learnables/learn2learn

✨ Added

• New vision example: MAML++. (@[Theo Morales](https://github.com/DubiousCactus))
• Add tutorial: "Demystifying Task Transforms", ([Varad Pimpalkhute](https://github.com/nightlessbaron/))
• Add `l2l.nn.MetaModule` and `l2l.nn.ParameterTransform` for parameter-efficient finetuning.
• Add `l2l.nn.freeze`and `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](https://github.com/DubiousCactus))
• Example for `detach_module`. ([Nimish Sanghi](https://github.com/nsanghi))
• 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.

📋 Changes

• 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](https://github.com/ZhaofengWu))
• 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](https://github.com/learnables/learn2learn/pull/283)) to incorporate multi-channelled inputs. ([Varad Pimpalkhute](https://github.com/nightlessbaron/))
• 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](https://github.com/learnables/learn2learn/issues/284)) ([Kevin Zhang](https://github.com/kzhang2))

📋 Changes

• 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`.
• Renamed and clarify dropout parameters for `ResNet12`.
• + 1 more

🐛 Fixed

• Fix setup.py for windows installs.

📋 Changes

• `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`
• + 2 more

📋 Changes

• `l2l.vision.datasets.CUBirds200`.
• Optimization transforms can be accessed directly through `l2l.optim`, e.g. `l2l.optim.KroneckerTransform`.
• All vision models adhere to the `.features` and `.classifier` interface.
• Fix `clone_module` for Modules whose submodules share parameters.

📋 Changes

• New example: [Meta-World](https://github.com/rlworkgroup/metaworld) example with MAML-TRPO with it's own env wrapper. (@[Kostis-S-Z](https://github.com/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`.
• clone_module supports non-Module objects.
• VGG flowers now relies on tarfile.open() instead of tarfile.TarFile().

✨ Added

• A CHANGELOG.md file.
• New vision datasets: FC100, tiered-Imagenet, FGVCAircraft, VGGFlowers102.
• New vision examples: Reptile & ANIL.
• Extensive benchmarks of all vision examples.

📋 Changed

• Re-wrote TaskDataset and task transforms in Cython, for a 20x speed-up.
• Travis testing with different versions of Python (3.6, 3.7), torch (1.1, 1.2, 1.3, 1.4), and torchvision (0.3, 0.4, 0.5).
• New Material doc theme with links to changelog and examples.

🐛 Fixed

• Support for `RandomClassRotation` with newer versions of torchvision.
• Various minor fixes in the examples.
• Add Dropbox download if GDrive fails for FC100.