文档建议反馈控制台
首页
学习
活动
专区
圈层
工具
MCP广场
文章/答案/技术大牛
发布
社区首页 >问答首页 >将faster_rcnn_r50预培训转换为托管在Triton模型服务器中的ONNX

问将faster_rcnn_r50预培训转换为托管在Triton模型服务器中的ONNX
EN

Stack Overflow用户
提问于 2021-12-01 23:41:33
回答 1查看 267关注 0票数 0

我浏览了to检测文档，将py手电模型转换为onnx，这里是链接。

所有安装都是正确的，我使用的是onnxruntime==1.8.1，用于ONNX运行时MMCV_WITH_OPS的自定义运算符。

我使用configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py用于更快的rcnn 链接和使用R-5-ffn预训练的链接模型。

我使用它将预先训练的模型转换为onnx，并成功地保存了一个名为fasterrcnn.onnx的onnx文件。

代码语言：javascript
复制
python tools/deployment/pytorch2onnx.py \
    configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py \
    checkpoints/faster_rcnn/faster_rcnn_r50_fpn_mstrain_3x_coco_20210524_110822-e10bd31c.pth \
    --output-file checkpoints/faster_rcnn/fasterrcnn.onnx \
    --input-img demo/demo.jpg \
    --test-img tests/data/color.jpg \
    --shape 608 608 \
    --dynamic-export \
    --cfg-options \
      model.test_cfg.deploy_nms_pre=-1 \

我正在使用onnx文件在NVIDIA triton模型服务器中托管模型。

代码语言：javascript
复制
 fasterrcnn_model | 1       | READY

Triton的onnx模型摘要如下所示

代码语言：javascript
复制
{
    "name": "fasterrcnn_model",
    "platform": "onnxruntime_onnx",
    "backend": "onnxruntime",
    "version_policy": {
        "latest": {
            "num_versions": 1
        }
    },
    "max_batch_size": 1,
    "input": [
        {
            "name": "input",
            "data_type": "TYPE_FP32",
            "dims": [
                3,
                -1,
                -1
            ]
        }
    ],
    "output": [
        {
            "name": "labels",
            "data_type": "TYPE_INT64",
            "dims": [
                -1
            ]
        },
        {
            "name": "dets",
            "data_type": "TYPE_FP32",
            "dims": [
                -1,
                5
            ]
        }
    ],
    "batch_input": [],
    "batch_output": [],
    "optimization": {
        "priority": "PRIORITY_DEFAULT",
        "input_pinned_memory": {
            "enable": true
        },
        "output_pinned_memory": {
            "enable": true
        },
        "gather_kernel_buffer_threshold": 0,
        "eager_batching": false
    },
    "instance_group": [
        {
            "name": "fasterrcnn_model",
            "kind": "KIND_CPU",
            "count": 1,
            "gpus": [],
            "profile": []
        }
    ],
    "default_model_filename": "model.onnx",
    "cc_model_filenames": {},
    "metric_tags": {},
    "parameters": {},
    "model_warmup": []
}

摘要概述了输出有“标签”和"dets“的类别。

在向triton发送带有示例图像的推理请求后，我将得到以下响应。标签

代码语言：javascript
复制
[[ 0.  1.  2.  3.  4.  5.  6.  7.  8.  9. 10. 11. 12. 13. 14. 15. 16. 17.
  18. 19. 20. 21. 22. 23. 24. 25. 26. 27. 28. 29. 30. 31. 32. 33. 34. 35.
  36. 37. 38. 39. 40. 41. 42. 43. 44. 45. 46. 47. 48. 49. 50. 51. 52. 53.
  54. 55. 56. 57. 58. 59. 60. 61. 62. 63. 64. 65. 66. 67. 68. 69. 70. 71.
  72. 73. 74. 75. 76. 77. 78. 79.  0.  1.  2.  3.  4.  5.  6.  7.  8.  9.
  10. 11. 12. 13. 14. 15. 16. 17. 18. 19.]]

dets

代码语言：javascript
复制
[[[-1.0000e+00 -1.0000e+00 -1.0000e+00 -1.0000e+00  0.0000e+00]
  [-3.0000e+02 -3.0000e+02 -3.0000e+02 -3.0000e+02  0.0000e+00]
  [-5.9900e+02 -5.9900e+02 -5.9900e+02 -5.9900e+02  0.0000e+00]
  [-8.9800e+02 -8.9800e+02 -8.9800e+02 -8.9800e+02  0.0000e+00]
  [-1.1970e+03 -1.1970e+03 -1.1970e+03 -1.1970e+03  0.0000e+00]
  [-1.4960e+03 -1.4960e+03 -1.4960e+03 -1.4960e+03  0.0000e+00]
  [-1.7950e+03 -1.7950e+03 -1.7950e+03 -1.7950e+03  0.0000e+00]
  [-2.0940e+03 -2.0940e+03 -2.0940e+03 -2.0940e+03  0.0000e+00]
  [-2.3930e+03 -2.3930e+03 -2.3930e+03 -2.3930e+03  0.0000e+00]
  [-2.6920e+03 -2.6920e+03 -2.6920e+03 -2.6920e+03  0.0000e+00]
  [-2.9910e+03 -2.9910e+03 -2.9910e+03 -2.9910e+03  0.0000e+00]
  [-3.2900e+03 -3.2900e+03 -3.2900e+03 -3.2900e+03  0.0000e+00]
  [-3.5890e+03 -3.5890e+03 -3.5890e+03 -3.5890e+03  0.0000e+00]
  [-3.8880e+03 -3.8880e+03 -3.8880e+03 -3.8880e+03  0.0000e+00]
  [-4.1870e+03 -4.1870e+03 -4.1870e+03 -4.1870e+03  0.0000e+00]
  [-4.4860e+03 -4.4860e+03 -4.4860e+03 -4.4860e+03  0.0000e+00]
  [-4.7850e+03 -4.7850e+03 -4.7850e+03 -4.7850e+03  0.0000e+00]
  [-5.0840e+03 -5.0840e+03 -5.0840e+03 -5.0840e+03  0.0000e+00]
  [-5.3830e+03 -5.3830e+03 -5.3830e+03 -5.3830e+03  0.0000e+00]
  [-5.6820e+03 -5.6820e+03 -5.6820e+03 -5.6820e+03  0.0000e+00]
  [-5.9810e+03 -5.9810e+03 -5.9810e+03 -5.9810e+03  0.0000e+00]
  [-6.2800e+03 -6.2800e+03 -6.2800e+03 -6.2800e+03  0.0000e+00]
  [-6.5790e+03 -6.5790e+03 -6.5790e+03 -6.5790e+03  0.0000e+00]
  [-6.8780e+03 -6.8780e+03 -6.8780e+03 -6.8780e+03  0.0000e+00]
  [-7.1770e+03 -7.1770e+03 -7.1770e+03 -7.1770e+03  0.0000e+00]
  [-7.4760e+03 -7.4760e+03 -7.4760e+03 -7.4760e+03  0.0000e+00]
  [-7.7750e+03 -7.7750e+03 -7.7750e+03 -7.7750e+03  0.0000e+00]
  [-8.0740e+03 -8.0740e+03 -8.0740e+03 -8.0740e+03  0.0000e+00]
  [-8.3730e+03 -8.3730e+03 -8.3730e+03 -8.3730e+03  0.0000e+00]
  [-8.6720e+03 -8.6720e+03 -8.6720e+03 -8.6720e+03  0.0000e+00]
  [-8.9710e+03 -8.9710e+03 -8.9710e+03 -8.9710e+03  0.0000e+00]
  [-9.2700e+03 -9.2700e+03 -9.2700e+03 -9.2700e+03  0.0000e+00]
  [-9.5690e+03 -9.5690e+03 -9.5690e+03 -9.5690e+03  0.0000e+00]
  [-9.8680e+03 -9.8680e+03 -9.8680e+03 -9.8680e+03  0.0000e+00]
  [-1.0167e+04 -1.0167e+04 -1.0167e+04 -1.0167e+04  0.0000e+00]
  [-1.0466e+04 -1.0466e+04 -1.0466e+04 -1.0466e+04  0.0000e+00]
  [-1.0765e+04 -1.0765e+04 -1.0765e+04 -1.0765e+04  0.0000e+00]
  [-1.1064e+04 -1.1064e+04 -1.1064e+04 -1.1064e+04  0.0000e+00]
  [-1.1363e+04 -1.1363e+04 -1.1363e+04 -1.1363e+04  0.0000e+00]
  [-1.1662e+04 -1.1662e+04 -1.1662e+04 -1.1662e+04  0.0000e+00]
  [-1.1961e+04 -1.1961e+04 -1.1961e+04 -1.1961e+04  0.0000e+00]
  [-1.2260e+04 -1.2260e+04 -1.2260e+04 -1.2260e+04  0.0000e+00]
  [-1.2559e+04 -1.2559e+04 -1.2559e+04 -1.2559e+04  0.0000e+00]
  [-1.2858e+04 -1.2858e+04 -1.2858e+04 -1.2858e+04  0.0000e+00]
  [-1.3157e+04 -1.3157e+04 -1.3157e+04 -1.3157e+04  0.0000e+00]
  [-1.3456e+04 -1.3456e+04 -1.3456e+04 -1.3456e+04  0.0000e+00]
  [-1.3755e+04 -1.3755e+04 -1.3755e+04 -1.3755e+04  0.0000e+00]
  [-1.4054e+04 -1.4054e+04 -1.4054e+04 -1.4054e+04  0.0000e+00]
  [-1.4353e+04 -1.4353e+04 -1.4353e+04 -1.4353e+04  0.0000e+00]
  [-1.4652e+04 -1.4652e+04 -1.4652e+04 -1.4652e+04  0.0000e+00]
  [-1.4951e+04 -1.4951e+04 -1.4951e+04 -1.4951e+04  0.0000e+00]
  [-1.5250e+04 -1.5250e+04 -1.5250e+04 -1.5250e+04  0.0000e+00]
  [-1.5549e+04 -1.5549e+04 -1.5549e+04 -1.5549e+04  0.0000e+00]
  [-1.5848e+04 -1.5848e+04 -1.5848e+04 -1.5848e+04  0.0000e+00]
  [-1.6147e+04 -1.6147e+04 -1.6147e+04 -1.6147e+04  0.0000e+00]
  [-1.6446e+04 -1.6446e+04 -1.6446e+04 -1.6446e+04  0.0000e+00]
  [-1.6745e+04 -1.6745e+04 -1.6745e+04 -1.6745e+04  0.0000e+00]
  [-1.7044e+04 -1.7044e+04 -1.7044e+04 -1.7044e+04  0.0000e+00]
  [-1.7343e+04 -1.7343e+04 -1.7343e+04 -1.7343e+04  0.0000e+00]
  [-1.7642e+04 -1.7642e+04 -1.7642e+04 -1.7642e+04  0.0000e+00]
  [-1.7941e+04 -1.7941e+04 -1.7941e+04 -1.7941e+04  0.0000e+00]
  [-1.8240e+04 -1.8240e+04 -1.8240e+04 -1.8240e+04  0.0000e+00]
  [-1.8539e+04 -1.8539e+04 -1.8539e+04 -1.8539e+04  0.0000e+00]
  [-1.8838e+04 -1.8838e+04 -1.8838e+04 -1.8838e+04  0.0000e+00]
  [-1.9137e+04 -1.9137e+04 -1.9137e+04 -1.9137e+04  0.0000e+00]
  [-1.9436e+04 -1.9436e+04 -1.9436e+04 -1.9436e+04  0.0000e+00]
  [-1.9735e+04 -1.9735e+04 -1.9735e+04 -1.9735e+04  0.0000e+00]
  [-2.0034e+04 -2.0034e+04 -2.0034e+04 -2.0034e+04  0.0000e+00]
  [-2.0333e+04 -2.0333e+04 -2.0333e+04 -2.0333e+04  0.0000e+00]
  [-2.0632e+04 -2.0632e+04 -2.0632e+04 -2.0632e+04  0.0000e+00]
  [-2.0931e+04 -2.0931e+04 -2.0931e+04 -2.0931e+04  0.0000e+00]
  [-2.1230e+04 -2.1230e+04 -2.1230e+04 -2.1230e+04  0.0000e+00]
  [-2.1529e+04 -2.1529e+04 -2.1529e+04 -2.1529e+04  0.0000e+00]
  [-2.1828e+04 -2.1828e+04 -2.1828e+04 -2.1828e+04  0.0000e+00]
  [-2.2127e+04 -2.2127e+04 -2.2127e+04 -2.2127e+04  0.0000e+00]
  [-2.2426e+04 -2.2426e+04 -2.2426e+04 -2.2426e+04  0.0000e+00]
  [-2.2725e+04 -2.2725e+04 -2.2725e+04 -2.2725e+04  0.0000e+00]
  [-2.3024e+04 -2.3024e+04 -2.3024e+04 -2.3024e+04  0.0000e+00]
  [-2.3323e+04 -2.3323e+04 -2.3323e+04 -2.3323e+04  0.0000e+00]
  [-2.3622e+04 -2.3622e+04 -2.3622e+04 -2.3622e+04  0.0000e+00]
  [-1.0000e+00 -1.0000e+00 -1.0000e+00 -1.0000e+00  0.0000e+00]
  [-3.0000e+02 -3.0000e+02 -3.0000e+02 -3.0000e+02  0.0000e+00]
  [-5.9900e+02 -5.9900e+02 -5.9900e+02 -5.9900e+02  0.0000e+00]
  [-8.9800e+02 -8.9800e+02 -8.9800e+02 -8.9800e+02  0.0000e+00]
  [-1.1970e+03 -1.1970e+03 -1.1970e+03 -1.1970e+03  0.0000e+00]
  [-1.4960e+03 -1.4960e+03 -1.4960e+03 -1.4960e+03  0.0000e+00]
  [-1.7950e+03 -1.7950e+03 -1.7950e+03 -1.7950e+03  0.0000e+00]
  [-2.0940e+03 -2.0940e+03 -2.0940e+03 -2.0940e+03  0.0000e+00]
  [-2.3930e+03 -2.3930e+03 -2.3930e+03 -2.3930e+03  0.0000e+00]
  [-2.6920e+03 -2.6920e+03 -2.6920e+03 -2.6920e+03  0.0000e+00]
  [-2.9910e+03 -2.9910e+03 -2.9910e+03 -2.9910e+03  0.0000e+00]
  [-3.2900e+03 -3.2900e+03 -3.2900e+03 -3.2900e+03  0.0000e+00]
  [-3.5890e+03 -3.5890e+03 -3.5890e+03 -3.5890e+03  0.0000e+00]
  [-3.8880e+03 -3.8880e+03 -3.8880e+03 -3.8880e+03  0.0000e+00]
  [-4.1870e+03 -4.1870e+03 -4.1870e+03 -4.1870e+03  0.0000e+00]
  [-4.4860e+03 -4.4860e+03 -4.4860e+03 -4.4860e+03  0.0000e+00]
  [-4.7850e+03 -4.7850e+03 -4.7850e+03 -4.7850e+03  0.0000e+00]
  [-5.0840e+03 -5.0840e+03 -5.0840e+03 -5.0840e+03  0.0000e+00]
  [-5.3830e+03 -5.3830e+03 -5.3830e+03 -5.3830e+03  0.0000e+00]
  [-5.6820e+03 -5.6820e+03 -5.6820e+03 -5.6820e+03  0.0000e+00]]]

标签响应看起来像普通的COCO类(80)，但是我很难解码dets响应。它看起来像包围框坐标4和置信阈值1，使形状(1100，5)。知道dets类别应该代表什么吗？输出通常取决于模型本身，但我认为onnx转换正在将输出更改为标签和dets。

onnx
onnxruntime
tritonserver
nvidia
EN

回答 1

Stack Overflow用户

回答已采纳

发布于 2021-12-03 21:00:34

看一下转换脚本，似乎dets是一个框加分数的组合。

代码语言：javascript
复制
boxes (Tensor): The bounding boxes of shape [N, num_boxes, 4]
scores (Tensor): The detection scores of shape [N, num_boxes, num_classes]
tuple[Tensor, Tensor]: dets of shape [N, num_det, 5]
            and class labels of shape [N, num_det].
dets = torch.cat([boxes, scores], dim=2)

helper.py#L197

票数 0
EN
页面原文内容由Stack Overflow提供。腾讯云小微IT领域专用引擎提供翻译支持
原文链接：

https://stackoverflow.com/questions/70192619

复制
相关文章

相似问题

领券

腾讯云开发者

扫码关注腾讯云开发者

扫码关注腾讯云开发者

领取腾讯云代金券

热门产品

热门推荐

更多推荐

Copyright © 2013 - 2026 Tencent Cloud. All Rights Reserved. 腾讯云 版权所有 

深圳市腾讯计算机系统有限公司 ICP备案/许可证号：粤B2-20090059 粤公网安备44030502008569号

腾讯云计算（北京）有限责任公司 京ICP证150476号 |  京ICP备11018762号

问题归档专栏文章快讯文章归档关键词归档开发者手册归档开发者手册 Section 归档