Deep Learning Basics with Python, TensorFlow and Keras | Hacker News Reader

Deep Learning Basics with Python, TensorFlow and Keras

pythonprogramming.net

160 pointsmrleinad12 comments

Sigh, I know there is interest in learning, but there are a few things that (IMHO) don't follow the best practices.

For example - pickle for saving numerical data. Please, use h5py. Pickle can save any Python objects... and execute any (also - not-python) code. Or in general - you load a pickle, and cannot guess what had happened. With h5py - you have a tree of numerical objects, nothing more, nothing less.

Fo those reasons, I use pickle only when there is no other way to go - i.e. Python objects that cannot be serialized into anything language-agnostic. For saving purely numerical data, h5py is my way to go (http://docs.h5py.org/en/latest/quick.html).

Check out Jeremy Howard and Fastai. I like sentex, their videos are good, but Jeremy is another level.

http://course.fast.ai/index.html

I have used this website when I had itch about doing Monte-Carlo. I went through the course at a good pace. The beauty was you can follow the video or the text or both depending on your style of learning.

I will definitely check the deep learning course.

note: I am not a python programmer.

Can anyone recommend good books for learning deep learning with python?

For a very practical introduction to deep learning with Keras, I recommend Deep Learning with Python by François Chollet, https://www.manning.com/books/deep-learning-with-python.

For a general context, my post "Learning Deep Learning with Keras" https://p.migdal.pl/2017/04/30/teaching-deep-learning.html or for a practical quick start - "Starting deep learning hands-on: image classification on CIFAR-10" https://deepsense.ai/deep-learning-hands-on-image-classifica....

Thank you!

In general I prefer reading technical material (to really learn something) from a piece of paper rather than from a screen. I'll order the book :)

Try the CS231n course from Stanford. Assignments, solutions, slides and lectures are available. Proper grad level deep learning course with excellent notes. The assignments are IMO, the best way to get good and in-depth knowledge of deep learning.

For books, Ian Goodfellow's deep learning is the best theory (as theoretical as Deep Learning gets ) oriented book.

I know a machine learning engineer from a certain massive lending company. I asked him “How do you decide which strategy to use and which algorithms to pay attention to? There are so many, and it’s hard to figure out which ones are production grade.”

He said gradient boosting. I pointed out that that implies there would be some period of time where the algorithm makes bad decisions. He said yup, that’s why it’s hard to start a lending company. Gotta find the data on the bad performers.

Apparently their original decision model was just a hand crafted set of if-statements.

I suspect most ML endeavors that seem impressive have origin stories like that, but it’s only anecdata.

> He said gradient boosting. I pointed out that that implies there would be some period of time where the algorithm makes bad decisions.

How are the first and second sentences related? Would other algorithms make better decisions?

Sure, self-play. Generate the data and search through the problem space for optimal decisions before they come up.

At the end of the day it’s a complex form of heuristic.

Even if it is an actual algorithm that is provably optimal (in some sense), it is probably slave to accurate and complete inputs...

Reply on news.ycombinator.com