logs16:RL test shorter reply is better
| Log Type | Detail |
|---|---|
| 1: What specific output am I working on right now? | Main purpose is seeing if RL works as expected. This should be reproducible. |
| 2: Thinking out loud - hypotheses about the current problem - what to work on next - how can I verify |
If average len goes down and it's reproducible, we can conclude RL is working. |
| 3: A record of currently ongoing runs along with a short reminder of what question each run is supposed to answer | Did avg len go down? If not, why do you think it didn't |
| 4: Results of runs and conclusion | Yes! it quickly converged to len == 1. |
| 5: Next steps | - (1) change this to len == 2 (most reward) to see if this is reproducible. (2) train seq2seq then see if this converge |
| 6: mega.nz | 20180516_test_medium25 |
| RL | ||
|---|---|---|
 |
 |
{'machine': 'client1', 'batch_size': 64, 'num_units': 512, 'num_layers': 2, 'vocab_size': 5000, 'embedding_size': 256, 'learning_rate': 0.5, 'learning_rate_decay': 0.99, 'use_attention': True, 'encoder_length': 28, 'decoder_length': 28, 'max_gradient_norm': 5.0, 'beam_width': 0, 'num_train_steps': 22, 'model_path': 'model/tweet_large'} dst {'machine': 'client1', 'batch_size': 64, 'num_units': 512, 'num_layers': 2, 'vocab_size': 5000, 'embedding_size': 256, 'learning_rate': 0.1, 'learning_rate_decay': 0.99, 'use_attention': True, 'encoder_length': 28, 'decoder_length': 28, 'max_gradient_norm': 5.0, 'beam_width': 0, 'num_train_steps': 1560, 'model_path': 'model/tweet_large_rl'} |