MetaTrainer

MetaTrainer is the key component for training a meta learning method.

Overall, we extend Trainer to MetaTrainer. If you want to implement a meta learning model, please extend this class and implement _train_epoch() method. eg. You can create MeLUTrainer(MetaTrainer) and implement its _train_epoch() method.

The extended modification can be listed briefly as following:

• [Override] self.evaluate(self, eval_data, load_best_model=True, model_file=None, show_progress=False): Evaluation.

• [Abstract] self.taskDesolve(task): Desolve a task into spt_x,spt_y,qrt_x,qrt_y.

• [Abstract] self._train_epoch(self, train_data, epoch_idx, loss_func=None, show_progress=False): An epoch of training.

self.evaluate(eval_data, load_best_model=True, model_file=None, show_progress=False)

We adapt the evaluation process with task circumstance in meta learning.

self.taskDesolve(task)

This is an abstract method which is waiting for specific model to implement. It desolves a task into spt_x,spt_y,qrt_x,qrt_y.

task(Task)

The object of class Task

[return] spt_x,spt_y,qrt_x,qrt_y

The four base parts of task in meta learning.

self._train_epoch(train_data, epoch_idx, loss_func=None, show_progress=False)

This is an abstract method which is waiting for specific model to implement. This method indicates for an epoch of training. It can be called by self.fit().