Aleksei Lymar

An attempt was made in #724

Create a Trainer class for learning rate scheduling

2

For now learning rate scheduling is implemented using the [LRScheduledModel](https://github.com/deepmipt/DeepPavlov/blob/613d265f7371ba05365a7d44485066293c169674/deeppavlov/core/models/lr_scheduled_model.py#L129-L132) class and its inheritors. It would be great to move this logic to a trainer class inherited from the [NNTrainer](https://github.com/deepmipt/DeepPavlov/blob/613d265f7371ba05365a7d44485066293c169674/deeppavlov/core/trainers/nn_trainer.py#L33-L38)...

Add ability to cache preprocessing for datasets

2

Running preprocessing every epoch slows SQUAD training 2x times. This is also an issue of #569 where preprocessing is embedded in a dataset iterator

`OPTION` requests are not allowed and `POST` responses do not have CORS headers.