recurrent-entity-networks
recurrent-entity-networks copied to clipboard
jimfleming
Task 16 accuracy
I have some issue to reach the expected accuracy on task 16. How should tune parameters in order to reach the expected accuracy?
Can you provide more information? What did you try, what version of the code, what do the logs say?
I cloned the actual trunk code and I tried different learning rate, but the accuracy is always around 0.5. Here the final log using default traing rate: INFO:tensorflow:global_step/sec: 6.85293 INFO:tensorflow:step = 34101, loss = 0.88134634 (14.653 sec) INFO:tensorflow:global_step/sec: 6.91063 INFO:tensorflow:step = 34201, loss = 0.86507523 (14.409 sec) INFO:tensorflow:global_step/sec: 6.64557 INFO:tensorflow:step = 34301, loss = 0.85566473 (15.049 sec) INFO:tensorflow:Saving checkpoints for 34321 into checkpointstrain\model.ckpt. Parameters: 49200 INFO:tensorflow:Starting evaluation at 2018-08-29-18:57:45 INFO:tensorflow:Graph was finalized. INFO:tensorflow:Restoring parameters from checkpointstrain\model.ckpt-34321 INFO:tensorflow:Running local_init_op. INFO:tensorflow:Done running local_init_op. INFO:tensorflow:Finished evaluation at 2018-08-29-18:57:47 INFO:tensorflow:Saving dict for global step 34321: accuracy = 0.491, global_step = 34321, loss = 0.8779338 Stopping... Best step: 18721 with accuracy = 0.5040000081062317. INFO:tensorflow:Loss for final step: 0.83799744. Parameters: 49200
In this moment I'm using windows and tensorflow 1.8.0 (CPU). Should be that impact on the result?
I have not tested it at all on Windows. It might be an issue, I'm not sure. Some of the tasks do not converge consistently, which may also be the issue.