#54314

@KellenSunderland @kevinch-nv Pls have alook

[lucasjinreal]

@kevinch-nv @KellenSunderland Daily address

[kevinch-nv]

TRT requires the weights for the conv to be initializers, unless it is being overwritten by an INT8 -> Float dequantize layer in QDQ networks. General support of tensor conv weights are unsupported at the moment.

[kevinch-nv]

Closing as duplicate of #609