tian
29d9ef5d0d
仓库结构: edge/ 设备端（原根目录设备端代码整体移入） control/ 管理端（清理后） docs/ 文档（PRD 移入 design/） README.md 根导航（新增） 清理: - control/.brainstorm 临时草稿删除 - control 根级重复文档(API表/PRD_04)并入 docs/design/ - control/plan.md -> docs/implementation/control-plan.md - control/safesightd-linux-arm64 二进制取消版本控制(.gitignore) - edge/transform 模型转换产物归入 models/，onnx/pt 大源文件取消跟踪(.gitignore) - Readme.md(PRD) -> docs/design/PRD_Product_v1.2.md（避开 README 大小写冲突） 更新: - 根 README.md 导航、docs/README.md 文档索引 - deployment.md/检查表路径加 edge/ 前缀 - .gitignore 重写（edge/control 分区规则）
31 lines
684 B
Python
31 lines
684 B
Python
#!/usr/bin/env python3
|
|
"""ONNX to RKNN converter for RetinaFace"""
|
|
|
|
from rknn.api import RKNN
|
|
|
|
# Create RKNN
|
|
rknn = RKNN(verbose=False)
|
|
|
|
# Config
|
|
rknn.config(
|
|
target_platform='rk3588',
|
|
mean_values=[[0, 0, 0]], # No mean subtraction
|
|
std_values=[[1, 1, 1]], # No normalization (already done in code)
|
|
quantized_dtype='w8a8', # INT8 quantization
|
|
)
|
|
|
|
# Load ONNX
|
|
print("Loading ONNX...")
|
|
rknn.load_onnx(model='RetinaFace_mobile320.onnx')
|
|
|
|
# Build
|
|
print("Building RKNN...")
|
|
rknn.build(do_quantization=True)
|
|
|
|
# Export
|
|
print("Exporting...")
|
|
rknn.export_rknn('RetinaFace_mobile320_new.rknn')
|
|
print("Done! Output: RetinaFace_mobile320_new.rknn")
|
|
|
|
rknn.release()
|