RYDE-WORK/lnp_ml
1
0
Fork 0
mirror of https://github.com/RYDE-WORK/lnp_ml.git synced 2026-07-22 05:49:55 +08:00
Code Issues Packages Projects Releases Wiki Activity
73 Commits 3 Branches 0 Tags
Commit Graph

73 Commits

This Branch
This Branch
All Branches
Author SHA1 Message Date
Michelle0574
22e83122e7 feat: Qwen SoftRAG（RAG prompt/soft token/QLoRA）+ Morgan 检索旁路与实验结果 2026-07-07 17:47:34 +00:00
Michelle0474
0fc72a9474 Add Qwen2.5-7B MolRAG encoder branch + dual-seed results 
- llm_prompt.py: RAG prompt encoding (use_rag), Qwen support, set_retrieval_pool (leak-safe, train-only pool)
- models.py: use_rag/rag_top_k passthrough to LLMPromptEncoder
- nested_cv_optuna.py: --use-rag CLI, per-fold retrieval pool (防泄漏), slim checkpoint (skip frozen LLM weights)
- Qwen+RAG results: seed42 delivery R2=0.2151, seed123=0.1558 (mean 0.185), best LLM encoder so far
 2026-06-28 20:42:27 +08:00
Michelle0474
c0aa30f2bb Add 6-variant ablation results (baseline/moe/llm_molt5/llm_biot5/both_molt5/both_biot5) 
- Updated ablation_summary.csv with full 6-variant comparison
- Updated run_all.sh to run 6 variants (added BioT5 variants)
- Added per-fold json artifacts (best_params/test_metrics/splits/history/summary)
- Model weights (.pt) excluded per .gitignore
 2026-06-24 15:33:07 +08:00
Michelle0474
182f7853d7 Merge branch 'feat/moe-layer' of github-michelle:RYDE-WORK/lnp_ml into feat/moe-layer 
# Conflicts:
#	lnp_ml/modeling/layers/llm_prompt.py
 2026-06-22 19:46:05 +08:00
Michelle0474
54c734dc16 feat: integrate MolT5 LLM encoder and fix desc dim 2026-06-22 19:39:24 +08:00
Michelle0574
33fe637364 feat(training): size 头正则化(dropout+weight_decay)与 size z-score，更新消融汇总 2026-06-20 02:45:00 +00:00
Michelle0574
cd9166d63c feat(moe): MoE + MolT5 LLM 融合层、消融流水线 2026-06-18 08:51:30 +00:00
Michelle0474
cd4534f05a Add Mixture-of-Experts (MoE) layer support across modeling pipeline 2026-05-26 15:26:18 +08:00
RYDE-WORK
804bda5576 Update score cal rules 2026-03-04 17:14:36 +08:00
RYDE-WORK
28b181e194 数据预处理时自动存档总荧光强度的均值与方差,以便于预测时反演荧光强度 2026-03-03 15:48:59 +08:00
RYDE-WORK
8e02bf7f05 增加特征重要度结果 2026-03-03 15:23:10 +08:00
RYDE-WORK
ac5f598484 移除无关任务 2026-03-03 15:04:13 +08:00
RYDE-WORK
3b38727053 增加特征重要性计算 2026-03-03 13:45:46 +08:00
RYDE-WORK
4b75b7406d Merge branch 'main' of github.com:RYDE-WORK/lnp_ml 2026-03-03 10:42:03 +08:00
RYDE-WORK
25938df4ce Add feature importance 2026-03-03 10:41:16 +08:00
RYDE-WORK
175b07473c Final Model 2026-03-02 14:51:54 +08:00
RYDE-WORK
722a9d59f3 Merge branch 'main' of github.com:RYDE-WORK/lnp_ml into main 2026-03-02 14:51:09 +08:00
RYDE-WORK
447f2543f7 增加baseline评估makefile命令 2026-03-02 11:12:54 +08:00
RYDE-WORK
26fe818905 Merge branch 'main' of github.com:RYDE-WORK/lnp_ml into main 2026-02-28 18:29:33 +08:00
RYDE-WORK
ead579b25c Fallback 2026-02-28 18:29:07 +08:00
RYDE-WORK
002aaebd6f Merge branch 'main' of github.com:RYDE-WORK/lnp_ml into main 2026-02-28 17:53:28 +08:00
RYDE-WORK
74fd012f13 尝试切换学习率策略为线性预热+余弦退火 2026-02-28 17:51:40 +08:00
RYDE-WORK
a7db8ffc15 缩小d_model/num_heads/n_attn_layers/hidden_dim或者增大正则化后,欠拟合,恢复默认参数 2026-02-28 17:32:49 +08:00
RYDE-WORK
e3f9d9e9db 统一所有脚本参数统计方式 2026-02-28 17:08:22 +08:00
RYDE-WORK
e925a338b9 Merge branch 'main' of github.com:RYDE-WORK/lnp_ml into main 2026-02-28 16:50:16 +08:00
RYDE-WORK
985f3a1bb0 加载模型时显式weight_only=False 2026-02-28 16:49:34 +08:00
RYDE-WORK
c8c33c1c75 撤回正则化更改 2026-02-28 16:48:45 +08:00
RYDE-WORK
ca459470df Add new final model 2026-02-28 16:28:56 +08:00
RYDE-WORK
bbc7c88f99 增大benchmark时的固定dropout和weight decay 2026-02-28 16:17:40 +08:00
RYDE-WORK
70b7a4c62a 移除旧版函数参数,以及抑制chemprop/torch影响终端输出查看的过多warning 2026-02-28 15:55:47 +08:00
RYDE-WORK
7c69d47238 微调阶段rdkit缓存共享 2026-02-28 15:40:29 +08:00
RYDE-WORK
f952033b09 Reuse cacje 2026-02-28 15:35:10 +08:00
RYDE-WORK
13b357ce05 增加5-fold训练的并行功能 2026-02-28 14:09:00 +08:00
RYDE-WORK
48de4d5d37 增加超参搜索范围 2026-02-28 13:57:31 +08:00
RYDE-WORK
9e4800b77e 内外部数据LNP四组分比例统一为百分比数,可电离脂质与mRNA比例的搜索范围与四组分解耦 2026-02-28 11:52:55 +08:00
RYDE-WORK
60e082af55 多任务微调时默认不加载预训练时quantified_delivery的预测头参数 2026-02-28 11:13:39 +08:00
RYDE-WORK
c52b82786d 对backbone与预测头采用分层学习率,缓解预训练任务数与任务定义都不一致的问题 2026-02-28 11:11:27 +08:00
RYDE-WORK
73921bb353 移除微调时对于预训练已固定参数的搜索,增加超参搜索时正则化搜索范围上限(加强正则化) 2026-02-27 11:39:58 +08:00
RYDE-WORK
3a45c0641c ... 2026-02-26 20:06:27 +08:00
RYDE-WORK
c6e0a974cc 移除其他无用脚本 2026-02-26 19:31:48 +08:00
RYDE-WORK
00f51f37f0 整合与baseline比对的相关脚本 2026-02-26 18:31:22 +08:00
RYDE-WORK
c7c33e3f48 ... 2026-02-26 18:10:49 +08:00
RYDE-WORK
b74f3ab4b2 按照数据预处理脚本变更修正其他相应脚本 2026-02-26 18:07:39 +08:00
RYDE-WORK
72c292a91f 整理Makefile结构 2026-02-26 18:06:07 +08:00
RYDE-WORK
0e917ef0d4 引入最新数据&预先修复数据错误&将luminescence为0的分布数据统一为0 2026-02-26 18:00:25 +08:00
RYDE
5730490642
Merge pull request #1 from RYDE-WORK/amine_split 
Amine split
 2026-02-11 16:55:20 +08:00
RYDE-WORK
0b90caef1d Resolve conflicts 2026-02-11 16:51:21 +08:00
RYDE-WORK
a9392aa780 Update models and UI 2026-02-11 16:49:28 +08:00
RYDE-WORK
3f33f9d233 Add lfs 2026-02-11 16:45:21 +08:00
RYDE-WORK
c225fc67a7 Dedicated docker-compose.yml for gpu deployment 2026-01-26 11:11:52 +08:00