还会有人训练这种东西

2023-08-23 13:06 作者:66662221 0人读过 | 我要投稿

一个基于gpt2的prompt生成器，采用的训练集是huggingface上提供的500万条从danbooru上收集的tag

这是其中几条tag:

huggingface上也有专门收集的stable diffusion和midjourney的prompt，但是其中有大量类似'high quality,masterpiece'之类的词，这些词完全可以用更简单的程序生成，并且如果将这些词加入训练集，最后的模型可能只会写这些词

但是这些从danbooru上收集的tag也不能直接拿去训练，因为其中有很多生僻词和人名，在训练的时候需要指定一个长度，如果tag小于这个长度需要填充，大于这个长度需要截断，这些词大部分都是按照个人喜好添加的，我希望模型可以生成描述类tag如衣服，头发等

删除这些词后，还要删除一些18+词汇和一些符号比如‘（’

然后要统计训练集中大部分tag的长度，我按照空格和逗号分词

统计发现大部分的tag长度集中在0-60，所以将长度设置为60，采用eos_token作为pad_token。

处理好的数据可以保存为npy文件。gpt2模型参数比较多，小型的也有1亿多，中型的更是有3亿，个人电脑训练很慢，所以我使用autodl上租的A5000训练。

再讲一下如何保存和使用huggingface上提供的模型，首先下载这些文件：

然后把这些文件放在文件夹，在使用from_pretrained方法时直接使用文件路径即可

训练不到一个epoch就可以达到70%多的准确率，这主要归功于gpt2

下面看一下生成效果，我设置最大长度为60：

生成tag后，你可以手动加入固定tag如“masterpiece"之类的，下面展示一些tag和图片：

(masterpiece),(best quality),(illustration) ,(extremely detailed CG unity 8k wallpaper),(best shadow), ultra-detailed ,finaly detail, solo,1girl, animal ears, blue theme, bow, capelet, dress, fire, glowing, glowing eyes, green bow, glowing eyes, hat, long sleeves, monochrome, short hair, sketch, skirt, solo, spot color, sword, sword of hisou, tabard,beautiful detailed eyes,(an extremely delicate and beautiful) , studio photography, extremely detailed wallpaper

(masterpiece),(best quality),(illustration) ,(extremely detailed CG unity 8k wallpaper),(best shadow), ultra-detailed ,finaly detail, solo,1girl, animal ears, blue dress, blue eyes, blush, braid, chakabo, cat ears, cat tail, dress, full moon, glowing, long hair, mob cap, moon, night, puffy sleeves, silver hair, skirt, solo, source request, tail,beautiful detailed eyes,(an extremely delicate and beautiful) , studio photography, extremely detailed wallpaper

(masterpiece),(best quality),(illustration) ,(extremely detailed CG unity 8k wallpaper),(best shadow), ultra-detailed ,finaly detail, solo,1girl, bow, breasts, brown hair, cleavage, dress, hair brush, hat, hat ornament, kijima ban, leaning forward, long hair, medium breasts, nature, open mouth, red eyes, ribbon, skirt, skirt set, smile, solo, water, white hair,rainy,street light,beautiful detailed eyes,(an extremely delicate and beautiful) , studio photography, extremely detailed wallpaper

(masterpiece),(best quality),(illustration) ,(extremely detailed CG unity 8k wallpaper),(best shadow), ultra-detailed ,finaly detail, solo,1girl, armor, breasts, choker, cloak, crossed arms, colored skin, dragon's crown, dress, earrings, fairy wings, gloves, high heels, jewelry, large breasts, magical girl, monochrome, necklace, no shoes, queen's blade, queen's blade rebellion,beautiful detailed eyes,(an extremely delicate and beautiful) , studio photography, extremely detailed wallpaper