Skip to content

Latest commit

 

History

History
304 lines (252 loc) · 15.5 KB

pytorch2onnx.md

File metadata and controls

304 lines (252 loc) · 15.5 KB
Raw

Tutorial 8: Pytorch to ONNX (Experimental)

How to convert models from Pytorch to ONNX

Prerequisite

  1. Install the prerequisites following get_started.md/Prepare environment.
  2. Build custom operators for ONNX Runtime and install MMCV manually following How to build custom operators for ONNX Runtime
  3. Install MMdetection manually following steps 2-3 in get_started.md/Install MMdetection.

Usage

python tools/deployment/pytorch2onnx.py \
    ${CONFIG_FILE} \
    ${CHECKPOINT_FILE} \
    --output-file ${OUTPUT_FILE} \
    --input-img ${INPUT_IMAGE_PATH} \
    --shape ${IMAGE_SHAPE} \
    --test-img ${TEST_IMAGE_PATH} \
    --opset-version ${OPSET_VERSION} \
    --cfg-options ${CFG_OPTIONS}
    --dynamic-export \
    --show \
    --verify \
    --simplify \

Description of all arguments

  • config : The path of a model config file.
  • checkpoint : The path of a model checkpoint file.
  • --output-file: The path of output ONNX model. If not specified, it will be set to tmp.onnx.
  • --input-img: The path of an input image for tracing and conversion. By default, it will be set to tests/data/color.jpg.
  • --shape: The height and width of input tensor to the model. If not specified, it will be set to 800 1216.
  • --test-img : The path of an image to verify the exported ONNX model. By default, it will be set to None, meaning it will use --input-img for verification.
  • --opset-version : The opset version of ONNX. If not specified, it will be set to 11.
  • --dynamic-export: Determines whether to export ONNX model with dynamic input and output shapes. If not specified, it will be set to False.
  • --show: Determines whether to print the architecture of the exported model and whether to show detection outputs when --verify is set to True. If not specified, it will be set to False.
  • --verify: Determines whether to verify the correctness of an exported model. If not specified, it will be set to False.
  • --simplify: Determines whether to simplify the exported ONNX model. If not specified, it will be set to False.
  • --cfg-options: Override some settings in the used config file, the key-value pair in xxx=yyy format will be merged into config file.

Example:

python tools/deployment/pytorch2onnx.py \
    configs/yolo/yolov3_d53_mstrain-608_273e_coco.py \
    checkpoints/yolo/yolov3_d53_mstrain-608_273e_coco.pth \
    --output-file checkpoints/yolo/yolov3_d53_mstrain-608_273e_coco.onnx \
    --input-img demo/demo.jpg \
    --test-img tests/data/color.jpg \
    --shape 608 608 \
    --show \
    --verify \
    --dynamic-export \
    --cfg-options \
      model.test_cfg.deploy_nms_pre=-1 \

How to evaluate the exported models

We prepare a tool tools/deplopyment/test.py to evaluate ONNX models with ONNXRuntime and TensorRT.

Prerequisite

  • Install onnx and onnxruntime (CPU version)

    pip install onnx onnxruntime==1.5.1

  • If you want to run the model on GPU, please remove the CPU version before using the GPU version.

    pip uninstall onnxruntime
pip install onnxruntime-gpu

    Note: onnxruntime-gpu is version-dependent on CUDA and CUDNN, please ensure that your environment meets the requirements.

  • Build custom operators for ONNX Runtime following How to build custom operators for ONNX Runtime

  • Install TensorRT by referring to How to build TensorRT plugins in MMCV (optional)

Usage

python tools/deployment/test.py \
    ${CONFIG_FILE} \
    ${MODEL_FILE} \
    --out ${OUTPUT_FILE} \
    --backend ${BACKEND} \
    --format-only ${FORMAT_ONLY} \
    --eval ${EVALUATION_METRICS} \
    --show-dir ${SHOW_DIRECTORY} \
    ----show-score-thr ${SHOW_SCORE_THRESHOLD} \
    ----cfg-options ${CFG_OPTIONS} \
    ----eval-options ${EVALUATION_OPTIONS} \

Description of all arguments

  • config: The path of a model config file.
  • model: The path of an input model file.
  • --out: The path of output result file in pickle format.
  • --backend: Backend for input model to run and should be onnxruntime or tensorrt.
  • --format-only : Format the output results without perform evaluation. It is useful when you want to format the result to a specific format and submit it to the test server. If not specified, it will be set to False.
  • --eval: Evaluation metrics, which depends on the dataset, e.g., "bbox", "segm", "proposal" for COCO, and "mAP", "recall" for PASCAL VOC.
  • --show-dir: Directory where painted images will be saved
  • --show-score-thr: Score threshold. Default is set to 0.3.
  • --cfg-options: Override some settings in the used config file, the key-value pair in xxx=yyy format will be merged into config file.
  • --eval-options: Custom options for evaluation, the key-value pair in xxx=yyy format will be kwargs for dataset.evaluate() function

Notes:

  • If the deployed backend platform is TensorRT, please add environment variables before running the file:

    export ONNX_BACKEND=MMCVTensorRT

  • If you want to use the --dynamic-export parameter in the TensorRT backend to export ONNX, please remove the --simplify parameter, and vice versa.

Results and Models

Model Config Metric PyTorch ONNX Runtime TensorRT
FCOS configs/fcos/fcos_r50_caffe_fpn_gn-head_4x4_1x_coco.py Box AP 36.6 36.5 36.3
FSAF configs/fsaf/fsaf_r50_fpn_1x_coco.py Box AP 36.0 36.0 35.9
RetinaNet configs/retinanet/retinanet_r50_fpn_1x_coco.py Box AP 36.5 36.4 36.3
SSD configs/ssd/ssd300_coco.py Box AP 25.6 25.6 25.6
YOLOv3 configs/yolo/yolov3_d53_mstrain-608_273e_coco.py Box AP 33.5 33.5 33.5
Faster R-CNN configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py Box AP 37.4 37.4 37.0
Mask R-CNN configs/mask_rcnn/mask_rcnn_r50_fpn_1x_coco.py Box AP 38.2 38.1 37.7
Mask AP 34.7 33.7 33.3
CornerNet configs/cornernet/cornernet_hourglass104_mstest_10x5_210e_coco.py Box AP 40.6 40.4 -
DETR configs/detr/detr_r50_8x2_150e_coco.py Box AP 40.1 40.1 -
PointRend configs/point_rend/point_rend_r50_caffe_fpn_mstrain_1x_coco.py Box AP 38.4 38.4 -
Mask AP 36.3 35.2 -

Notes:

  • All ONNX models are evaluated with dynamic shape on coco dataset and images are preprocessed according to the original config file. Note that CornerNet is evaluated without test-time flip, since currently only single-scale evaluation is supported with ONNX Runtime.

  • Mask AP of Mask R-CNN drops by 1% for ONNXRuntime. The main reason is that the predicted masks are directly interpolated to original image in PyTorch, while they are at first interpolated to the preprocessed input image of the model and then to original image in other backend.

List of supported models exportable to ONNX

The table below lists the models that are guaranteed to be exportable to ONNX and runnable in ONNX Runtime.

Model Config Dynamic Shape Batch Inference Note
FCOS configs/fcos/fcos_r50_caffe_fpn_gn-head_4x4_1x_coco.py Y Y
FSAF configs/fsaf/fsaf_r50_fpn_1x_coco.py Y Y
RetinaNet configs/retinanet/retinanet_r50_fpn_1x_coco.py Y Y
SSD configs/ssd/ssd300_coco.py Y Y
YOLOv3 configs/yolo/yolov3_d53_mstrain-608_273e_coco.py Y Y
Faster R-CNN configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py Y Y
Mask R-CNN configs/mask_rcnn/mask_rcnn_r50_fpn_1x_coco.py Y Y
CornerNet configs/cornernet/cornernet_hourglass104_mstest_10x5_210e_coco.py Y N no flip, no batch inference, tested with torch==1.7.0 and onnxruntime==1.5.1.
DETR configs/detr/detr_r50_8x2_150e_coco.py Y Y batch inference is not recommended
PointRend configs/point_rend/point_rend_r50_caffe_fpn_mstrain_1x_coco.py Y Y

Notes:

  • Minimum required version of MMCV is 1.3.5

  • All models above are tested with Pytorch==1.6.0 and onnxruntime==1.5.1, except for CornerNet. For more details about the torch version when exporting CornerNet to ONNX, which involves mmcv::cummax, please refer to the Known Issues in mmcv.

  • Though supported, it is not recommended to use batch inference in onnxruntime for DETR, because there is huge performance gap between ONNX and torch model (e.g. 33.5 vs 39.9 mAP on COCO for onnxruntime and torch respectively, with a batch size 2). The main reason for the gap is that these is non-negligible effect on the predicted regressions during batch inference for ONNX, since the predicted coordinates is normalized by img_shape (without padding) and should be converted to absolute format, but img_shape is not dynamically traceable thus the padded img_shape_for_onnx is used.

  • Currently only single-scale evaluation is supported with ONNX Runtime, also mmcv::SoftNonMaxSuppression is only supported for single image by now.

The Parameters of Non-Maximum Suppression in ONNX Export

In the process of exporting the ONNX model, we set some parameters for the NMS op to control the number of output bounding boxes. The following will introduce the parameter setting of the NMS op in the supported models. You can set these parameters through --cfg-options.

  • nms_pre: The number of boxes before NMS. The default setting is 1000.

  • deploy_nms_pre: The number of boxes before NMS when exporting to ONNX model. The default setting is 0.

  • max_per_img: The number of boxes to be kept after NMS. The default setting is 100.

  • max_output_boxes_per_class: Maximum number of output boxes per class of NMS. The default setting is 200.

Reminders

  • When the input model has custom op such as RoIAlign and if you want to verify the exported ONNX model, you may have to build mmcv with ONNXRuntime from source.
  • mmcv.onnx.simplify feature is based on onnx-simplifier. If you want to try it, please refer to onnx in mmcv and onnxruntime op in mmcv for more information.
  • If you meet any problem with the listed models above, please create an issue and it would be taken care of soon. For models not included in the list, please try to dig a little deeper and debug a little bit more and hopefully solve them by yourself.
  • Because this feature is experimental and may change fast, please always try with the latest mmcv and mmdetecion.

FAQs

  • None