TensorRT engines for TensorFlow 2 Object Detection's Single-Shot Detector and Faster R-CNN models.
For best results, we recommend running these scripts on an environment with TensorRT >= 8.0.1 and TensorFlow 2.5.
Install TensorRT as per the TensorRT Install Guide. You will need to make sure the Python bindings for TensorRT are also installed correctly, these are available by installing the python3-libnvinfer and python3-libnvinfer-dev packages on your TensorRT download.
Install all dependencies listed in requirements.txt:
pip install -r requirements.txt
You will also need the latest onnx_graphsurgeon python module. If not already installed by TensorRT, you can install it manually by running:
pip install onnx-graphsurgeon==0.3.10 --index-url https://pypi.ngc.nvidia.com
NOTE: Please make sure that the onnx-graphsurgeon module installed by pip is version == 0.3.10.
Finally, you may want to clone the code from the AutoML Repository to use some helper utilities from it (this will be improved later):
git clone https://github.com/google/automl
The workflow to convert an EfficientDet model is basically TensorFlow → ONNX → TensorRT, and so parts of this process require TensorFlow to be installed. If you are performing this conversion to run inference on the edge, such as for NVIDIA Jetson devices, it might be easier to do the ONNX conversion on a PC first (NVIDIA Jetson not tested yet).
The starting point of conversion is a TensorFlow saved model. This can be exported from your own trained models, or you can download a pre-trained model.
You can download one of the pre-trained TFOD models from the TF2 Detection Model Zoo, such as:
wget http://download.tensorflow.org/models/object_detection/tf2/20200711/ssd_mobilenet_v2_320x320_coco17_tpu-8.tar.gz
When extracted:
tar -xvf ssd_mobilenet_v2_320x320_coco17_tpu-8.tar.gz
this package holds a directory named saved_model which holds the saved model ready for conversion.
However, if you are working with your own trained model, or if you need to re-export the saved model, you can do so from the training checkpoint. The downloaded package above also contains a pre-trained checkpoint. The structure is similar to this:
ssd_mobilenet_v2_320x320_coco17_tpu-8
├── checkpoint
│ ├── ckpt-0.data-00000-of-00001
│ └── ckpt-0.index
├── pipeline.config
└── saved_model
└── saved_model.pb
Mandatory! To (re-)export a saved model from here, clone and install the TFOD API from the TF Models Repository repository, and run:
cd /path/to/models/research/object_detection
python exporter_main_v2.py \
--input_type float_image_tensor \
--trained_checkpoint_dir /path/to/ssd_mobilenet_v2_320x320_coco17_tpu-8/checkpoint \
--pipeline_config_path /path/to/ssd_mobilenet_v2_320x320_coco17_tpu-8/pipeline.config \
--output_directory /path/to/export
Where --trained_checkpoint_dir and --pipeline_config_path point to the corresponding paths in the training checkpoint, --input_type must be set to float_image_tensor, otherwise TensorRT won't be able to convert uint8 type of input tensor. On the path pointed by --output_directory you will then find the newly created saved model in a directory aptly named saved_model.
Keep in mind that TFOD is no always honest about model resolutions, see truthful supported model resolution breakdown here:
| Model | Resolution |
|---|---|
| SSD MobileNet v2 320x320 | 300x300 |
| SSD MobileNet V1 FPN 640x640 | 640x640 |
| SSD MobileNet V2 FPNLite 320x320 | 320x320 |
| SSD MobileNet V2 FPNLite 640x640 | 640x640 |
| SSD ResNet50 V1 FPN 640x640 (RetinaNet50) | 640x640 |
| SSD ResNet50 V1 FPN 1024x1024 (RetinaNet50) | 1024x1024 |
| SSD ResNet101 V1 FPN 640x640 (RetinaNet101) | 640x640 |
| SSD ResNet101 V1 FPN 1024x1024 (RetinaNet101) | 1024x1024 |
| SSD ResNet152 V1 FPN 640x640 (RetinaNet152) | 640x640 |
| SSD ResNet152 V1 FPN 1024x1024 (RetinaNet152) | 1024x1024 |
| Faster R-CNN ResNet50 V1 640x640 | 640x640 |
| Faster R-CNN ResNet50 V1 1024x1024 | 1024x1024 |
| Faster R-CNN ResNet101 V1 640x640 | 640x640 |
| Faster R-CNN ResNet101 V1 1024x1024 | 1024x1024 |
| Faster R-CNN ResNet152 V1 640x640 | 640x640 |
| Faster R-CNN ResNet152 V1 1024x1024 | 1024x1024 |
| Faster R-CNN Inception ResNet V2 640x640 | 640x640 |
If TF saved model ready to be converted (i.e. you ran exporter_main_v2.py script), run:
python create_onnx.py \
--pipeline_config /path/to/exported/pipeline.config \
--saved_model /path/to/exported/saved_model \
--onnx /path/to/save/model.onnx
This will create the file model.onnx which is ready to convert to TensorRT.
The script has a few optional arguments, including:
--nms_threshold [...]allows overriding the default NMS score threshold parameter, as the runtime latency of the NMS plugin is sensitive to this value. It's a good practice to set this value as high as possible, while still fulfilling your application requirements, to reduce inference latency (not fully working yet. do not use).--batch_sizeallows selection of various batch sizes, default is 1.--debugallows to add an extra output to debug a particular node.--tf2onnxallows to save an intermediate ONNX graph generated by t2onnx.
Optionally, you may wish to visualize the resulting ONNX graph with a tool such as Netron.
The input to the graph is a float32 tensor with the selected input shape, containing RGB pixel data in the range of 0 to 255. All preprocessing will be performed inside the Model graph, so it is not required to further pre-process the input data.
The outputs of the graph are the same as the outputs of the EfficientNMS plugin.
It is possible to build the TensorRT engine directly with trtexec using the ONNX graph generated in the previous step. However, the script build_engine.py is provided for convenience, as it has been tailored to Model engine building and calibration. Run python build_engine.py --help for details on available settings.
To build the TensorRT engine file with FP16 precision, run:
python build_engine.py \
--onnx /path/to/saved/model.onnx \
--engine /path/to/save/engine.trt \
--precision fp16
The file engine.trt will be created, which can now be used to infer with TensorRT.
For best results, make sure no other processes are using the GPU during engine build, as it may affect the optimal tactic selection process.
To build and calibrate an engine for INT8 precision, run:
python build_engine.py \
--onnx /path/to/model.onnx \
--engine /path/to/engine.trt \
--precision int8 \
--calib_input /path/to/calibration/images \
--calib_cache /path/to/calibration.cache
Where --calib_input points to a directory with several thousands of images. For example, this could be a subset of the training or validation datasets that were used for the model. It's important that this data represents the runtime data distribution relatively well, therefore, the more images that are used for calibration, the better accuracy that will be achieved in INT8 precision. For models trained for the COCO dataset, we have found that 5,000 images gives a good result.
The --calib_cache controls where the calibration cache file will be written to. This is useful to keep a cached copy of the calibration results. Next time you need to build the engine for the same network, if this file exists, it will skip the calibration step and use the cached values instead.
Optionally, you can obtain execution timing information for the built engine by using the trtexec utility, as:
trtexec \
--loadEngine=/path/to/engine.trt \
--useCudaGraph --noDataTransfers \
--iterations=100 --avgRuns=100
If it's not already in your $PATH, the trtexec binary is usually found in /usr/src/tensorrt/bin/trtexec, depending on your TensorRT installation method.
An inference benchmark will run, with GPU Compute latency times printed out to the console. Depending on your environment, you should see something similar to:
GPU Compute Time: min = 1.55835 ms, max = 1.91591 ms, mean = 1.58719 ms, median = 1.578 ms, percentile(99%) = 1.90668 ms
For optimal performance, inference should be done in a C++ application that takes advantage of CUDA Graphs to launch the inference request. Alternatively, the TensorRT engine built with this process can also be executed through either Triton Inference Server or DeepStream SDK.
However, for convenience, a python inference script is provided here for quick testing of the built TensorRT engine.
To perform object detection on a set of images with TensorRT, run:
python infer.py \
--engine /paht/to/saved/engine.trt \
--input /path/to/images \
--output /path/to/output \
--preprocessor fixed_shape_resizer
Where the input path can be either a single image file, or a directory of jpg/png/bmp images. Argument --preprocessor corresponds with image preprocessor set in your pipeline.config file, usually it is under image_resizer section, only two are now supported, namely fixed_shape_resizer and keep_aspect_ratio_resizer.
The detection results will be written out to the specified output directory, consisting of a visualization image, and a tab-separated results file for each input image processed.
Given a validation dataset (such as COCO val2017 data) and ground truth annotations (such as COCO instances_val2017.json), you can get the mAP metrics for the built TensorRT engine. This will use the mAP metrics calculation script from the AutoML repository.
python eval_coco.py \
--engine /path/to/engine.trt \
--input /path/to/coco/val2017 \
--annotations /path/to/coco/annotations/instances_val2017.json \
--automl_path /path/to/automl \
--preprocessor fixed_shape_resizer
Where the --automl_path argument points to the root of the AutoML repository and --preprocessor corresponds with image preprocessor set in your pipeline.config file.
The mAP metric is sensitive to the NMS score threshold used, as using a high threshold will reduce the model recall, resulting in a lower mAP value. Ideally, mAP should be measured with a threshold of 0, but such a low value will impact the runtime latency of the EfficientNMS plugin. It may be a good idea to build separate TensorRT engines for different purposes. That is, one engine with a low threshold (like 0) dedicated for mAP validation, and another engine with your application specific threshold (like 0.4) for deployment. This is why we keep the NMS threshold as a configurable parameter in the create_onnx.py script.
To compare how the TensorRT detections match the original TensorFlow model results, you can run:
python compare_tf.py \
--engine /path/to/saved/engine.trt \
--saved_model /path/to/exported/saved_model \
--input /path/to/images \
--output /path/to/output \
--preprocessor fixed_shape_resizer
This script will process the images found in the given input path through both TensorFlow and TensorRT using the corresponding saved model and engine. It will then write to the output path a set of visualization images showing the inference results of both frameworks for visual qualitative comparison. Argument --preprocessor corresponds with image preprocessor set in your pipeline.config file.
If you run this on COCO val2017 images, you may also add the parameter --annotations /path/to/coco/annotations/instances_val2017.json to further compare against COCO ground truth annotations.