can you export onnx?

[you-old]

can you export onnx?

[biubug6]

can you export onnx?

It's easy. I'll provide it later .

[piotr-anyvision]

@biubug6 If you could do it, that would be amazing!

[SnowRipple]

@biubug6 Do you have any timetable for the onnx model please?

[biubug6]

@SnowRipple @wangpupanjing So sorry to have kept you waiting. Now I provide script "convert_to_onnx.py" to export onnx.

[SnowRipple]

I think I had the same error, the problem was with onnx not pytorch, just updated onnx (pytorch can have different onnx version) ;)

[SnowRipple]

Yes, but it was a while ago so I don't remember specifics - it is known problem with onnx, but there was simple fix - try google;)

[121649982]

can you tell me how to load .onnx model with tensor RT , thank you very much.

I load the model like this ：
std::string dataDirs = "E:/gc/Pytorch_Retinaface-master";
std::vectorstd::string dir;
dir.push_back(dataDirs);
auto parsed = parser->parseFromFile(
locateFile("FaceDetector.onnx", dir).c_str(), static_cast(gLogger.getReportableSeverity()));
if (!parsed)
{
return false;
}

but,I get this errors:

sorRT does not natively support INT64. Attempting to cast down to INT32.
[01/07/2020-19:19:10] [W] [TRT] onnx2trt_utils.cpp:198: Your ONNX model has been generated with INT64 weights, while TensorRT does not natively support INT64. Attempting to cast down to INT32.
[01/07/2020-19:19:10] [W] [TRT] onnx2trt_utils.cpp:198: Your ONNX model has been generated with INT64 weights, while TensorRT does not natively support INT64. Attempting to cast down to INT32.
[01/07/2020-19:19:10] [W] [TRT] onnx2trt_utils.cpp:198: Your ONNX model has been generated with INT64 weights, while TensorRT does not natively support INT64. Attempting to cast down to INT32.
[01/07/2020-19:19:10] [W] [TRT] onnx2trt_utils.cpp:198: Your ONNX model has been generated with INT64 weights, while TensorRT does not natively support INT64. Attempting to cast down to INT32.
[01/07/2020-19:19:10] [W] [TRT] onnx2trt_utils.cpp:198: Your ONNX model has been generated with INT64 weights, while TensorRT does not natively support INT64. Attempting to cast down to INT32.
[01/07/2020-19:19:11] [W] [TRT] onnx2trt_utils.cpp:198: Your ONNX model has been generated with INT64 weights, while TensorRT does not natively support INT64. Attempting to cast down to INT32.
[01/07/2020-19:19:11] [W] [TRT] onnx2trt_utils.cpp:198: Your ONNX model has been generated with INT64 weights, while TensorRT does not natively support INT64. Attempting to cast down to INT32.
While parsing node number 106 [Upsample]:
ERROR: builtin_op_importers.cpp:3240 In function importUpsample:
[8] Assertion failed: scales_input.is_weights()

[xsacha]

I think torch2trt is better than going pytorch -> onnx -> tensorrt

[zzxaijs]

[01/07 / 2020-19：19：11] [W] [TRT] onnx2trt_utils.cpp：198：您的ONNX模型是使用INT64权重生成的，而TensorRT本身不支持INT64。尝试转换为INT32。
解析节点号106 [Upsample]时：
错误：builtin_op_importers.cpp：3240在函数importUpsample中：
[8]断言失败：scales_input.is_weights（）,同样的这个错怎么解决

[denisvmedyantsev]

I faced the problem of the pytorch -> onnx -> tensorrt approach as above. I used simplifier, it helped, but I found a new problem: the output is different for different batch size for the trt engine. Fix the interpolation instead of using the simplifier in

Pytorch_Retinaface/models/net.py

Line 89 in b984b4b

up3 = F.interpolate(output3, size=[output2.size(2), output2.size(3)], mode="nearest")

and

Pytorch_Retinaface/models/net.py

Line 93 in b984b4b

up2 = F.interpolate(output2, size=[output1.size(2), output1.size(3)], mode="nearest")

It helped. Just cast size to int, eg. F.interpolate(output3, size=[output2.size(2), output2.size(3)], mode="nearest") => F.interpolate(output3, size=(int(output2.size(2)), int(output2.size(3))), mode="nearest")