This PR adds a benchmark for Recurrent Networks. This benchmark is for a simple network, so we know how our future modifications are influencing the model compared to this simple form.
For some reason 10 units perform about 1-2% better than 512 units, so I've left 10 units in here. There's some imperfections to this benchmark, for example we should have an AUC and F1 score everywhere here. We should be doing cross validation of 10 too. Rerunning this with these perfections would be a good idea later, there's not enough time before the end of the sprint (and this stuff takes a while to run)
This PR adds a benchmark for Recurrent Networks. This benchmark is for a simple network, so we know how our future modifications are influencing the model compared to this simple form.
For some reason 10 units perform about 1-2% better than 512 units, so I've left 10 units in here. There's some imperfections to this benchmark, for example we should have an AUC and F1 score everywhere here. We should be doing cross validation of 10 too. Rerunning this with these perfections would be a good idea later, there's not enough time before the end of the sprint (and this stuff takes a while to run)