This week

• [x] Improve and test the rnn-mix model
• [x] Conclude the experiments on the synthetic dataset

Day 1-5

This week has been spent testing and experimenting, it does not make sense to report each individual run, better to report the results and the conclusions:

Google Colaboratory

Last week we ended with a baseline model from which start experimenting new hyperparameters, It performed worst than our base population model, can we improve it?

What we could change is listable with some points:

Hyper-parameters

What can we tweak?

1. Is embedding better than encoding?
2. Change just the Sequence length [3,4,5] and run for 300 epochs. Is it better with a longer SQL?
3. With the best SQL, change the number of RNN neurons [8, 16, 32, 64, 128], which is the best?
4. With all that in place, try to add a dropout layer
5. Now that the first RRN layer is optimized, let's change the NN, let's start by changing the number of hidden layers [1,2] then change their number of neurons [16:128]
6. Test different LR and Batch sizes

To testing we define a new model: