GSoC

Repository for Google Summer of Code 2019

Evaluation	Task	Link	Status	Pull Request
E1	Autoencoder Migration	Here	Complete	#68
E1	Boosted Trees Migration	Here	WIP - Minor Bug Fixes	Branch
E1	Mask R-CNN Migration	Here	Part Migrated Working on Testing and Migration Guide	-
E1	Custom ResNet with TinyImageNet - Part 1	Here	Model Demo Ready Working on CLR, Cutout/Occlusion and Progressive Resizing	Branch
E1	Hyperparameter Search	Here	WIP - Keras Tuner, tf.keras wrappers	Branch

RetinaNet (Was last updated on July 13, 2018) - Medium blogs, Focal Loss paper by Yi-Lin et al, Alpha weighted Focal Loss
GANs - The TFGAN repo has a few varieties of GANs but they were updated 7-11 months ago. So, I looked into and worked on SRGANs, StackGANs, CycleGANs to add a folder for them
Training strategies from benchmarks.ai and Jeremy Howard's blog including SuperConvergence, Progressive Resizing, Exponential CLR, LR decay...etc.
Adding tf.data pipelines to load images apart from using flow_from_directory with keras' ImageDataGenerator class
Will be looking into the current implementations of wide_deep and boosted trees to recreate them. (Maybe towards the end of next month)

Adding tf.distribute.startegy or tf.distribute.MirroredStrategy for scoping tf.keras models for distributed leaning and adding tf.function decorators wherever required to optimise the execution.
Subclassing the models and layers is considered best practice. So, will change some of them from The Functional API to subclassing API
Applied for beta access to use KeasTuner to tune parameters quickly and effectively