时间序列

work

1. 时间线

• 19/06: 定下方向 - nlp
• 19/06 - 19/07:
  
  • 继续nlp基础，复习Coursera，课程项目(Text Generator)，重新理解seq2seq
  • 学习LM, Embeddings，笔记
  • LSTM + Attention, CNN, BERT开始做简单项目练手，感性-理性
  • 翻译知识储备，语料库收集
• 19/08 - current：
  
  • Attention模型: 语料预处理，word-piece分词，模型训练
  • Transformer模型: 语料预处理，moses tokenizer，jieba cut，bpe，模型训练
  • Api化：模型部署，nematus代码贡献

2. 技术栈

2.1 Deep Learning

• CNN
• RNN
• seq2seq
• Encoder-Decoder
• Attention
• Training & Tuning

2.2 NLP

• Tokenize: Bpe, jieba, moses
• LM: skip-gram, cBow
• Word Embedding
• Attention
• Transformer
• Bert

2.2 Toolkit

• Python
• Tensorflow
• CUDA GPU
• Linux, Shell
• Jieba
• nematus
• WMT

3. Key Result

3.1 API

3.2 Test Case

• 例句：
  ["yeah", ",", "I", "'m", "an", "engineer", "working", "on", "artificial", "intelligence", "algorithms", "."]，12个token
• 数量：500
• 硬件：gtx 1080 Ti gpu * 1
• 速度：110 sents/sec, 1200 words/sec(approximate)
  Translated 500 sents in 4.594204902648926 sec. Speed 108.83276009559567 sents/sec

4. Support

• Texts Process
• Embeddings
• LM
• seq2seq
• Bert
• 语音相关
• OCR