facebook/levit-256 on Ascend NPU

1. 简介

将 ModelScope 上的 facebook/levit-256 图片分类模型适配到华为昇腾 NPU (Ascend910B) 上运行。

pip install -r requirements.txt
python inference.py

推理结果 (NPU Top-5):

Top-1: sandbar, sand bar (0.6391)
Top-2: seashore, coast, seacoast, sea-coast (0.3176)
Top-3: lakeside, lakeshore (0.0055)
Top-4: geyser (0.0024)
Top-5: parachute, chute (0.0018)

日志保存在 logs/inference.log。

对单张测试图片进行 CPU 与 NPU 一致性验证：

CPU Top-1: sandbar, sand bar
NPU Top-1: sandbar, sand bar
CPU Top-5: sandbar, sand bar, seashore, coast, seacoast, sea-coast, lakeside, lakeshore, geyser, parachute, chute
NPU Top-5: sandbar, sand bar, seashore, coast, seacoast, sea-coast, lakeside, lakeshore, geyser, parachute, chute
Top-1 match: True
Top-5 match: True

测试条件：2 次预热 + 10 次正式测量，batch_size=1，单卡 NPU。

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

见 screenshots/self_verification.png。

使用 ModelScope snapshot_download 下载权重，禁止 HuggingFace 自动下载
推理使用 torch.npu.synchronize() 确保 NPU 计算完成
精度验证使用官方 Ascend 一致性指标：max_abs_error、mean_abs_error、relative_error、cosine_similarity

4. 精度验证

对单张测试图片进行 CPU 与 NPU 一致性验证：

CPU Top-1: sandbar, sand bar

NPU Top-1: sandbar, sand bar

CPU Top-5: sandbar, sand bar, seashore, coast, seacoast, sea-coast, lakeside, lakeshore, geyser, parachute, chute

NPU Top-5: sandbar, sand bar, seashore, coast, seacoast, sea-coast, lakeside, lakeshore, geyser, parachute, chute

Top-1 match: True

Top-5 match: True

指标

数值

Avg latency

Min latency

Max latency

P50 latency

P90 latency

P95 latency

Throughput

images/sec