TensorRT 6.0 ONNX_Parser doesn't support the ONNX model exported by PyTorch 1.3.1

[RizhaoCai]

Description

TensorRT 6.0 ONNX Parser doesn't support the ONNX model exported by PyTorch 1.3.1.

The TensorRT ONNX Parser seems not well compatible with the new PyTorch version. 1.3 or 1.4.

I have a Jetson TX2 (Jetpack 4.3, TensorRT6) to deploy my model.
This is my workflow:
1) Run training on GTX 1080 with PyTorch 1.3.1
2) Export the model to an ONNX model with the ONNX exporter shipped with PyTorch 1.3.1.
3) Transfer the ONNX model to my TX2
4) Use the ONNX parser of TensorRT (Python API) to build my engine.

However, it will tell when building the engine that
[TensorRT] ERROR: Network must have at least one output

Although there are some issues related to this error, e.g. #319, #286, but they did not take the PyTorch version into account. So here I point it out. This issue is open for reminding those who have the same problem but don't know how to solve.

At the very beginning, I thought to change to opset 7 may help as it is mentioned at the TensorRT doc

In general, the newer version of the ONNX Parser is designed to be backward compatible up to opset 7

Then, I tried
torch.onnx.export(model, dummy_input, onnx_model_path, verbose=True, input_names=input_names, output_names=output_names, opset_version=7)

When I used PyTorch 1.3.1, the problem was still there, and the size of the exported model is 13,599 KB.
Interestingly, when I used PyTorch 1.2.0, the size was 13,986 KB, meaning that different versions PyTorch with the same versions opset doesn't guarantee you can export the same ONNX model.

Therefore. using PyTorch 1.2 may help you solve the problem (1.1 also)

Environment

TensorRT Version: 6.0.1
GPU Type: TX2
CUDA Version: 10.0
CUDNN Version: 7.6
Operating System + Version: ubuntu 18.04
Python Version (if applicable): 3.6
PyTorch Version (if applicable): 1.3.1

[rmccorm4]

Hi @RizhaoCai ,

PyTorch 1.3 + TensorRT 6 is a known incompatability. Please use PyTorch <= 1.2 for TensorRT 6, or upgrade to TensorRT 7 (when available for Jetson since you mentioned TX2).

[Akshaysharma29]

I am trying to convert onnx to tensorrt on jetson nano with pytorch 1.1 and tensorRT 6.
[TensorRT] ERROR: Network must have at least one output Traceback (most recent call last): File "onnx2trt.py", line 61, in <module> context = engine.create_execution_context() AttributeError: 'NoneType' object has no attribute 'create_execution_context'

[rmccorm4]

Hi @Akshaysharma29 ,

The ONNX parser failed to parse your model, you need to print out the errors to see why.

Something like this:

import sys
import tensorrt as trt

if __name__ == "__main__":
    # Building engine
    with trt.Builder(TRT_LOGGER) as builder, \
         builder.create_network() as network, \
         trt.OnnxParser(network, TRT_LOGGER) as parser:

        # Fill network atrributes with information by parsing model
        with open("model.onnx", "rb") as f:
            if not parser.parse(f.read()):
                print('ERROR: Failed to parse the ONNX file: {}'.format(args.onnx))
                for error in range(parser.num_errors):
                    print(parser.get_error(error))
                sys.exit(1)

[RizhaoCai]

I am trying to convert onnx to tensorrt on jetson nano with pytorch 1.1 and tensorRT 6.
[TensorRT] ERROR: Network must have at least one output Traceback (most recent call last): File "onnx2trt.py", line 61, in <module> context = engine.create_execution_context() AttributeError: 'NoneType' object has no attribute 'create_execution_context'

Hello, @Akshaysharma29
I had similar experiences.
You got 'NoneType' for the engine because TensorRT did not manage to parse your model.
Let's check the ERROR msg:
[TensorRT] ERROR: Network must have at least one output Traceback
In such case, firstly, you should make sure that your ONNX model is exported in a way that is compatible with TensorRT. You can refer to this example .

Sometimes, the error can be due to that the forward function of your model has dynamic operations, like tensor.view(-1).
An ONNX model exported by PyTorch is not necessarily to be compatible with the ONNX parser shipped with TensorRT.

[Akshaysharma29]

I have referred your link @RizhaoCai
I have also replace
x = torch.cat([o.view(o.size(0), -1, 4) for o in loc], 1)
with
x=torch.cat(
(
loc[0].view(loc[0].size(0), 4800, 4),
loc[1].view(loc[1].size(0), 800, 4),
loc[2].view(loc[2].size(0), 200, 4),
loc[3].view(loc[3].size(0),75, 4)
)
, 1)
to remove dynamic operation
but nothing work any suggestion?

@rmccorm4
error while tensorrt parsing is
ERROR: Failed to parse the ONNX file: faceDetector.onnx node 0 (importModel): INVALID_GRAPH: Assertion failed: tensors.count(input_name)

[rmccorm4]

For that error, I would try building this repo's 19.12 (TRT 6) branch from source, and then try parsing the model again. I don't really have experience building it on Jetson, but it seems to be documented in the README now for Jetpack users.

[Akshaysharma29]

Finally bug is solved Thanks for the help @RizhaoCai and @rmccorm4
Things I used
1) remove dynamic operations, like tensor.view(-1). as stated by @RizhaoCai
2) used torch=1.1 and tersorrt =6.0.1.10 as stated by @rmccorm4
3) All conversion is performed on GPU.