Xiaonan Li
+1. I'm wondering how I can get the exported structural data from devdocs.
It's a pity the authors do not release original dataset
好像是哦，谢谢，可能这正好是另一个issue的原因
sorry for late reply. You should manually init fitlog in your project dictionary,please refer to https://fitlog.readthedocs.io/zh/latest/ And if you want to run the main quickly, you can choice one of...
@w5688414 maybe it' renamed by fastNLP.io.loader, and I have fixed it in code.
> flat_main.py： param_ = [{'params':non_embedding_param}, {'params':embedding_param, 'lr':args.lr*args.embed_lr_rate}] > 1、默认args.embed_lr_rate=1，相当于没有设置单独的学习率； > 2、bigram_embedding和lattice_embedding都是由预训练词向量文件所载入，思考是不是使用较小学习率微调即可 是一个调参的思路
> flat_main.py：lrschedule_callback = LRScheduler(lr_scheduler=LambdaLR(optimizer, lambda ep: 1 / (1 + 0.05*ep) )) > 请问这一段的作用是什么呢？学习率调整不是已经在前面定义了么？望指教，谢谢 请问前面定义的学习率调整是指？
weight_decay对应的应该是l2正则项，LRScheduler对应的学习率逐渐变小。我实验下来SGD比ADAM好一点
1.可以在命令行参数中设置--use_bigram为False 2.https://github.com/jiesutd/LatticeLSTM 这里有谷歌硬盘的链接