Facebook RegNet-Y-0320(width=320)图像分类模型针对华为昇腾 NPU 推理的适配。
RegNetForImageClassification)RegNetY-0320 采用基于阶段的架构,具体包括:
| 指标 | 数值 |
|---|---|
| 平均延迟 | 6.44 毫秒 |
| 最小延迟 | 6.17 毫秒 |
| 最大延迟 | 6.73 毫秒 |
| 吞吐量 | 155.32 张/秒 |
| 设备 | Ascend910 NPU |
CPU 与 NPU 推理对比:
facebook-regnet-y-320-NPU/
├── model_files/ # Model weights and config (not in git)
│ ├── config.json
│ ├── pytorch_model.bin
│ └── preprocessor_config.json
├── assets/ # Test images
│ └── test.jpg
├── logs/ # Execution logs
│ ├── env_check.log
│ ├── inference.log
│ ├── accuracy.log
│ ├── benchmark.log
│ └── paths.txt
├── screenshots/ # Screenshots
├── model_utils.py # Model loading utilities
├── inference.py # NPU inference with Top-5 output
├── eval_accuracy.py # CPU vs NPU accuracy comparison
├── benchmark.py # Performance benchmarking
├── requirements.txt # Python dependencies
└── README.md # This filepip install -r requirements.txtpython inference.pypython eval_accuracy.pypython benchmark.pyfacebook/regnet-y-0320-sw。Apache 2.0