Best Practice Data Science Deep Learning Machine-Learning Scientific Libraries (Numpy/Pandas/SciKit/...)
Tensorflow is one of the most powerful machine learning libraries.
However, one could argue that it is as well one of the most convoluted libraries to utilise. Or it used to be.
Tensorflow has been updating its API, simplifying the pipeline while keeping it flexible. You could either predict your class label using a pre-trained canned estimator or doing some fancy operations with the inner tensors of your neural network. It's a choose your own adventure game.
But it's challenging to make sense of all the new changes. Fear not. Here, we learned what changes we should adopt.
This talk will walk you through the following new features:
- `Estimators`: complete representation of a model. We will learn about the different levels of abstraction of an `estimator`.
- `Datasets`: efficient input pipelines. We need large datasets to train models, and we should be conscious about how we are handling them.
- `Tf-hub`: a repository of Pre-trained modules. Starting each task from zero is not the best option. Particularly, when we have pre-trained models ready to boost the performance of our system.
The objective is at the end of this talk you will learn how to design a neural network following the best practices.
Basic knowledge of neural networks and the naming conventions in Tensorflow will be useful for understanding this talk, but I will try to minimise the gibberish to a local minimum.
Type: Talk (30 mins); Python level: Intermediate; Domain level: Intermediate
Oh hello! I'm Mai, and I'm a PhD student, and I research machine learning algorithms in different fields but mostly for Natural Language Understanding in Social Media, using, of course, Python, because the Neural Networks might be complex but your code shouldn't.
I 'm a board member of the Spanish Python Association, helped in organising several PyConES conferences, and a proud member of the Pyladies.
Empathy is my superpower; I know, it's the worst superpower ever.