Like 0.2.0 but with qpth made optional.

Added

• New vision example: MAML++. (@Theo Morales)
• Add tutorial: "Demystifying Task Transforms", (Varad Pimpalkhute)
• Add l2l.nn.MetaModule and l2l.nn.ParameterTransform for parameter-efficient finetuning.
• Add l2l.nn.freezeand l2l.nn.unfreeze.
• Add Adapters and LoRA examples.
• Add TasksetSampler, compatible with PyTorch's Dataloaders.

Changed

• Documentation: uses mkdocstrings instead of pydoc-markdown.
• Remove text/news_topic_classification.py example.
• Rename TaskDataset to Taskset.

Fixed

• MAML Toy example. (@Theo Morales)
• Example for detach_module. (Nimish Sanghi)
• Loading duplicate FGVC Aircraft images.
• Move vision datasets to Zenodo. (mini-ImageNet, tiered-ImageNet, FC100, CIFAR-FS, CUB200)
• mini-ImageNet targets are now ints (not np.float64).
• Swap family for variants in FGVCAircraft, as in MetaDataset.

v0.1.7

Added

• Bounding box cropping for Aircraft and CUB200.
• Pretrained weights for vision models with: l2l.vision.models.get_pretrained_backbone().
• Add keep_requires_grad flag to detach_module. (Zhaofeng Wu)

Fixed

• Fix arguments when instantiating l2l.nn.Scale.
• Fix train_loss logging in LightningModule implementations with PyTorch-Lightning 1.5.
• Fix RandomClassRotation (https://github.com/learnables/learn2learn/pull/283) to incorporate multi-channelled inputs. (Varad Pimpalkhute)
• Fix memory leak in maml.py and meta-sgd.py and add tests to maml_test.py and metasgd_test.py to check for possible future memory leaks. (https://github.com/learnables/learn2learn/issues/284) (Kevin Zhang)

v0.1.6

Added

• PyTorch Lightning interface to MAML, ANIL, ProtoNet, MetaOptNet.
• Automatic batcher for Lightning: l2l.data.EpisodicBatcher.
• l2l.nn.PrototypicalClassifier and l2l.nn.SVMClassifier.
• Add l2l.vision.models.WRN28.
• Separate modules for CNN4Backbone, ResNet12Backbone, WRN28Backbones w/ pretrained weights.
• Add l2l.data.OnDeviceDataset and implement device parameter for benchmarks.
• (Beta) Add l2l.data.partition_task and l2l.data.InfiniteIterator.

Changed

• Renamed and clarify dropout parameters for ResNet12.

Fixed

• Improved support for 1D inputs in l2l.nn.KroneckerLinear. (@timweiland)

v0.1.5

Fixed

• Fix setup.py for windows installs.

v0.1.4

Added

• FilteredMetaDatasest filter the classes used to sample tasks.
• UnionMetaDatasest to get the union of multiple MetaDatasets.
• Alias MiniImageNetCNN to CNN4 and add embedding_size argument.
• Optional data augmentation schemes for vision benchmarks.
• l2l.vision.models.ResNet12
• l2l.vision.datasets.DescribableTextures
• l2l.vision.datasets.Quickdraw
• l2l.vision.datasets.FGVCFungi
• Add labels_to_indices and indices_to_labels as optional arguments to l2l.data.MetaDataset.

Changed

• Updated reference for citations.

Added

• l2l.vision.datasets.CUBirds200.

Changed

• Optimization transforms can be accessed directly through l2l.optim, e.g. l2l.optim.KroneckerTransform.
• All vision models adhere to the .features and .classifier interface.

Fixed

• Fix clone_module for Modules whose submodules share parameters.

Added

• New example: Meta-World example with MAML-TRPO with it's own env wrapper. (@Kostis-S-Z)
• l2l.vision.benchmarks interface.
• Differentiable optimization utilities in l2l.optim. (including l2l.optim.LearnableOptimizer for meta-descent)
• General gradient-based meta-learning wrapper in l2l.algorithms.GBML.
• Various nn.Modules in l2l.nn.
• l2l.update_module as a more general alternative to l2l.algorithms.maml_update.

Fixed

• clone_module supports non-Module objects.
• VGG flowers now relies on tarfile.open() instead of tarfile.TarFile().