Open tomergi opened 5 years ago
Are you using synthetic data as cooked traces?
No, cooked traces are generated from a broadband and a cellular network trace datasets. More details in the paper section 5.1, "a broadband dataset provided by the FCC [10] and a 3G/HSDPA mobile dataset collected in Norway [37]."
I'm trying to reproduce the results of the paper. I crated cooked traces using the same datasets used in the paper, and used a 80/20 split for training and testing. It seems like MPC performs better than both the pretrained model and models I tried to train. I added the fix in issue #20 and I use the following schedule for the entropy weight: 5 until epoch 20000 4 until epoch 30000 3 until epoch 40000 2 until epoch 50000 1 until epoch 60000 0.5 until epoch 70000 0.2 until epoch 80000 0.1 until epoch 90000 0.01 until epoch 100000 where the training ends. I get these results:![problem_with_pensieve_training](https://user-images.githubusercontent.com/5486408/59979807-26ae6400-95f5-11e9-8ff1-a7b573c0b665.png)
What am I doing wrong? Is the total number of epochs correct?