# Degree project notes

## Evaluation and model selection
We all know that we can have hyper-parameters tuning on our developement set or validation set.

The bellow is the summary that how people work with it in the field of
Question generation task.

1. [du2017learning](https://arxiv.org/abs/1705.00106)
In their implementation, they use the loss of the model on validation set as the criterion for hyper-parameter tuning
2. [devlin2018bert](https://arxiv.org/abs/1810.04805v2)
check `master` branch `bert/run_pretraining.py#L187`
They use both loss and accuracy

## Length prediction
Check [adi2016fine](https://groups.csail.mit.edu/sls/publications/2017/ICLR17_Belinkov.pdf)

### Poisson regression
We would like to use poisson regression and
We can find a implmentation [here](https://discuss.pytorch.org/t/poisson-loss-function/44301/6)

The loss function is the negative log-likelihood of the generalized possion distribution. See [equation](https://en.wikipedia.org/wiki/Poisson_regression#Maximum_likelihood-based_parameter_estimation) and the [video](https://www.youtube.com/watch?v=0XfXHYDYoBA)