feie9454/kronos-trader
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

update readme & figs

Browse Source
This commit is contained in:
zhangboyu1 2025-10-09 16:24:35 +08:00
parent 2f7d3484cf
commit 38b5176cb7
2 changed files with 140 additions and 95 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

106
finetune_csv/README.md
View File

@ -1,50 +1,60 @@
# Kronos Finetuning on Your Custom csv Dataset # Kronos Fine-tuning on Custom CSV Datasets
Supports fine-tuning training with custom CSV data using configuration files This module provides a comprehensive pipeline for fine-tuning Kronos models on your own CSV-formatted financial data. It supports both sequential training (tokenizer followed by predictor) and individual component training, with full distributed training capabilities.
## 1. Prepare Your Data
**Data Format**: Ensure CSV file contains the following columns: `timestamps`, `open`, `high`, `low`, `close`, `volume`, `amount` ## 1. Data Preparation
A good csv data should be like: ### Required Data Format
Your CSV file must contain the following columns in this exact order:
- `timestamps`: DateTime stamps for each data point
- `open`: Opening price
- `high`: Highest price
- `low`: Lowest price
- `close`: Closing price
- `volume`: Trading volume
- `amount`: Trading amount
(volume and amount can be 0 if not available)
### Sample Data Format
| timestamps | open | close | high | low | volume | amount | | timestamps | open | close | high | low | volume | amount |
|------------|------|-------|------|-----|--------|--------| |------------|------|-------|------|-----|--------|--------|
| 2019/11/26 9:35 | 182.45215 | 184.45215 | 184.95215 | 182.45215 | 15136000 | 0 | | 2019/11/26 9:35 | 182.45215 | 184.45215 | 184.95215 | 182.45215 | 15136000 | 0 |
| 2019/11/26 9:40 | 184.35215 | 183.85215 | 184.55215 | 183.45215 | 4433300 | 0 | | 2019/11/26 9:40 | 184.35215 | 183.85215 | 184.55215 | 183.45215 | 4433300 | 0 |
| ... | ... | ... | ... | ... | ... | ... | | 2019/11/26 9:45 | 183.85215 | 183.35215 | 183.95215 | 182.95215 | 3070900 | 0 |
| ... | ... | ... | ... | ... | ... | ... |
You can check "data/HK_ali_09988_kline_5min_all.csv" to find out the proper format. > **Reference**: Check `data/HK_ali_09988_kline_5min_all.csv` for a complete example of the proper data format.
## 2. Training
### Configuration Setup ## 2. Config Preparation
First edit the `config.yaml` file to set the correct paths and parameters:
Please edit the correct data path and set your training parameters.
```yaml ```yaml
# Data configuration # Data configuration
data: data:
data_path: "/path/to/your/data.csv" data_path: "/path/to/your/data.csv"
lookback_window: 512 lookback_window: 512 # Historical data points to use
predict_window: 48 predict_window: 48 # Future points to predict
# ... other parameters max_context: 512 # Maximum context length
...
# Model path configuration
model_paths:
pretrained_tokenizer: "/path/to/pretrained/tokenizer"
pretrained_predictor: "/path/to/pretrained/predictor"
base_save_path: "/path/to/save/models"
# ... other paths
``` ```
There are some other settings here, please see `configs/config_ali09988_candle-5min.yaml` for more comments.
### Run Training ## 3. Training
Using train_sequential ### Method 1: Sequential Training (Recommended)
The `train_sequential.py` script handles the complete training pipeline automatically:
```bash ```bash
# Complete training # Complete training (tokenizer + predictor)
python train_sequential.py --config configs/config_ali09988_candle-5min.yaml python train_sequential.py --config configs/config_ali09988_candle-5min.yaml
# Skip existing models # Skip existing models
@ -53,36 +63,58 @@ python train_sequential.py --config configs/config_ali09988_candle-5min.yaml --s
# Only train tokenizer # Only train tokenizer
python train_sequential.py --config configs/config_ali09988_candle-5min.yaml --skip-basemodel python train_sequential.py --config configs/config_ali09988_candle-5min.yaml --skip-basemodel
# Only train basemodel # Only train predictor
python train_sequential.py --config configs/config_ali09988_candle-5min.yaml --skip-tokenizer python train_sequential.py --config configs/config_ali09988_candle-5min.yaml --skip-tokenizer
``` ```
Run each stage separately ### Method 2: Individual Component Training
Train each component separately for more control:
```bash ```bash
# Only train tokenizer # Step 1: Train tokenizer
python finetune_tokenizer.py --config configs/config_ali09988_candle-5min.yaml python finetune_tokenizer.py --config configs/config_ali09988_candle-5min.yaml
# Only train basemodel (requires fine-tuned tokenizer first) # Step 2: Train predictor (requires fine-tuned tokenizer)
python finetune_base_model.py --config configs/config_ali09988_candle-5min.yaml python finetune_base_model.py --config configs/config_ali09988_candle-5min.yaml
``` ```
DDP Training ### DDP Training
For faster training on multiple GPUs:
```bash ```bash
# Choose communication protocol yourself, nccl can be replaced with gloo # Set communication backend (nccl for NVIDIA GPUs, gloo for CPU/mixed)
DIST_BACKEND=nccl \ DIST_BACKEND=nccl \
torchrun --standalone --nproc_per_node=8 train_sequential.py --config configs/config_ali09988_candle-5min.yaml torchrun --standalone --nproc_per_node=8 train_sequential.py --config configs/config_ali09988_candle-5min.yaml
``` ```
## 2. Training Results
![HK_ali_09988_kline_5min_all_historical_20250919_073929](examples/HK_ali_09988_kline_5min_all_historical_20250919_073929.png) ## 4. Training Results
![HK_ali_09988_kline_5min_all_historical_20250919_073944](examples/HK_ali_09988_kline_5min_all_historical_20250919_073944.png) The training process generates several outputs:
![HK_ali_09988_kline_5min_all_historical_20250919_074012](examples/HK_ali_09988_kline_5min_all_historical_20250919_074012.png) ### Model Checkpoints
- **Tokenizer**: Saved to `{base_save_path}/{exp_name}/tokenizer/best_model/`
- **Predictor**: Saved to `{base_save_path}/{exp_name}/basemodel/best_model/`
![HK_ali_09988_kline_5min_all_historical_20250919_074042](examples/HK_ali_09988_kline_5min_all_historical_20250919_074042.png) ### Training Logs
- **Console output**: Real-time training progress and metrics
- **Log files**: Detailed logs saved to `{base_save_path}/logs/`
- **Validation tracking**: Best models are saved based on validation loss
## 5. Prediction Vis
The following images show example training results on alibaba (HK stock) data:
![Training Result 1](examples/HK_ali_09988_kline_5min_all_historical_20250919_073929.png)
![Training Result 2](examples/HK_ali_09988_kline_5min_all_historical_20250919_073944.png)
![Training Result 3](examples/HK_ali_09988_kline_5min_all_historical_20250919_074012.png)
![Training Result 4](examples/HK_ali_09988_kline_5min_all_historical_20250919_074042.png)
![Training Result 5](examples/HK_ali_09988_kline_5min_all_historical_20250919_074251.png)
![HK_ali_09988_kline_5min_all_historical_20250919_074251](examples/HK_ali_09988_kline_5min_all_historical_20250919_074251.png)

129
finetune_csv/README_CN.md
View File

@ -1,36 +1,58 @@
# 自定义数据集的Kronos微调训练 # Kronos微调-支持自定义CSV数据集
支持使用配置文件进行自定义csv数据的微调训练 这是一个在自定义的CSV格式数据上微调Kronos模型的完整流程。包含顺序训练（先训练tokenizer再训练predictor）和单独模块训练，同时支持分布式训练。
## 快速开始
### 1. 配置设置 ## 1. 准备数据
首先编辑 `config.yaml` 文件，设置正确的路径和参数： ### 数据格式
您的CSV文件必须按以下确切顺序包含以下列：
- `timestamps`: 每个数据点的时间戳
- `open`: 开盘价
- `high`: 最高价
- `low`: 最低价
- `close`: 收盘价
- `volume`: 交易量
- `amount`: 交易金额
(volume和amount可以全0如果没有这部分的数据)
### 示例数据格式
| timestamps | open | close | high | low | volume | amount |
|------------|------|-------|------|-----|--------|--------|
| 2019/11/26 9:35 | 182.45215 | 184.45215 | 184.95215 | 182.45215 | 15136000 | 0 |
| 2019/11/26 9:40 | 184.35215 | 183.85215 | 184.55215 | 183.45215 | 4433300 | 0 |
| 2019/11/26 9:45 | 183.85215 | 183.35215 | 183.95215 | 182.95215 | 3070900 | 0 |
> **标准数据样例**: `data/HK_ali_09988_kline_5min_all.csv`
## 2. 准备config文件
data_path需要改成正确的数据路径，训练参数可以自己调节
```yaml ```yaml
# 数据配置 # 数据配置
data: data:
data_path: "/path/to/your/data.csv" data_path: "/path/to/your/data.csv"
lookback_window: 512 lookback_window: 512 # 要使用的历史数据点
predict_window: 48 predict_window: 48 # 要预测的未来点数
# ... 其他参数 max_context: 512 # 最大上下文长度
...
# 模型路径配置
model_paths:
pretrained_tokenizer: "/path/to/pretrained/tokenizer"
pretrained_predictor: "/path/to/pretrained/predictor"
base_save_path: "/path/to/save/models"
# ... 其他路径
``` ```
这里还有其他一些设置， `configs/config_ali09988_candle-5min.yaml` 有更详细的注释。
### 2. 运行训练 ## 3. 训练
### 方法1: 直接顺序训练
使用train_sequential `train_sequential.py` 脚本自动处理完整的训练流程：
```bash ```bash
# 完整训练 # 完整训练（tokenizer + predictor）
python train_sequential.py --config configs/config_ali09988_candle-5min.yaml python train_sequential.py --config configs/config_ali09988_candle-5min.yaml
# 跳过已存在的模型 # 跳过已存在的模型
@ -39,67 +61,58 @@ python train_sequential.py --config configs/config_ali09988_candle-5min.yaml --s
# 只训练tokenizer # 只训练tokenizer
python train_sequential.py --config configs/config_ali09988_candle-5min.yaml --skip-basemodel python train_sequential.py --config configs/config_ali09988_candle-5min.yaml --skip-basemodel
# 只训练basemodel # 只训练predictor
python train_sequential.py --config configs/config_ali09988_candle-5min.yaml --skip-tokenizer python train_sequential.py --config configs/config_ali09988_candle-5min.yaml --skip-tokenizer
``` ```
单独运行各个阶段 ### 方法2: 单独组件训练
可以单独训练每个组件：
```bash ```bash
# 只训练tokenizer # 步骤1: 训练tokenizer
python finetune_tokenizer.py --config configs/config_ali09988_candle-5min.yaml python finetune_tokenizer.py --config configs/config_ali09988_candle-5min.yaml
# 只训练basemodel（需要先有微调后的tokenizer） # 步骤2: 训练predictor（需要微调后的tokenizer）
python finetune_base_model.py --config configs/config_ali09988_candle-5min.yaml python finetune_base_model.py --config configs/config_ali09988_candle-5min.yaml
``` ```
DDP训练 ### DDP训练
如果有多卡，可以开启ddp加速训练：
```bash ```bash
# 通信协议自行选择，nccl可替换gloo # 设置通信后端（NVIDIA GPU用nccl，CPU/混合用gloo）
DIST_BACKEND=nccl \ DIST_BACKEND=nccl \
torchrun --standalone --nproc_per_node=8 train_sequential.py --config configs/config_ali09988_candle-5min.yaml torchrun --standalone --nproc_per_node=8 train_sequential.py --config configs/config_ali09988_candle-5min.yaml
``` ```
## 配置说明 ## 4. 训练结果
### 主要配置项 训练过程生成以下输出：
- **data**: 数据相关配置 ### 模型检查点
- `data_path`: CSV数据文件路径 - **Tokenizer**: 保存到 `{base_save_path}/{exp_name}/tokenizer/best_model/`
- `lookback_window`: 回望窗口大小 - **Predictor**: 保存到 `{base_save_path}/{exp_name}/basemodel/best_model/`
- `predict_window`: 预测窗口大小
- `train_ratio/val_ratio/test_ratio`: 数据集分割比例
- **training**: 训练相关配置 ### 训练日志
- `epochs`: 训练轮数 - **控制台输出**: 实时训练进度和指标
- `batch_size`: 批次大小 - **日志文件**: 详细日志保存到 `{base_save_path}/logs/`
- `tokenizer_learning_rate`: Tokenizer学习率 - **验证跟踪**: 基于验证损失保存最佳模型
- `predictor_learning_rate`: Predictor学习率
- **model_paths**: 模型路径配置 ## 5. 预测可视化
- `pretrained_tokenizer`: 预训练tokenizer路径
- `pretrained_predictor`: 预训练predictor路径
- `base_save_path`: 模型保存根目录
- `finetuned_tokenizer`: 微调后tokenizer路径（用于basemodel训练）
- **experiment**: 实验控制 以下图像显示了kronos在阿里巴巴股票数据上微调后的示例训练结果：
- `train_tokenizer`: 是否训练tokenizer
- `train_basemodel`: 是否训练basemodel
- `skip_existing`: 是否跳过已存在的模型
## 训练流程 ![训练结果 1](examples/HK_ali_09988_kline_5min_all_historical_20250919_073929.png)
1. **Tokenizer微调阶段** ![训练结果 2](examples/HK_ali_09988_kline_5min_all_historical_20250919_073944.png)
- 加载预训练tokenizer
- 在自定义数据上微调
- 保存微调后的tokenizer到 `{base_save_path}/tokenizer/best_model/`
2. **Basemodel微调阶段** ![训练结果 3](examples/HK_ali_09988_kline_5min_all_historical_20250919_074012.png)
- 加载微调后的tokenizer和预训练predictor
- 在自定义数据上微调 ![训练结果 4](examples/HK_ali_09988_kline_5min_all_historical_20250919_074042.png)
- 保存微调后的basemodel到 `{base_save_path}/basemodel/best_model/`
![训练结果 5](examples/HK_ali_09988_kline_5min_all_historical_20250919_074251.png)
**数据格式**: 确保CSV文件包含以下列：`timestamps`, `open`, `high`, `low`, `close`, `volume`, `amount`