Implement MAML meta-opt

[wj-Mcat]

Description

Try to refactor MAML meta learning algriothm to make it more reusable in paddle-based applications.

Consideration

In MAML module, there are three things which is wired against the normal model application code：

• clone module which retrain the computation graph
• accumulate the gradient on cloned model
• backward the gradient based on query set data

Design

class BaseLearner(ABC):
    """Abstract Base Learner Class"""
    def __init__(self, module: Layer, optimizer: Optimizer) -> None:
        """The constructor of BaseLearner

        Args:
            module (Layer): the model to be trained
        """
        super().__init__()
        self._source_module = module
        self.cloned_module = None
        self.optimizer = optimizer

    def new_cloned_model(self,) -> Layer:
        """get the cloned model and keep the computation gragh

        Returns:
            Layer: the cloned model
        """
        self.cloned_module = clone_model(self._source_module)
        return self.cloned_module

    @abstractmethod
    def adapt(self, train_loss: Tensor) -> None:
        """Adapt the model to the current training loss

        Args:
            train_loss (Tensor): the current training loss
        """
        raise NotImplementedError


    @abstractmethod
    def step(self) -> None:
        """Perform a step of training

        Args:
            loss (float): _description_

        Raises:
            NotImplementedError: _description_
        """
        raise NotImplementedError

    def clear_grad(self):
        """clear the gradient in the computation graph
        """
        self.optimizer.clear_grad()

• adapt: accumulate the gradient on cloned model
• new_cloned_model: clone and save the model
• step: run step on optimizer in parameters of source model
• clear_grad: clear the gradient

[wj-Mcat]

Try to accomplish the task : #17

[wj-Mcat]

Changes in Learner Structure

class BaseLearner(Layer):
    """Abstract Base Learner Class"""
    def __init__(self, module: Layer) -> None:
        """The constructor of BaseLearner

        Args:
            module (Layer): the model to be trained
        """
        super().__init__()
        self.module = module

    @abstractmethod
    def adapt(self, loss: Tensor) -> None:
        """Adapt the model to the current training loss

        Args:
            loss (Tensor): the current training loss
        """
        raise NotImplementedError

-    def new_cloned_model(self,) -> Layer:
+   def clone(self: Type[Learner]) -> Learner:
        """create cloned module and keep the computation gragh

        Args:
            self (Type[Learner]): the sub-learner

        Returns:
            Learner: the cloned model
        """
        raise NotImplementedError

    def forward(self, *args, **kwargs):
        return self.module(*args, **kwargs)

-   @abstractmethod
-   def step(self) -> None:
-        """Perform a step of training
-
-        Args: 
-           loss (float): _description_
-
-        Raises:
-            NotImplementedError: _description_
-        """
-       raise NotImplementedError
-
-   def clear_grad(self):
-        """clear the gradient in the computation graph
-        """
-        self.optimizer.clear_grad()

As the above code shown, there are mainly two changes:

• change new_clone_model to clone which can reuse the meta algorithm.
• remove optimzier from learner which handle the outer loop

[tata1661]

@wj-Mcat Can you provide a detailed README.md to compare your empirical results with existing ones provided by PaddleFSL? You can put it in examples/optim. Thanks for the contribution!

[wj-Mcat]

Oh sorry, I have relaxed myself a few days. I will run the experiments to get the empirical results in the next few days.

[wj-Mcat]

To verify the effectiveness of the optim method, I have done some experiments based on model-zoon algo and optim algo. Below are my conclusion:

Omniglot - MAML

• model-zoo

• optim

Omniglot - ANIL

• model-zoon

• optim

MiniImagenet - ANIL

• model-zoo

• optim

CIFAR-FS - MAML

• model-zoo

• optim

[wj-Mcat]

Metric Overview

Dataset	Algo	model zoo(first order)	Optim(first order)
Omniglot	MAML	97.25 ± 1.7	97.07 ± 2.4
Omniglot	ANIL	93.62 ± 2.08	94.80 ± 3.7
MiniImageNet	ANIL	52.56 ± 3.5	57.50 ± 3.2
CIFAR-FS	MAML	46.88 ± 3.4	49.44 ± 4.7

[wj-Mcat]

Thanks for merging, I will try to fix #28 with another PR.