Parses ONNX models for execution with TensorRT.
See also the TensorRT documentation.
For the list of recent changes, see the changelog.
For a list of commonly seen issues and questions, see the FAQ.
For business inquiries, please contact [email protected]
For press and other inquiries, please contact Hector Marinez at [email protected]
Development on the Master branch is for the latest version of TensorRT 8.2.3.0 with full-dimensions and dynamic shape support.
For previous versions of TensorRT, refer to their respective branches.
Building INetwork objects in full dimensions mode with dynamic shape support requires calling the following API:
C++
const auto explicitBatch = 1U << static_cast<uint32_t>(nvinfer1::NetworkDefinitionCreationFlag::kEXPLICIT_BATCH);
builder->createNetworkV2(explicitBatch)
Python
import tensorrt
explicit_batch = 1 << (int)(tensorrt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)
builder.create_network(explicit_batch)
For examples of usage of these APIs see:
Current supported ONNX operators are found in the operator support matrix.
For building within docker, we recommend using and setting up the docker containers as instructed in the main (TensorRT repository)[https://github.com/NVIDIA/TensorRT#setting-up-the-build-environment] to build the onnx-tensorrt library.
Once you have cloned the repository, you can build the parser libraries and executables by running:
cd onnx-tensorrt
mkdir build && cd build
cmake .. -DTENSORRT_ROOT=<path_to_trt> && make -j
// Ensure that you update your LD_LIBRARY_PATH to pick up the location of the newly built library:
export LD_LIBRARY_PATH=$PWD:$LD_LIBRARY_PATH
Note that this project has a dependency on CUDA. By default the build will look in /usr/local/cuda for the CUDA toolkit installation. If your CUDA path is different, overwrite the default path by providing -DCUDA_TOOLKIT_ROOT_DIR=<path_to_cuda_install> in the CMake command.
For building only the libraries, append -DBUILD_LIBRARY_ONLY=1 to the CMake build command.
All experimental operators will be considered unsupported by the ONNX-TRT's supportsModel() function.
NonMaxSuppression is available as an experimental operator in TensorRT 8. It has the limitation that the output shape is always padded to length [max_output_boxes_per_class, 3], therefore some post processing is required to extract the valid indices.
ONNX models can be converted to serialized TensorRT engines using the onnx2trt executable:
onnx2trt my_model.onnx -o my_engine.trt
ONNX models can also be converted to human-readable text:
onnx2trt my_model.onnx -t my_model.onnx.txt
ONNX models can also be optimized by ONNX's optimization libraries (added by dsandler).
To optimize an ONNX model and output a new one use -m to specify the output model name and -O to specify a semicolon-separated list of optimization passes to apply:
onnx2trt my_model.onnx -O "pass_1;pass_2;pass_3" -m my_model_optimized.onnx
See more all available optimization passes by running:
onnx2trt -p
See more usage information by running:
onnx2trt -h
Python bindings for the ONNX-TensorRT parser are packaged in the shipped .whl files. Install them with
python3 -m pip install <tensorrt_install_dir>/python/tensorrt-8.x.x.x-cp<python_ver>-none-linux_x86_64.whl
TensorRT 8.2.1.8 supports ONNX release 1.8.0. Install it with:
python3 -m pip install onnx==1.8.0
The ONNX-TensorRT backend can be installed by running:
python3 setup.py install
The TensorRT backend for ONNX can be used in Python as follows:
import onnx
import onnx_tensorrt.backend as backend
import numpy as np
model = onnx.load("/path/to/model.onnx")
engine = backend.prepare(model, device='CUDA:1')
input_data = np.random.random(size=(32, 3, 224, 224)).astype(np.float32)
output_data = engine.run(input_data)[0]
print(output_data)
print(output_data.shape)The model parser library, libnvonnxparser.so, has its C++ API declared in this header:
NvOnnxParser.h
After installation (or inside the Docker container), ONNX backend tests can be run as follows:
Real model tests only:
python onnx_backend_test.py OnnxBackendRealModelTest
All tests:
python onnx_backend_test.py
You can use -v flag to make output more verbose.
Pre-trained models in ONNX format can be found at the ONNX Model Zoo