TensorFlow Hub Module Creation and Support for TensorFlow.js

Training of High end Deep Learning Models takes a Long time to Train, Hours, Days and Even Weeks. So, as a Developer/researchers it slows down the creativity and implementation cycle. The Whole Idea of Transfer Learning provides a nice solution where a model trained on a similar dataset can be trained partially or fully, on the custom dataset, yielding similar result. TFHub is a platform by Tensorflow where pretrained models are shared, so that developers and researchers can use transfer learning on these models. In this Project I'll be working on the following tasks:

• Implement Tensorflow Hub sample for Cloud TPU
• Convert ShuffleNet from ONNX to TF Hub Module
• Implement and Train ESRGAN(Enhanced Super Resolution Generative Adversarial Network) from Scratch (https://arxiv.org/pdf/1809.00219.pdf)
• Implement GAN Distillation to compress ESRGAN (https://arxiv.org/pdf/1902.00159.pdf)
• Add Support for SavedModel 2.0 Functions in saved_model_cli
• Add Sample Notebook demonstrating usage of ESRGAN TF Hub Module
• Create proof of concept Video player for real time video frame Super Resolution (using "distilled" ESRGAN)

Github: https://github.com/captain-pool/GSOC