albert_base couldn't find checkpoint

brightmart / albert_zh

A LITE BERT FOR SELF-SUPERVISED LEARNING OF LANGUAGE REPRESENTATIONS, 海量中文预训练ALBERT模型

https://arxiv.org/pdf/1909.11942.pdf

3.94k stars 753 forks source link

Closed Zessay closed 4 years ago

Zessay commented 4 years ago

albert_base_zh(额外训练了1.5亿个实例即 36k steps * batch_size 4096) 转换成pytorch.bin的过程中会发生 "couldn't find checkpoint"的错误，其他模型都没有问题。

GanjinZero commented 4 years ago

出现了同样的问题

Zessay commented 4 years ago

出现了同样的问题

已经解决了，将文件夹下的checkpoint文件里面的内容修改为：

model_checkpoint_path: "albert_model.ckpt"
all_model_checkpoint_paths: "albert_model.ckpt"

就可以了！