/Keras-model-zoo

Keras Model Zoo

This repository contains pre-trained models that are readily available to use in tf.keras or Keras. There are two things that we are focusing on right now: Computer Vision and NLP, each of which is a sub-package.

Objective and Motivation

Deep learning has really taken off in the last few years. Since AlexNet won the 2012 ImageNet competition, the amount of research and the papers/neural network architectures coming out has seen an exponential rise. It is also well known that transfer learning and fine-tuning both have been a great success and are used on a daily basis by Researchers and ML practitioners. In order to experiment fast, people rely on ready-made models. For example, if someone is trying to make some changes to the Transformer architecture, that person wants a ready-made model with a certain amount of flexibility. This not only helps saving time but it also encourages people to try more and more things.

Existing work

There have been many efforts made in this direction. For example,

  • keras-applications
  • slim-model-zoo
  • tensor2tensor
  • TFHub

Why a new keras_model_zoo then?

All of the existing works listed above have certain limitations.

  • keras-applications is limited to a very small number of models and doesn't provide pre-trained models for NLP, or Object Detection, or Semantic Segmentation.
  • TFHub is even more limiting. It doesn't provide the flexibility researchers need. As of now, it is well suited for inferencing or little changes but not much.
  • slim isn't the recommended API in TF2.0
  • tensor2tensor has its own learning curve and again, it doesn't provide all the functionalities

Requirements

  • The models should be compatible with both, tf.keras and Keras.
  • Models should be described as Lego blocks that are usable and which can be modified easily.
  • Models focusing on computer vision should go under KerasCV and others, focusing on NLP, should go under KerasNLP.
  • Please refer to the API signatures defined in each sub-package to for more details