关于embeddings文件夹下的设置

[jianjianxi0612]

看到里面是关于github大文件存储的下载，是不是里面放的是词向量长度和神经网络层大小？？

[RomanGao]

并不是，网络层深度跟词向量大小没关系。网络层输入大小，跟embedding和输入的词汇个数有关 | | RomanGao | | 邮箱：gaochl2014@126.com | Signature is customized by Netease Mail Master 在2019年10月14日 16:55，jianjianxi0612 写道： 看到里面是关于github大文件存储的下载，是不是里面放的是词向量长度和神经网络层大小？？ — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[jianjianxi0612]

该文件里放的是某个链接，不明白里面的是关于什么的设置

[zhaodongh]

embeddings文件夹下是训练好的50维词向量

[cuter44]

Hello @zhaodongh. While fetching this embedding file, we encountered "This repository is over its data quota". We have done some search to find an alternative but no luck. Could you give some clue how we can get this file?
Thanks.

[Ameoi]

we are sorry. it is too long for us to find this file in our disk. we use embedding in https://github.com/aseveryn/deep-qa. but the embedding link is also invalid. Maybe you can replace it by other embedding

[shuishen112]

你好，你所说的embedding向量可以直接通过以下网址下载，我这里是可以下载成功的，如果你依然无法下载，请再次联系我。
https://github.com/zhaodongh/Quantum-QA/blob/master/embeddings/aquaint%2Bwiki.txt.gz.ndim%3D50.bin

另外如果你想复现论文中的实验结果，有几点需要注意：
第一：这个代码是基于theano开发，你需要配置好一个正确的theano开发环境。
第二：实验的初始化随机值对结果影响很大，可能你需要进行微调。
第三：由于该模型涉及到密度矩阵的训练，对超参比较敏感，比如batchsize，adam训练步长。

[cuter44]

For whom may concern,

We had tried glove.6b.50d, which at least make it run. There is a 10084/58617 word vectors deficit on TREC(significantly harm performance), and 4378/36781 on WIKI.

This snippet converts text-format embedding to NNQLM flavor.

Also thank for the original embedding.