markdown Atlas300实验:模型验证caffe_npu推理误差

Posted

tags:

篇首语:本文由小常识网(cha138.com)小编为大家整理,主要介绍了markdown Atlas300实验:模型验证caffe_npu推理误差相关的知识,希望对你有一定的参考价值。

注:
1. 图像预处理部分存在差异`ImagePreProcess_1使用DVPP进行resize`和`ImagePreProcessPillow_1使用OpenCV进行resize`
2. 模型差异:npu推理是模型进过AIPP进行融合预处理后的模型

|模型|余弦距离|最大绝对误差|相对欧氏距离|标准差|
|--|--|--|--|--|
|Pedestrian|0.9823884|0.027|0.18713343|(0.027,0.027)(0.026,0.028)|
|PedestrainGlobal|0.9898|0 - 0.239|||
| index | CompareFile1                                                 | CompareFile2                                                 | CosineSimilarity | MaxAbsoluteError | RelativeEuclideanDistance | StandardDeviation          |
| ----- | ------------------------------------------------------------ | ------------------------------------------------------------ | ---------------- | ---------------- | ------------------------- | -------------------------- |
| 1     | PedestrainGlobal\npu\davinci_000115_output_0_prob_1_0.float | PedestrainGlobal\npu_int8\davinci_000115_output_0_prob_1_0.float | 1                | 0.001            | 8.13E-04                  | (0.182,0.326)(0.182,0.326) |
| 2     | PedestrainGlobal\npu\davinci_000115_output_1_prob_2_0.float | PedestrainGlobal\npu_int8\davinci_000115_output_1_prob_2_0.float | 0.999969         | 0.007            | 0.011062                  | (0.091,0.214)(0.091,0.216) |
| 3     | PedestrainGlobal\npu\davinci_000115_output_2_prob_3_0.float | PedestrainGlobal\npu_int8\davinci_000115_output_2_prob_3_0.float | 1                | 0.001            | 0.001132                  | (0.091,0.284)(0.091,0.285) |
| 4     | PedestrainGlobal\npu\davinci_000115_output_3_prob_4_0.float | PedestrainGlobal\npu_int8\davinci_000115_output_3_prob_4_0.float | 1                | 0                | 1.29E-04                  | (0.333,0.469)(0.333,0.469) |
| 5     | PedestrainGlobal\npu\davinci_000115_output_4_prob_5_0.float | PedestrainGlobal\npu_int8\davinci_000115_output_4_prob_5_0.float | 0.999964         | 0.008            | 0.011475                  | (0.500,0.431)(0.500,0.423) |
| 6     | PedestrainGlobal\npu\davinci_000115_output_5_prob_6_0.float | PedestrainGlobal\npu_int8\davinci_000115_output_5_prob_6_0.float | 0.999386         | 0.024            | 0.041245                  | (0.500,0.276)(0.500,0.300) |
| 7     | PedestrainGlobal\npu\davinci_000115_output_6_prob_7_0.float | PedestrainGlobal\npu_int8\davinci_000115_output_6_prob_7_0.float | 1                | 0                | 1.10E-05                  | (0.333,0.471)(0.333,0.471) |
| 8     | PedestrainGlobal\npu\davinci_000115_output_7_prob_8_0.float | PedestrainGlobal\npu_int8\davinci_000115_output_7_prob_8_0.float | 0.999841         | 0.013            | 0.021463                  | (0.333,0.356)(0.333,0.364) |
| 9     | PedestrainGlobal\npu\davinci_000115_output_8_prob_9_0.float | PedestrainGlobal\npu_int8\davinci_000115_output_8_prob_9_0.float | 0.998428         | 0.041            | 0.069086                  | (0.333,0.349)(0.334,0.374) |
| 10    | PedestrainGlobal\npu\davinci_000115_output_9_prob_Kid_0.float | PedestrainGlobal\npu_int8\davinci_000115_output_9_prob_Kid_0.float | 1                | 0                | 0                         | (0.500,0.500)(0.500,0.500) |
| 11    | PedestrainGlobal\npu\davinci_000115_output_10_prob_plaid_lattice_0.float | PedestrainGlobal\npu_int8\davinci_000115_output_10_prob_plaid_lattice_0.float | 1                | 0                | 6.40E-06                  | (0.250,0.433)(0.250,0.433) |

| index | CompareFile1                                       | CompareFile2                                                 | CosineSimilarity | MaxAbsoluteError | RelativeEuclideanDistance | StandardDeviation          |
| ----- | -------------------------------------------------- | ------------------------------------------------------------ | ---------------- | ---------------- | ------------------------- | -------------------------- |
| 1     | PedestrainGlobal\caffe\davinci_000115_prob_1.float | PedestrainGlobal\npu_int8\davinci_000115_output_0_prob_1_0.float | 0.968041         | 0.239            | 0.253727                  | (0.182,0.358)(0.182,0.326) |
| 2     | PedestrainGlobal\caffe\davinci_000115_prob_2.float | PedestrainGlobal\npu_int8\davinci_000115_output_1_prob_2_0.float | 0.996503         | 0.095            | 0.128354                  | (0.091,0.244)(0.091,0.216) |
| 3     | PedestrainGlobal\caffe\davinci_000115_prob_3.float | PedestrainGlobal\npu_int8\davinci_000115_output_2_prob_3_0.float | 0.999993         | 0.004            | 0.005254                  | (0.091,0.286)(0.091,0.285) |
| 4     | PedestrainGlobal\caffe\davinci_000115_prob_4.float | PedestrainGlobal\npu_int8\davinci_000115_output_3_prob_4_0.float | 1                | 0.001            | 0.001475                  | (0.333,0.469)(0.333,0.469) |
| 5     | PedestrainGlobal\caffe\davinci_000115_prob_5.float | PedestrainGlobal\npu_int8\davinci_000115_output_4_prob_5_0.float | 0.99867          | 0.047            | 0.068067                  | (0.500,0.470)(0.500,0.423) |
| 6     | PedestrainGlobal\caffe\davinci_000115_prob_6.float | PedestrainGlobal\npu_int8\davinci_000115_output_5_prob_6_0.float | 0.980581         | 0.155            | 0.228325                  | (0.500,0.454)(0.500,0.300) |
| 7     | PedestrainGlobal\caffe\davinci_000115_prob_7.float | PedestrainGlobal\npu_int8\davinci_000115_output_6_prob_7_0.float | 1                | 0                | 2.43E-04                  | (0.333,0.471)(0.333,0.471) |
| 8     | PedestrainGlobal\caffe\davinci_000115_prob_8.float | PedestrainGlobal\npu_int8\davinci_000115_output_7_prob_8_0.float | 0.989415         | 0.12             | 0.176296                  | (0.333,0.443)(0.333,0.364) |
| 9     | PedestrainGlobal\caffe\davinci_000115_prob_9.float | PedestrainGlobal\npu_int8\davinci_000115_output_8_prob_9_0.float | 0.918076         | 0.252            | 0.490395                  | (0.333,0.248)(0.334,0.374) |
| 10    | PedestrainGlobal\caffe\davinci_000115_prob_a.float | PedestrainGlobal\npu_int8\davinci_000115_output_9_prob_Kid_0.float | 1                | 0                | 0                         | (0.500,0.500)(0.500,0.500) |
| 11    | PedestrainGlobal\caffe\davinci_000115_prob_b.float | PedestrainGlobal\npu_int8\davinci_000115_output_10_prob_plaid_lattice_0.float | 1                | 0                | 2.22E-04                  | (0.250,0.433)(0.250,0.433) |

以上是关于markdown Atlas300实验:模型验证caffe_npu推理误差的主要内容,如果未能解决你的问题,请参考以下文章

markdown [问题] Atlas300性能和精度误差实验

c_cpp Atlas300代码片段

极智AI | 多场景覆盖 看昇腾新推出的 Atlas300 系列新卡

工控机如何搭载华为ATLAS-300推理卡?

Atlas实现mysql主从分离

板卡AI昇腾atlas300 docker开发环境搭建