align with PL 1.0

0.2.5

Update __init__.py

Reinforcement learning

[0.2.3] - 2020-10-12

Added

• Enabled PyTorch Lightning 0.10 compatibility (#264)
• Added dummy datasets (#266)
• Added KittiDataModule (#248)
• Added UNet (#247)
• Added reinforcement learning models, losses and datamodules (#257)

pertained models

[0.2.1] - 2020-09-13

Added

• Added pretrained VAE with resnet encoders and decoders
• Added pretrained AE with resnet encoders and decoders
• Added CPC pretrained on CIFAR10 and STL10
• Verified BYOL implementation

Changed

• Dropped all dependencies except PyTorch Lightning and PyTorch
• Decoupled datamodules from GAN (#206)
• Modularize AE & VAE (#196)

Fixed

• Fixed gym (#221)
• Fix L1/L2 regularization (#216)
• Fix max_depth recursion crash in AsynchronousLoader (#191)

Self Supervised Learning

[0.1.1] - 2020-08-23

Added

• Added Faster RCNN + Pscal VOC DataModule (#157)
• Added a better lars scheduling LARSWrapper (#162)
• Added CPC finetuner (#158)
• Added BinaryMNISTDataModule (#153)
• Added learning rate scheduler to BYOL (#148)
• Added Cityscapes DataModule (#136)
• Added learning rate scheduler LinearWarmupCosineAnnealingLR (#138)
• Added BYOL (#144)
• Added ConfusedLogitCallback (#118)
• Added an asynchronous single GPU dataloader (#1521)

Fixed

• Fixed simclr finetuner (#165)
• Fixed STL10 finetuner (#164)
• Fixed Image GPT (#108)
• Fixed unused MNIST transforms in tran/val/test (#109)

Changed

• Enhanced train batch function (#107)

Initial release

[0.1.0] - 2020-07-02

Added

• Added setup and repo structure
• Added requirements
• Added docs
• Added Manifest
• Added coverage
• Added MNIST template
• Added VAE template
• Added GAN + AE + MNIST
• Added Linear Regression
• Added Moco2g
• Added simclr
• Added RL module
• Added Loggers
• Added Transforms
• Added Tiny Datasets
• Added regularization to linear + logistic models
• Added Linear and Logistic Regression tests
• Added Image GPT
• Added Recommenders module

Changed

• Device is no longer set in the DQN model init
• Moved RL loss function to the losses module
• Moved rl.common.experience to datamodules
• train_batch function to VPG model to generate batch of data at each step (POC)
• Experience source no longer gets initialized with a device, instead the device is passed at each step()
• Refactored ExperienceSource classes to be handle multiple environments.

Removed

• Removed N-Step DQN as the latest version of the DQN supports N-Step by setting the n_step arg to n
• Deprecated common.experience

Fixed

• Documentation
• Doct tests
• CI pipeline
• Imports and pkg
• CPC fixes