关于转onnx的问题

[BaoBaoJianqiang]

我看过您的convertor代码，已经成功转换为cpu版本，可以运行，大约11s一张图片。 为了进一步提升速度，我尝试转onnx，但是遇到了问题，还请指教，给出正确的转换方法，代码如下（写在TestModel.py的init方法的model.load_state_dict(d)之后）： import onnx import onnxruntime export_onnx_file = './net.onnx' torch.onnx.export(model, torch.randn(1,1,224,224,device='cuda'), export_onnx_file, verbose=False, input_names = ["inputs"]+["params_%d"%i for i in range(120)], output_names = ["outputs"], opset_version = 10 do_constant_folding = True, dynamic_axes = {"inputs":{0:"batch_size"}, 2:"h", 3:"w", "outputs":{0: "batch_size"}})

        net = onnx.load('./net.onnx') 
        onnx.checker.check_model(net) 
        onnx.helper.printable_graph(net.graph) 

[Zerohertz]

I think it's because tensor tracking is impossible in the upsampling process.

Try this!

fpem_v2.py

   def _upsample_add(self, x, y):
        # _, _, H, W = y.size()
        # return F.interpolate(x, size=(H, W), mode='bilinear') + y
        _, _, H, W = y.size()
        upsample = nn.Upsample(size=(H, W), mode='bilinear')#, align_corners=True)
        return upsample(x) + y

pan_pp.py

    def _upsample(self, x, size, scale=1):
        # _, _, H, W = size
        # return F.interpolate(x, size=(H // scale, W // scale), mode='bilinear')
        _, _, H, W = size
        upsample = nn.Upsample(size=(H // scale, W // scale), mode='bilinear')#, align_corners=True)
        return upsample(x)

export2onnx

    dynamic_axes = {
        'in': {
            0: 'batch',
            2: 'Width',
            3: 'Height'
        },
        'out': {
            0: 'batch',
            2: 'Height',
            3: 'Width'
        }
    }

    torch.onnx.export(
        model,
        inputData,
        "test.onnx",
        input_names=["in"],
        output_names=["out"],
        dynamic_axes=dynamic_axes,
    )

[BaoBaoJianqiang]

请问这样的改动，是否需要重新训练？然后再生成onnx？

[BaoBaoJianqiang]

您这个了的inputData，是我之前在代码中提供的值吗？

[BaoBaoJianqiang]

能否同时支持cpu和gpu？

[Zerohertz]

Please check this code!