0 Fork 0

Files

T

mckero f4f6ec7db5 feat(cw): ship the full fp32 DeepCW model instead of the int8 build

用户要求内置完整版模型, 不要量化版。

assets/deepcw/model.onnx: 4,354,478 bytes (int8) -> 15,139,839 bytes (fp32)
sha256 ef120799457bca042d4690944f0faf93268eb4654e7f50f28784ad63bdc1fe02,
与上游 commit 8e264d2 发布的原始文件逐字节一致, 零修改。

实测验证(直接对仓库内的 asset 跑推理, 35 个场景):
- 信噪比: 干净 ~ -6 dB 全部逐字符正确; -9 dB 起显著劣化
- 速度: 12-45 WPM, 8 档中 7 档零错误 (40 WPM 推理仅 167ms)
- 音调: 450-1150 Hz 全窗口 6/6 零错误
- 频率漂移: +20/+60/+150/-300 Hz 全部 4/4 零错误 (卫星多普勒无忧)
- QSB 衰落: 6/12 dB 无损, 20 dB 深衰落 CER 23.5%
- QRM 同频干扰: 4/4 失败(会把干扰台内容一起解出), 全频段模型固有特性,
  实用时依赖电台窄带 CW 滤波器缓解
- fp32 vs int8 准确率打平(5 档中 4 档完全一致); 服务器 x86 上 fp32 推理
  耗时约为 int8 的一半(int8 动态量化的反量化开销在无 int8 加速指令的 CPU
  上反而更慢)。手机 ARM 侧表现待装机确认。

NOTICE.md / DEEPCW.md / README.md 同步更新: 移除 int8 量化派生的记录与复现
步骤, 改为声明未修改照搬上游。

代价: APK 体积约 59MB -> 70MB, 运行内存峰值上升。此前真机闪退的根因是 R8
缺 -keep ai.onnxruntime.** 规则(已修), 与模型大小无关。

2026-08-13 14:48:56 +00:00

2.9 KiB

Raw Blame History

Third-Party Notices — CW Decode Module

DeepCW neural CW decoding model

The CW (Morse code) decoder in this module performs inference with a neural network model obtained from the DeepCW project.


Component	`src/main/assets/deepcw/model.onnx` and `model.onnx.json`
Upstream project	DeepCW / deepcw-engine
Source repository	https://github.com/e04/deepcw-engine
Author / copyright	e04
License	GNU Affero General Public License v3.0 only (AGPL-3.0-only)
License text	`DeepCW-AGPL-3.0.txt`
Obtained at commit	`8e264d243bbd4467bd19f3f28292219405b47e0e`
File size	15,139,839 bytes
SHA-256	`ef120799457bca042d4690944f0faf93268eb4654e7f50f28784ad63bdc1fe02`
Modifications	None. The full fp32 model is vendored byte-for-byte as published upstream.

Related upstream repositories by the same author (not vendored here):

https://github.com/e04/web-deep-cw-decoder — reference web application
https://github.com/e04/HamNoise — neural noise reduction (not used)

License compatibility

Look4Sat is licensed under the GNU General Public License v3.0 or later (GPL-3.0-or-later). The DeepCW model is licensed under AGPL-3.0-only.

Section 13 of the GPL version 3 expressly permits combining GPL-3.0 covered work with AGPL-3.0 covered work; the resulting combination may be conveyed, with the AGPL's additional network-interaction requirement applying to the AGPL-covered portion. Accordingly:

The Look4Sat source code remains under GPL-3.0-or-later.
The DeepCW model remains under AGPL-3.0-only.
Distributions of the combined application are accompanied by complete corresponding source, satisfying both licenses.

AGPL section 13 (network interaction)

Inference runs entirely on the local device via ONNX Runtime. The application does not offer the model's functionality to users interacting with it remotely over a network, so the additional network-source-offer requirement of AGPL-3.0 section 13 is not triggered by this usage. The complete corresponding source for both the application and the vendored model remains publicly available at the repository hosting this file.

Reproducing the vendored files

SHA=8e264d243bbd4467bd19f3f28292219405b47e0e
curl -sLO https://raw.githubusercontent.com/e04/deepcw-engine/$SHA/model.onnx
curl -sLO https://raw.githubusercontent.com/e04/deepcw-engine/$SHA/model.onnx.json
curl -sL -o DeepCW-AGPL-3.0.txt \
  https://raw.githubusercontent.com/e04/deepcw-engine/$SHA/LICENSE
sha256sum model.onnx
# expected: ef120799457bca042d4690944f0faf93268eb4654e7f50f28784ad63bdc1fe02
# copy model.onnx and model.onnx.json into assets/deepcw/ unchanged

ONNX Runtime

Inference engine: com.microsoft.onnxruntime:onnxruntime-android, licensed under the MIT License. Consumed as a published Maven artifact; not vendored in this repository.

2.9 KiB Raw Blame History