learn2learn.algorithms - learn2learn (2024)

[Source]

Description

High-level implementation of Model-Agnostic Meta-Learning.

This class wraps an arbitrary nn.Module and augments it with clone() and adapt()methods.

For the first-order version of MAML (i.e. FOMAML), set the first_order flag to Trueupon initialization.

Arguments

• model (Module) - Module to be wrapped.
• lr (float) - Fast adaptation learning rate.
• first_order (bool, optional, default=False) - Whether to use the first-order approximation of MAML. (FOMAML)
• allow_unused (bool, optional, default=None) - Whether to allow differentiation of unused parameters. Defaults to allow_nograd.
• allow_nograd (bool, optional, default=False) - Whether to allow adaptation with parameters that have requires_grad = False.

References

1. Finn et al. 2017. "Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks."

Example

123456

linear = l2l.algorithms.MAML(nn.Linear(20, 10), lr=0.01)clone = linear.clone()error = loss(clone(X), y)clone.adapt(error)error = loss(clone(X), y)error.backward()

[Source]

Description

High-level implementation of Meta-SGD.

This class wraps an arbitrary nn.Module and augments it with clone() and adaptmethods.It behaves similarly to MAML, but in addition a set of per-parameters learning ratesare learned for fast-adaptation.

Arguments

• model (Module) - Module to be wrapped.
• lr (float) - Initialization value of the per-parameter fast adaptation learning rates.
• first_order (bool, optional, default=False) - Whether to use the first-order version.
• lrs (list of Parameters, optional, default=None) - If not None, overrides lr, and uses the list as learning rates for fast-adaptation.

References

1. Li et al. 2017. “Meta-SGD: Learning to Learn Quickly for Few-Shot Learning.” arXiv.

Example

123456

linear = l2l.algorithms.MetaSGD(nn.Linear(20, 10), lr=0.01)clone = linear.clone()error = loss(clone(X), y)clone.adapt(error)error = loss(clone(X), y)error.backward()

[Source]

Description

General wrapper for gradient-based meta-learning implementations.

A variety of algorithms can simply be implemented by changing the kindof transform used during fast-adaptation.For example, if the transform is Scale we recover Meta-SGD [2] with adapt_transform=Falseand Alpha MAML [4] with adapt_transform=True.If the transform is a Kronecker-factored module (e.g. neural network, or linear), we recoverKFO from [5].

Arguments

• module (Module) - Module to be wrapped.
• tranform (Module) - Transform used to update the module.
• lr (float) - Fast adaptation learning rate.
• adapt_transform (bool, optional, default=False) - Whether to update the transform's parameters during fast-adaptation.
• first_order (bool, optional, default=False) - Whether to use the first-order approximation.
• allow_unused (bool, optional, default=None) - Whether to allow differentiation of unused parameters. Defaults to allow_nograd.
• allow_nograd (bool, optional, default=False) - Whether to allow adaptation with parameters that have requires_grad = False.

References

1. Finn et al. 2017. “Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks.”
2. Li et al. 2017. “Meta-SGD: Learning to Learn Quickly for Few-Shot Learning.”
3. Park & Oliva. 2019. “Meta-Curvature.”
4. Behl et al. 2019. “Alpha MAML: Adaptive Model-Agnostic Meta-Learning.”
5. Arnold et al. 2019. “When MAML Can Adapt Fast and How to Assist When It Cannot.”

Example

 1 2 3 4 5 6 7 8 910111213141516171819

model = SmallCNN()transform = l2l.optim.ModuleTransform(torch.nn.Linear)gbml = l2l.algorithms.GBML( module=model, transform=transform, lr=0.01, adapt_transform=True,)gbml.to(device)opt = torch.optim.SGD(gbml.parameters(), lr=0.001)# Training with 1 adaptation stepfor iteration in range(10): opt.zero_grad() task_model = gbml.clone() loss = compute_loss(task_model) task_model.adapt(loss) loss.backward() opt.step()

[Source]

Description

A PyTorch Lightning module for MAML.

Arguments

• model (Module) - A PyTorch nn.Module.
• loss (Function, optional, default=CrossEntropyLoss) - Loss function which maps the cost of the events.
• ways (int, optional, default=5) - Number of classes in a task.
• shots (int, optional, default=1) - Number of samples for adaptation.
• adaptation_steps (int, optional, default=1) - Number of steps for adapting to new task.
• lr (float, optional, default=0.001) - Learning rate of meta training.
• adaptation_lr (float, optional, default=0.1) - Learning rate for fast adaptation.
• scheduler_step (int, optional, default=20) - Decay interval for lr.
• scheduler_decay (float, optional, default=1.0) - Decay rate for lr.

References

1. Finn et al. 2017. "Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks."

Example

123456

tasksets = l2l.vision.benchmarks.get_tasksets('omniglot')model = l2l.vision.models.OmniglotFC(28**2, args.ways)maml = LightningMAML(classifier, adaptation_lr=0.1, **dict_args)episodic_data = EpisodicBatcher(tasksets.train, tasksets.validation, tasksets.test)trainer = pl.Trainer.from_argparse_args(args)trainer.fit(maml, episodic_data)

[Source]

Description

A PyTorch Lightning module for ANIL.

Arguments

• features (Module) - A nn.Module to extract features, which will not be adaptated.
• classifier (Module) - A nn.Module taking features, mapping them to classification.
• loss (Function, optional, default=CrossEntropyLoss) - Loss function which maps the cost of the events.
• ways (int, optional, default=5) - Number of classes in a task.
• shots (int, optional, default=1) - Number of samples for adaptation.
• adaptation_steps (int, optional, default=1) - Number of steps for adapting to new task.
• lr (float, optional, default=0.001) - Learning rate of meta training.
• adaptation_lr (float, optional, default=0.1) - Learning rate for fast adaptation.
• scheduler_step (int, optional, default=20) - Decay interval for lr.
• scheduler_decay (float, optional, default=1.0) - Decay rate for lr.

References

1. Raghu et al. 2020. "Rapid Learning or Feature Reuse? Towards Understanding the Effectiveness of MAML"

Example

123456

tasksets = l2l.vision.benchmarks.get_tasksets('omniglot')model = l2l.vision.models.OmniglotFC(28**2, args.ways)anil = LightningANIL(model.features, model.classifier, adaptation_lr=0.1, **dict_args)episodic_data = EpisodicBatcher(tasksets.train, tasksets.validation, tasksets.test)trainer = pl.Trainer.from_argparse_args(args)trainer.fit(anil, episodic_data)

[Source]

Description

A PyTorch Lightning module for Prototypical Networks.

Arguments

• features (Module) - Feature extractor which classifies input tasks.
• loss (Function, optional, default=CrossEntropyLoss) - Loss function which maps the cost of the events.
• distance_metric (str, optional, default='euclidean') - Distance metric between samples. ['euclidean', 'cosine']
• train_ways (int, optional, default=5) - Number of classes in for train tasks.
• train_shots (int, optional, default=1) - Number of support samples for train tasks.
• train_queries (int, optional, default=1) - Number of query samples for train tasks.
• test_ways (int, optional, default=5) - Number of classes in for test tasks.
• test_shots (int, optional, default=1) - Number of support samples for test tasks.
• test_queries (int, optional, default=1) - Number of query samples for test tasks.
• lr (float, optional, default=0.001) - Learning rate of meta training.
• scheduler_step (int, optional, default=20) - Decay interval for lr.
• scheduler_decay (float, optional, default=1.0) - Decay rate for lr.

References

1. Snell et al. 2017. "Prototypical Networks for Few-shot Learning"

Example

123456

tasksets = l2l.vision.benchmarks.get_tasksets('mini-imagenet')features = Convnet() # init modelprotonet = LightningPrototypicalNetworks(features, **dict_args)episodic_data = EpisodicBatcher(tasksets.train, tasksets.validation, tasksets.test)trainer = pl.Trainer.from_argparse_args(args)trainer.fit(protonet, episodic_data)

[Source]

Description

A PyTorch Lightning module for MetaOptNet.

Arguments

• features (Module) - Feature extractor which classifies input tasks.
• svm_C_reg (float, optional, default=0.1) - Regularization weight for SVM.
• svm_max_iters (int, optional, default=15) - Maximum number of iterations for SVM convergence.
• loss (Function, optional, default=CrossEntropyLoss) - Loss function which maps the cost of the events.
• train_ways (int, optional, default=5) - Number of classes in for train tasks.
• train_shots (int, optional, default=1) - Number of support samples for train tasks.
• train_queries (int, optional, default=1) - Number of query samples for train tasks.
• test_ways (int, optional, default=5) - Number of classes in for test tasks.
• test_shots (int, optional, default=1) - Number of support samples for test tasks.
• test_queries (int, optional, default=1) - Number of query samples for test tasks.
• lr (float, optional, default=0.001) - Learning rate of meta training.
• scheduler_step (int, optional, default=20) - Decay interval for lr.
• scheduler_decay (float, optional, default=1.0) - Decay rate for lr.

References

1. Lee et al. 2019. "Meta-Learning with Differentiable Convex Optimization"

Example

123456

tasksets = l2l.vision.benchmarks.get_tasksets('mini-imagenet')features = Convnet() # init modelmetaoptnet = LightningMetaOptNet(features, **dict_args)episodic_data = EpisodicBatcher(tasksets.train, tasksets.validation, tasksets.test)trainer = pl.Trainer.from_argparse_args(args)trainer.fit(metaoptnet, episodic_data)