Kaloscope-2.0 (LSNet-XL-448) - Ascend NPU Adaptation

# 昇腾环境
CANN >= 8.5.1
torch_npu >= 2.9.0
torch >= 2.9.0

# 其他依赖
pip install torch torchvision pillow numpy timm

# 从 ModelScope 下载模型
pip install modelscope
modelscope download --model Heathcliff02/Kaloscope-2.0 --local_dir ./models/Kaloscope-2.0

python inference.py --image path/to/artwork.jpg --topk 5

Image: artwork.jpg
Inference time: 43.51 ms
Top-5 predictions:
  1. [12345] artist_name_1 (87.32%)
  2. [23456] artist_name_2 (5.14%)
  3. [34567] artist_name_3 (2.81%)
  4. [45678] artist_name_4 (1.63%)
  5. [56789] artist_name_5 (0.92%)

python inference.py --image_dir ./test_images --output results.csv

python inference.py --image artwork.jpg --device cpu

python inference.py --benchmark --iterations 100 --warmup 10

python eval.py  # 运行完整评测套件

from inference import LSNetArtist, load_class_mapping, preprocess_image, infer

# 加载模型（自动切换至 NPU）
model, cfg = load_model('best_checkpoint.pth', device='npu')

# 加载类别映射
class_mapping = load_class_mapping('class_mapping.csv')

# 预处理图像
tensor, pil_img = preprocess_image('artwork.jpg')

# 推理
indices, probs, elapsed = infer(model, tensor, device='npu', topk=5)

# 输出结果
for i, (idx, prob) in enumerate(zip(indices, probs)):
    artist = class_mapping.get(int(idx), f"class_{idx}")
    print(f"{i+1}. [{idx}] {artist} ({prob*100:.2f}%)")

kaloscope-npu/
├── inference.py                          # NPU 推理脚本
├── eval.py                               # 精度与性能评测脚本
├── README.md                             # 本文档
└── eval_output/                          # 评测输出目录
    ├── EVALUATION_REPORT.md              # 评测报告
    ├── evaluation.log                    # 运行日志
    ├── accuracy_results.csv              # 精度对比结果
    ├── performance_results.json          # 性能基准数据
    ├── precision_analysis.csv            # 100样本精度分析
    └── test_images/                      # 评测用测试图像

=========================================================
NPU vs CPU Accuracy Validation:
  Avg Relative Error:  0.0017%  ✅ (< 1%)
  Avg Cosine Similarity: 0.999994  ✅ (> 0.999)
  Top-1 Match Rate:    100.0%   ✅

NPU Performance (Ascend 910):
  Mean Latency:  43.51 ms
  Throughput:    23.3 FPS (bs=1), 139.7 FPS (bs=8)
  Memory:        6,925 MB
=========================================================

@misc{wang2025lsnetlargefocussmall,
      title={LSNet: See Large, Focus Small},
      author={Ao Wang and Hui Chen and Zijia Lin and Jungong Han and Guiguang Ding},
      year={2025},
      eprint={2503.23135},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.23135},
}