sync local with remote by vscode

I want to use vscode as my main code editor.I need to sync my code from local to the remote GPU server.I choose to use sftp extensions in vscode.Aftering installing and reloading the extension, for M

Setup the LaTeX environment in vscode

I want to give up texstudio in my Mac due to its ugly GUI. OS macOS 10.14.2 Setup Steps Install MacTex. Do not update package in default setting especially you are SJTUer. If you update, the font wil

Extending Neural Generative Conversational Model using External Knowledge Sources

文章总结的两点还挺有意思的 用基于连接的方法来做对话是因为有大规模的语料 之前几年的对话系统研究都集中在把所有的组件都替换一遍。 工作中用了 wikipedia summaries NELL knowledge base related work中列觉了很多利用外部知识的对话系统的工作 这个工作中把context中的非停用词的相关knowledge都抽取出来，然后用Bag-of-words

PyTorch调参

在写CVAE模型的过程中，遇到一个loss突然变大的过程，看到网络上说由于Adam的原因，所以会导致收敛不稳定。可以把amdgrad参数打开。1optimizer = Adam(filter(lambda p: p.requires_grad, model.parameters()), lr=args.lr, weight_decay=args.wd, amsgrad=True)

Inductive Representation Learning on Large Graphs

这是一篇发表在NIPS2017上的工作，提出了一个模型叫做GraphSAGE，用来解决图表示的。提出了一个问题However, most existing approaches require that all nodes in the graph are present during training of the embeddings;作者把这些模型叫做transductive的，而他提出来的

Incorporating Loose-Structured Knowledge into LSTM with Recall Gate for Conversation Modeling

Introduction使用了Recall gate去将knowledge整合进LSTM中。其中由sentence modeling, knowledge triggering和conversation modeling components。想法看着还比较简单，就是先将knowldege pair先映射成dense vector，然后直接拼接到LSTM的cell中。

Monte Carlo Methods

Sampling and Monte Carlo MethodsWhy Sampling蒙特卡洛方法主要是为了计算出积分或者某些目标的近似值，通过随机采样的方式，这种在可求导的情况下可以提高计算速度。比如说用minibatch的方式来估计training loss。除此之外，有些时候还需要我们去估计intractable的值。 Basics of Monte Carlo Sampling$$s =

FastGCN Fast Learning with Graph Convolutional Networks via Importance Sampling

IntroductionBorrowing the concept of a convolution filter for image pixels or a linear array of signals, GCN uses the connectivity structure of the graph as the filter to perform neighborhood mixing.

Reinforcement Learning

入门Environment是确定的，还有一个Replay Memory来保存已被观测到的状态转移他们的目标是训练一个discounted并且cumulative的Reward函数。他让更久之后的reward的重要性降低。 Q-learning的主要思想是，如果我们有了一个Q函数 $Q^*: State \times Action \rightarrow \mathbb{R}$那么我们就能够知道我们

Commonsense Knowledge Aware Conversation Generation with Graph Attention