misc notes

• [done] 6/17 Make old chatbot compatible with tensorflow 1.0
• Make basic chatbot
  • 6/11 Prepare pycharm env with python 3.3 and latest TensorFlow
  • 6/12 Port data.py to Python3
    • 6/12 fix O(1) slowness
    • 6/12 Compare generated file with the original ones.
  • 6/12 Investigate why data.py is so slow
  • 6/17 Write own train.py
    • 6/13 make an empty file
    • 6/13 commit it
    • 6/14 Read https://www.tensorflow.org/install/migration
    • 6/17 Make model
    • 6/17 Comment for each unknown lines
    • 6/17 write model function
    • 6/17 See if it's easy to port the old chatbot to 0.1.0 using https://github.com/tensorflow/models/blob/master/tutorials/rnn/translate/seq2seq_model.py
  • 6/17 Very naive implmentation with the same data as CS20SI
  • 6/17 Compare the two bots
  • 6/17 Probably rewrite it in new seq2seq API
• 6/18 Find a best way to record trial and error
  • 6/18 model parameters (do we have to keep all the parameter files?)
    • Yes, with directory in tag name
  • 6/18 How we construct model
  • 6/18 Maybe tag or release?
  • 6/18 Think what should be done while training? Read papers?
• 6/18 Re-read http://web.stanford.edu/class/cs20si/assignments/a3.pdf and make a list of todo again
• 6/18 Re-read seq2seq model API https://www.tensorflow.org/tutorials/seq2seq
• 6/18 Improvement 1: Support Tensorboard
• 6/19 Try AdagradOptimizer and compare hom much faster it converges
• 6/19 Check if dataset is on memory
• Improvement 2: Construct the response in a non-greedy way
  • 6/24 Understand general beam search
  • 6/25 Read wikipedia
  • 6/25 Find actual tensorflow beam search example
  • 6/25 Make list of items to immplement
    • 6/25 wait, maybe implment in the graph is the right way?
    • 6/25 Understand current greedy model throughly?