Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/pinto0309/yolact_edge_onnx_tensorrt_myriad
Provides a conversion flow for YOLACT_Edge to models compatible with ONNX, TensorRT, OpenVINO and Myriad (OAK). My own implementation of post-processing allows for e2e inference. Support for Multi-Class NonMaximumSuppression, CombinedNonMaxSuppression.
https://github.com/pinto0309/yolact_edge_onnx_tensorrt_myriad
myriad oak oak-d onnx openvino tensorrt yolact yolactedge
Last synced: 2 months ago
JSON representation
Provides a conversion flow for YOLACT_Edge to models compatible with ONNX, TensorRT, OpenVINO and Myriad (OAK). My own implementation of post-processing allows for e2e inference. Support for Multi-Class NonMaximumSuppression, CombinedNonMaxSuppression.
- Host: GitHub
- URL: https://github.com/pinto0309/yolact_edge_onnx_tensorrt_myriad
- Owner: PINTO0309
- License: mit
- Created: 2022-06-09T09:16:03.000Z (over 2 years ago)
- Default Branch: main
- Last Pushed: 2022-11-01T14:48:55.000Z (about 2 years ago)
- Last Synced: 2024-10-04T12:50:06.583Z (3 months ago)
- Topics: myriad, oak, oak-d, onnx, openvino, tensorrt, yolact, yolactedge
- Language: Python
- Homepage:
- Size: 25.5 MB
- Stars: 27
- Watchers: 3
- Forks: 3
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# yolact_edge_onnx_tensorrt_myriad
Provides a conversion flow for **`YOLACT_Edge`** to models compatible with ONNX, TensorRT, OpenVINO and Myriad (OAK). My own implementation of post-processing allows for e2e inference. Support for `Multi-Class NonMaximumSuppression`, `CombinedNonMaxSuppression`.
[](https://github.com/PINTO0309/yolact_edge_onnx_tensorrt_myriad/actions?query=workflow%3ACodeQL)
- [x] Replace `ReduceMax` and `ArgMax`.
- [x] Multi-Class NonMaximumSuppression, CombinedNonMaxSuppression for ONNX
- [x] Demo Code https://github.com/PINTO0309/PINTO_model_zoo/tree/main/085_Yolact_Edge/32_Multi-class_NonMaximumSuppression_compliant_version
- [x] Multi-Class NonMaximumSuppression, CombinedNonMaxSuppression ONNX sample
https://user-images.githubusercontent.com/33194443/180679579-8b8e8345-57d7-4286-9d5d-cc1a380654d9.mp4
# Official Repo
https://github.com/haotian-liu/yolact_edge
# Tools
1. https://github.com/PINTO0309/tflite2tensorflow
2. https://github.com/PINTO0309/simple-onnx-processing-tools
# Convert
See sequence below.
https://github.com/PINTO0309/yolact_edge_onnx_tensorrt_myriad/blob/main/convert_script.txt
# Conversion to TensorFlow, TensorFlow Lite
https://github.com/PINTO0309/onnx2tf
# Benchmark
## ONNX + TensorRT
```bssh
$ sit4onnx --input_onnx_file_path yolact_edge_mobilenetv2_550x550.onnx
INFO: file: yolact_edge_mobilenetv2_550x550.onnx
INFO: providers: ['TensorrtExecutionProvider', 'CPUExecutionProvider']
INFO: input_name.1: input shape: [1, 3, 550, 550] dtype: float32
INFO: test_loop_count: 10
INFO: total elapsed time: 44.979095458984375 ms
INFO: avg elapsed time per pred: 4.4979095458984375 ms
INFO: output_name.1: x1y1x2y2_score_class shape: [1, 0, 6] dtype: float32
INFO: output_name.2: final_masks shape: [0, 138, 138] dtype: float32
```
# How to change NMS parameters
https://github.com/PINTO0309/simple-onnx-processing-tools
```bash
### fnms_max_output_boxes_per_class
sam4onnx \
--op_name fnms_nonmaxsuppression11 \
--input_onnx_file_path yolact_edge_mobilenetv2_550x550.onnx \
--output_onnx_file_path yolact_edge_mobilenetv2_550x550.onnx \
--input_constants fnms_max_output_boxes_per_class int64 [10]
### iou_threshold
sam4onnx \
--op_name fnms_nonmaxsuppression11 \
--input_onnx_file_path yolact_edge_mobilenetv2_550x550.onnx \
--output_onnx_file_path yolact_edge_mobilenetv2_550x550.onnx \
--input_constants fnms_iou_threshold float32 [0.6]
### score_threshold
sam4onnx \
--op_name fnms_nonmaxsuppression11 \
--input_onnx_file_path yolact_edge_mobilenetv2_550x550.onnx \
--output_onnx_file_path yolact_edge_mobilenetv2_550x550.onnx \
--input_constants fnms_score_threshold float32 [0.7]
```
# Model Structure
- INPUTS:
- `input`: `float32 [1, 3, 550, 550]`
- OUTPUTS:
- `x1y1x2y2_score_class`: `float32 [1, N, 6]`
- `N` = The number of objects detected, filtered by NMS, and therefore less than 1600. max_output_boxes_per_class=20 x 80classes
- `6` = x1, y1, x2, y2, score, classid
- `final_masks`: `float32 [N, 138, 138]`
- `N` = The number of objects detected, filtered by NMS, and therefore less than 1600. max_output_boxes_per_class=20 x 80classes
# Acknowledgments
https://github.com/yujin6056/yolactedge-onnx-conversion