cubeai/cv_level1_protected_animals_classification 在昇腾 NPU 上的部署

1. 简介

本项目将 cubeai/cv_level1_protected_animals_classification 模型适配到昇腾 NPU（Ascend910B）上运行。

模型架构：ViTForImageClassification
任务类型：多分类图像分类（保护动物识别）
类别数：66
输入尺寸：224 x 224 RGB
框架：PyTorch + Transformers
权重来源：ModelScope snapshot_download 本地加载

2. 验证环境

项目	版本/信息
NPU 驱动	npu-smi 25.5.2
NPU 设备	Ascend910_9362
Python	3.11.14
PyTorch	(环境自带)
torch_npu	(环境自带)
transformers	4.50.0

NPU 健康状态：OK

3. 推理运行

pip install -r requirements.txt
python inference.py

推理结果 (NPU Top-5):

Top-1: Sousa chinensis (0.778753)
Top-2: Dugong dugon (0.115194)
Top-3: Tragopan satyra (0.006850)
Top-4: Grus antigone (0.006030)
Top-5: Bos mutus (0.003553)

完整推理日志 (logs/inference.log):

=== cubeai/cv_level1_protected_animals_classification NPU Inference ===
Model: cubeai/cv_level1_protected_animals_classification
Loader type: transformers
Input shape: torch.Size([1, 3, 224, 224])
Output shape: torch.Size([1, 66])
Num classes: 66

NPU Top-5 Predictions:
  Top-1: Sousa chinensis (0.778753)
  Top-2: Dugong dugon (0.115194)
  Top-3: Tragopan satyra (0.006850)
  Top-4: Grus antigone (0.006030)
  Top-5: Bos mutus (0.003553)

All class probabilities:
  Ailuropoda melanoleuca: 0.000981
  Alligator sinensis: 0.002298
  Arctictis binturong: 0.001136
  Axis porcinus: 0.000463
  Bos gaurus: 0.001325
  Bos mutus: 0.003553
  Budorcas taxicolor: 0.000563
  Camelus bactrianus: 0.000567
  Campanumoea javanica Blume: 0.002011
  Castor fiber: 0.000472
  Cervus eldii: 0.000629
  Cervus nippon: 0.000791
  Crossoptilon mantchuricum: 0.002289
  Dugong dugon: 0.115194
  Elaphurus davidianus: 0.001126
  Elephas maximus Linnaeus: 0.002705
  Equus ferus: 0.000931
  Equus hemionus Pallas: 0.001836
  Equus kiang: 0.001193
  Grus antigone: 0.006030
  Grus japonensis: 0.002083
  Grus leucogeranus: 0.001867
  Grus monacha: 0.000283
  Grus nigricollis: 0.000796
  Gulo gulo: 0.001674
  Haliaeetus albicilla: 0.000652
  Helarctos malayanus: 0.001917
  Hemitragus jemlahicus: 0.003282
  Hylobatidae: 0.003304
  Larus relictus: 0.000150
  Lophophorus: 0.001187
  Lophura swinhoii: 0.001823
  Macaca assamensis: 0.001466
  Macaca cyclopis Swinhoe: 0.001057
  Macaca leonina: 0.001711
  Martes zibellina: 0.001287
  Mergus squamatus: 0.000363
  Muntiacus crinifrons: 0.001932
  Naemorhedus baileyi: 0.001159
  Naemorhedus swinhoei: 0.001098
  Neofelis nebulosa: 0.001322
  Nipponia nippon: 0.000316
  Nycticebus bengalensis: 0.000543
  Otididae: 0.001318
  Panthera pardus: 0.001657
  Panthera tigris: 0.001534
  Panthera uncia: 0.001044
  Pantholops hodgsonii: 0.002549
  Pavo muticus: 0.002298
  Pelochelys cantorii: 0.002063
  Polyplectron: 0.002021
  Presbytis: 0.001150
  Procapra przewalskii: 0.002836
  Przewalskium albirostris: 0.001096
  Rhinopithecus: 0.001981
  Saiga tatarica: 0.001631
  Sousa chinensis: 0.778753
  Syrmaticus ellioti: 0.000336
  Syrmaticus humiae: 0.002766
  Syrmaticus mikado: 0.002361
  Testudo horsfieldii: 0.001168
  Tragopan blythii: 0.001831
  Tragopan caboti: 0.001524
  Tragopan melanocephalus: 0.003207
  Tragopan satyra: 0.006850
  Tragulus kanchil: 0.000663

4. 精度验证

对单张测试图片进行 CPU 与 NPU 一致性验证：

指标	数值
max_abs_error	0.021020
mean_abs_error	0.006207
relative_error	0.8088%
cosine_similarity	0.999981
threshold	1.0%
结果	PASS

CPU Top-1: Sousa chinensis
NPU Top-1: Sousa chinensis
Top-1 match: True

5. 性能参考

指标	数值
平均延迟	6.0498 ms
最小延迟	5.9850 ms
最大延迟	6.1195 ms
p50 延迟	6.0430 ms
p90 延迟	6.1055 ms
p95 延迟	6.1125 ms
图片/秒	165.29

测试配置: 预热 2 次 + 正式 10 次，单卡 NPU。

6. 精度评测说明

本项目包含单图 smoke consistency 验证，非官方完整验证集评测。详细指标见第 4 节。

7. 自验证截图

见 screenshots/self_verification.png。

8. 日志文件

日志	说明
`logs/inference.log`	NPU 推理输出
`logs/accuracy.log`	CPU-NPU 精度一致性
`logs/benchmark.log`	NPU 性能基准测试

9. 注意事项

权重通过 ModelScope snapshot_download 下载，不提交到仓库。
推理时使用 local_files_only=True 避免 HuggingFace 自动下载。
测试图片为随机网络图片，实际性能可能因输入不同而略有差异。

10. 标签

#NPU #Ascend #Ascend910 #ViT #ImageClassification #ProtectedAnimals

1. 简介

模型架构：ViTForImageClassification

任务类型：多分类图像分类（保护动物识别）

类别数：66

输入尺寸：224 x 224 RGB

框架：PyTorch + Transformers

权重来源：ModelScope snapshot_download 本地加载

项目

版本/信息

NPU 驱动

npu-smi 25.5.2

NPU 设备

Ascend910_9362

Python

3.11.14

PyTorch

(环境自带)

torch_npu

(环境自带)

transformers

4.50.0

3. 推理运行

pip install -r requirements.txt
python inference.py

推理结果 (NPU Top-5):

Top-1: Sousa chinensis (0.778753)

Top-2: Dugong dugon (0.115194)

Top-3: Tragopan satyra (0.006850)

Top-4: Grus antigone (0.006030)

Top-5: Bos mutus (0.003553)

完整推理日志 (logs/inference.log):

=== cubeai/cv_level1_protected_animals_classification NPU Inference ===
Model: cubeai/cv_level1_protected_animals_classification
Loader type: transformers
Input shape: torch.Size([1, 3, 224, 224])
Output shape: torch.Size([1, 66])
Num classes: 66

NPU Top-5 Predictions:
  Top-1: Sousa chinensis (0.778753)
  Top-2: Dugong dugon (0.115194)
  Top-3: Tragopan satyra (0.006850)
  Top-4: Grus antigone (0.006030)
  Top-5: Bos mutus (0.003553)

All class probabilities:
  Ailuropoda melanoleuca: 0.000981
  Alligator sinensis: 0.002298
  Arctictis binturong: 0.001136
  Axis porcinus: 0.000463
  Bos gaurus: 0.001325
  Bos mutus: 0.003553
  Budorcas taxicolor: 0.000563
  Camelus bactrianus: 0.000567
  Campanumoea javanica Blume: 0.002011
  Castor fiber: 0.000472
  Cervus eldii: 0.000629
  Cervus nippon: 0.000791
  Crossoptilon mantchuricum: 0.002289
  Dugong dugon: 0.115194
  Elaphurus davidianus: 0.001126
  Elephas maximus Linnaeus: 0.002705
  Equus ferus: 0.000931
  Equus hemionus Pallas: 0.001836
  Equus kiang: 0.001193
  Grus antigone: 0.006030
  Grus japonensis: 0.002083
  Grus leucogeranus: 0.001867
  Grus monacha: 0.000283
  Grus nigricollis: 0.000796
  Gulo gulo: 0.001674
  Haliaeetus albicilla: 0.000652
  Helarctos malayanus: 0.001917
  Hemitragus jemlahicus: 0.003282
  Hylobatidae: 0.003304
  Larus relictus: 0.000150
  Lophophorus: 0.001187
  Lophura swinhoii: 0.001823
  Macaca assamensis: 0.001466
  Macaca cyclopis Swinhoe: 0.001057
  Macaca leonina: 0.001711
  Martes zibellina: 0.001287
  Mergus squamatus: 0.000363
  Muntiacus crinifrons: 0.001932
  Naemorhedus baileyi: 0.001159
  Naemorhedus swinhoei: 0.001098
  Neofelis nebulosa: 0.001322
  Nipponia nippon: 0.000316
  Nycticebus bengalensis: 0.000543
  Otididae: 0.001318
  Panthera pardus: 0.001657
  Panthera tigris: 0.001534
  Panthera uncia: 0.001044
  Pantholops hodgsonii: 0.002549
  Pavo muticus: 0.002298
  Pelochelys cantorii: 0.002063
  Polyplectron: 0.002021
  Presbytis: 0.001150
  Procapra przewalskii: 0.002836
  Przewalskium albirostris: 0.001096
  Rhinopithecus: 0.001981
  Saiga tatarica: 0.001631
  Sousa chinensis: 0.778753
  Syrmaticus ellioti: 0.000336
  Syrmaticus humiae: 0.002766
  Syrmaticus mikado: 0.002361
  Testudo horsfieldii: 0.001168
  Tragopan blythii: 0.001831
  Tragopan caboti: 0.001524
  Tragopan melanocephalus: 0.003207
  Tragopan satyra: 0.006850
  Tragulus kanchil: 0.000663

指标

数值

max_abs_error

0.021020

mean_abs_error

0.006207

relative_error

0.8088%

cosine_similarity

0.999981

threshold

1.0%

结果

PASS

指标

数值

平均延迟

6.0498 ms

最小延迟

5.9850 ms

最大延迟

6.1195 ms

p50 延迟

6.0430 ms

p90 延迟

6.1055 ms

p95 延迟

6.1125 ms

图片/秒

165.29

日志

说明

logs/inference.log

NPU 推理输出

logs/accuracy.log

CPU-NPU 精度一致性

logs/benchmark.log

NPU 性能基准测试