如何将HappyTransformer模型保存到指定目录而非默认~/.cache

问题描述 #

使用HappyTransformer的HappyTextToText时，模型默认会保存到.cache/目录，我希望将其改存到指定路径"/home/ubuntu/storage1/various_transformer_models/"，尝试把cache_dir传给TTSettings后执行报错：

TypeError: __init__() got an unexpected keyword argument 'cache_dir'

错误代码示例：

from happytransformer import HappyTextToText, TTSettings

cache_dir = "/home/ubuntu/storage1/various_transformer_models/"
tt_settings = TTSettings(cache_dir=cache_dir)
happy_tt = HappyTextToText(model_type="gpt", tt_settings=tt_settings)

解决方案 #

cache_dir是HappyTextToText类初始化时的直接参数，不属于TTSettings的配置项。正确做法是将cache_dir直接传给HappyTextToText的构造函数：

from happytransformer import HappyTextToText, TTSettings

cache_dir = "/home/ubuntu/storage1/various_transformer_models/"
# 基础用法：仅指定缓存目录
happy_tt = HappyTextToText(model_type="gpt", cache_dir=cache_dir)

# 如果需要配置生成参数，可单独定义TTSettings后传入
tt_settings = TTSettings(max_length=50, temperature=0.7)  # 添加你的生成相关配置
happy_tt = HappyTextToText(model_type="gpt", cache_dir=cache_dir, tt_settings=tt_settings)

说明 #

TTSettings仅用于配置文本生成的参数（如最大长度、温度、top_k等），和模型缓存目录无关。模型缓存路径的配置需要直接在初始化HappyTextToText时传入cache_dir参数，框架会自动将下载的模型文件保存到你指定的目录中。

内容的提问来源于stack exchange，提问作者littleworth