中文：
一、整体框架及模型侧
50+排序（包含序列、多任务）、召回模型，tf和torch双语言支持
召回模型实现多分类负采样策略
序列模型统一实现可选的add和label_attention策略，使listwise模型转为可用的pointwise模型
FM类模型的高度统一实现和拓展
layer和model分离，用户可引入不同layer组合新的模型架构
统一seed，使得模型训练效果稳定可复现
模型参数尽可能多开放，并可以个性化配置
二、特征工程侧
统一特征数据输入规范：数值特征（必选）+类别特征（必选）+多值特征（可选）+序列特征（可选）
个性化生成测试训练数据。平等的使各种设备、各种配置都能运行、测试、体验深度模型。
实现部分常用的特征工程函数。
三、语言性能侧
tf模型充分利用tf.function的优势
统一tf和torch的模型权重初始化和默认应对策略
torch模型通过自建build函数自适应输入尺寸，使得torch和tf统一
torch模型拓展自身函数，如BN、LN等，使得torch和tf统一
torch模型统一增加L1、L2正则
English:
I. Overall Framework and Model Side
- Over 50 sorts (including sequences, multi-task), recall models, dual-language support for TensorFlow and PyTorch
- Multi-class negative sampling strategy implemented in recall models
- Unified implementation of optional add and label_attention strategies in sequence models, converting listwise models to usable pointwise models
- Highly unified implementation and expansion of FM-like models
- Separation of layers and models, allowing users to combine different layers to create new model architectures
- Unified seed for stable and reproducible model training effects
- As many model parameters as possible are open and can be individually configured
II. Feature Engineering Side
- Unified feature data input standards: Numerical features (mandatory) + Categorical features (mandatory) + Multi-valued features (optional) + Sequence features (optional)
- Personalized generation of test and training data. Equally enabling various devices and configurations to run, test, and experience deep models.
- Implementation of some commonly used feature engineering functions.
III. Language Performance Side
- TensorFlow models fully utilize the advantages of tf.function
- Unified weight initialization and default strategies for TensorFlow and PyTorch models
- PyTorch models adapt to input sizes through self-built build functions, unifying PyTorch and TensorFlow
- PyTorch models expand their own functions, such as BN (Batch Normalization), LN (Layer Normalization), etc., to unify PyTorch and TensorFlow
- Unified addition of L1 and L2 regularization to PyTorch models