songsenand
  • Joined on 2024-01-17
songsenand pushed to main at songsenand/SUInput 2026-02-15 01:07:06 +08:00
94b44e6f71 添加损失权重支持并重构部分模块结构
songsenand pushed to main at songsenand/SUInput 2026-02-15 00:25:48 +08:00
515f261824 修复模型加载方法,使用正确的实例方法加载状态字典
songsenand pushed to main at songsenand/SUInput 2026-02-15 00:08:59 +08:00
fd913748ca 调整残差块和分类头的 dropout 概率,并新增残差模块到 MoE 模型
songsenand pushed to main at songsenand/SUInput 2026-02-14 23:34:41 +08:00
e91f823d65 feat: 优化模型输入处理与专家数量,增强训练与推理兼容性
songsenand pushed to main at songsenand/SUInput 2026-02-14 17:07:42 +08:00
9fad2bf1d4 修复损失计算方式,使用NLLLoss替代原始criterion
songsenand pushed to main at songsenand/SUInput 2026-02-14 15:51:11 +08:00
f89635b201 添加 package-data 配置以包含 trainer 和 suinput 模块的额外数据文件
songsenand pushed to main at songsenand/SUInput 2026-02-14 15:43:31 +08:00
d60997438e 更新评估数据集样本文件
songsenand pushed to main at songsenand/SUInput 2026-02-14 15:29:35 +08:00
b68f75b09d 修复 char_info.pinyin 访问方式,使用字典形式确保兼容性
songsenand pushed to main at songsenand/SUInput 2026-02-14 15:27:16 +08:00
d2d65c7efa 调整导入顺序并修复pickle保存逻辑
134c8a09cf feat: 重构拼音输入数据集与 MoE 模型结构,优化专家网络配置及评估逻辑
Compare 2 commits »
songsenand pushed to main at songsenand/SUInput 2026-02-13 16:13:21 +08:00
7eb00c6207 feat(model): 优化专家输出结构并添加专家偏置支持
songsenand pushed to main at songsenand/SUInput 2026-02-13 15:44:52 +08:00
f4be47df78 feat(trainer): 使用 hidden_size 代替 d_model 计算输出维度并添加池化层
songsenand pushed to main at songsenand/SUInput 2026-02-13 14:20:15 +08:00
d82c80f3a9 修复分类头输出维度,使用 d_model 替代 hidden_size
songsenand pushed to main at songsenand/SUInput 2026-02-13 14:15:51 +08:00
6923870171 修复输出维度计算错误,使用 d_model 代替 input_dim
songsenand pushed to main at songsenand/SUInput 2026-02-13 12:58:38 +08:00
0e3418798e 添加自定义学习率调度支持并优化默认优化器配置
songsenand pushed to main at songsenand/SUInput 2026-02-13 12:12:45 +08:00
335540d8c2 调整学习率阈值并优化日志输出精度
songsenand pushed to main at songsenand/SUInput 2026-02-13 11:31:09 +08:00
02f851205f 修复周期性评估时平均损失计算错误
songsenand pushed to main at songsenand/SUInput 2026-02-13 11:05:58 +08:00
92b12ef703 调整数据集打乱缓冲区大小并优化样本处理逻辑
songsenand pushed to main at songsenand/SUInput 2026-02-13 10:55:17 +08:00
54ac5af876 feat: 优化数据加载与训练逻辑,增加自定义学习率调度支持
songsenand pushed to main at songsenand/SUInput 2026-02-13 01:45:22 +08:00
982d0521d5 添加日志记录和确保模型处于训练模式
songsenand pushed to main at songsenand/SUInput 2026-02-13 01:32:03 +08:00
35e835f618 使用 hint 字段替代原始 input_ids 和 attention_mask 进行推理