Home

Seq2Seq RL Chatbot implementation notes

Links

• Algorithm for Reinforcement Learning
• Log Template

IN PROGRESS

• logs30:Show sampled reply for large data
• logs31:Get logits and log_prob

ideas & concerns

• client id on the chart is broken.
• log policy entropy
• limit iteration to see if the learning works.
• save iterator state

History

• logs0:old
• logs1:Use easy tf log
• logs2:Revisit auth on PyDrive
• logs3:How to save model
• logs4:See if RL framework is working
• logs5:Consolidate logs
• logs6:Make RL model work with small model
• logs7:Test RL with small model
• logs8:Understand policy entropy TO REVISIT
• logs9:Fix msec data issue
• logs10:See if RL works with medium model
• logs11:See if RL works with medium model2
• logs12:Train seq2seq with Large Data
• logs13:Understand Policy Gradient
• logs13:Understand Policy Gradient
• logs14:Normalize Reward
• logs15:See if new RL is working
• logs16:RL test shorter reply is better
• logs17:RL test len equals 2 is the best
• logs18:RL test train seq2seq first
• logs19: Random negative reward to avoid 0 loss
• logs20: Make a list of how we implement RL for seq2seq
• logs21: Steps to make small RL work
• logs22:Refactoring ideas
• Logs23:Train large seq2seq
• Logs24: Run RL on large seq2seq
• Logs25:Fix sample helper
• Logs26: Confirm dull response work for small tweets
• Logs27: Mutual Information
• Logs28: Mutual Information Observe
• Logs29: Wrap up Mutual Information Beam