Currently, keras.applications hosts a total of 38 fully-trained models that can be used for any kind of image task without spending any time on training. Due to its presence and integration with other Keras tools, it makes for a fully-integrated experience in developing with them.
We explore the vast options available to users for developing cutting-edge applications powered by ML using TensorFlow and Keras, and how they can be prepared for inference/fine-tuned for specific use-cases. Following model-building, we explore options to deploy the model using simple services like PythonAnywhere.
In this session, we will start with what the use of the 'keras.applications' package is, followed by an explanation of the different types of pre-trained models present in the package, what is the use-case and specialty of each of them, and a thumb rule on what to use for which problem. We then walk through a use-case for simple image classification on a multi-class dataset and follow it up by a complete hands-on demo of fine-tuning and displaying the same, on-stage. Next, we show how to prepare the model for strict inference and predictions by exposing Flask-based APIs for the model, followed by a demo of putting it up on a demo platform like Google Cloud Platform or PythonAnywhere.
After this talk, an attendee should ideally be able to make use of pre-trained models from native Keras and be able to deploy models for inference independently on any given platform with the right resources.
This session is for anyone who has simple/preliminary knowledge of Deep Learning (What are neural networks, What are Convolutional Neural Networks, What is Keras) as well as for beginners to Keras who want to be able to prototype/build models faster and make them available to the public to use.
Students and working professionals are all welcome, alike!