TensorFlow Tutorials | Hacker News Reader

TensorFlow Tutorials

github.com

175 pointsbootload24 comments

I really like their high-level tutorials as in here: https://www.tensorflow.org/versions/0.6.0/tutorials/ (with descriptions and general intro to neural networks, not only - a specific library).

There are also Jupyter Notebooks: https://github.com/tensorflow/tensorflow/tree/master/tensorf... (used as exercises in the Udacity course: https://www.udacity.com/course/deep-learning--ud730).

This is a great start.

However, I love to attach narratives to problems, and have always enjoyed the coding competition questions where they tell a little scenario (albeit contrived) that usually have to do with something like a farmer and her cows in a barn or bees flying to pollen.

It seems to somehow engage my mind more with the problem.

Anyways, It would be great if there was some sort of 'problem description' and a background story context to really get the learner excited about applying the machine learning tool to solve real (fictional) life problems that are affecting real (fictional) people.

And just to dream, maybe a comic edition like codecomics does with react, or learning mysql with nostarchpress manga comics.

These look like machine learning cases - does anyone have any other examples - network design? Crypto?

Some other applications: https://github.com/aymericdamien/TensorFlow-Examples https://www.tensorflow.org/versions/0.6.0/tutorials/pdes/ind... https://www.tensorflow.org/versions/0.6.0/tutorials/mandelbr...

Does anyone have experience how TensorFlow work on Raspberry pi? I have an idea for a simple robot-car that's learns itself how to drive and and transport sensitive things (say, full glass of water). Would it put too much heavy load operations on Raspi 2?

you could get it running on a raspberry pi. it's quick to run image recognition and so forth. it's training it initially that takes a lot of computing resources.

I don't believe anyone has it working yet. I've been watching this issue for progress: https://github.com/tensorflow/tensorflow/issues/254

I would use NVidia Jetson TK1 or X1 instead, you can run same CUDA accelerated ML frameworks as elsewhere. Runs Ubuntu. TK1 is about $190 with 192 CUDA cores and enough IO for whatever you need. http://elinux.org/Jetson/GPIO

Train on a desktop/cloud GPU and use the trained network on the Jetson, you should get realtime perf for most simple autonomous driving related stuff.

Look into ROS and Ardupilot Rover as well for a head start into autonomous RC cars.

Here's a tutorial on getting Tensorflow running on Jetson - http://cudamusing.blogspot.co.uk/2015/11/building-tensorflow...

This is great!

If anyone's interested, yesterday we put up a fairly detailed guide to installing TensorFlow on EC2 GPU instances with Python 3.4.

http://eatcodeplay.com/installing-gpu-enabled-tensorflow-wit...

I'm trying to read the official docs for a computer vision gp and my mind froze. Will try these ones. Thanks :)

Half the tutorial is about how to set up virtualEnv and other config.

This is why I hate developing in python. I will just wait for go and node.js to come up with their own APIs for tensor flow.

You could write that API...

Use the docker image instead - https://www.tensorflow.org/versions/0.6.0/get_started/os_set...

I'd successfully finished the Andrew Ng Coursera class using Matlab on my old MacBook Air, and moved on to TensorFlow. Turns out Docker is just too much for my 2 Gb Mac - it totally hangs. I got it working pretty easily on a web IDE codeenv, but it has limitations and I couldn't load the sample files or use ipython notebook. I'd love to run through these tutorials on a remote hosted environment with minimal configuration work. Anyone have advice?

Nodejs, nodejs-legacy, npm, gulp, bower and tsd. Not a big deal of course.

"Half the tutorial is about how to set up virtualEnv and other config."

No, it's not. The README.md for the repo has a link to the tensorflow 'Download and Setup' instructions (in another repo).

The files in the tensorflow_tutorials repo are all code samples (and corresponding ipython notebooks) showing how to use Tensorflow.

Humbug! In my day you had to resolve linker dependencies, and god forbid if you mixed debug and release or one of the things you were linking against accidently set your process to single-threaded apartment mode, kids these days have it easy!

Or you could install it with easy install and pip. Literally 2 commands in terminal.

Virtual environments have their benefits, but they are far from necessary.

I discovered it was nearly impossible to set up pip on windows. It required installing virtual environment stuff first which was very difficult. I eventually discovered someone who had a binary wrapper of it.

Python on windows is a special case. Don't try to use pip/easy_install(setup tools) for anything that has extension modules, or needs compilation rather. Just find the binary distribution for your CPU architecture for windows and install using that.

Pip and setup tools were not made for a windows-type environment. I.e. Where you have don't library dependencies automatically installed and you require a compiler. Windows is a second-class citizen when it comes to those things, but it's getting better with cygwin and all that.

Pip comes installed by default with python 3.4 and 3.5. Not sure about the latest 2.7 builds, but I expect it's bundled there too.

EDIT: For the general case of Python on windows, it might be worth looking into [Anaconda][https://www.continuum.io/downloads].

Setup tools doesn't have distributions for 2.5 anymore. You have to download a "bootstrap.py" type of file and run that to get setup tools instaled.

I agree (about Python, not node.js). This was my introduction to Python. What a mess! Nothing quite worked, and the instructions were very incomplete. Install with or without sudo? System python, homebrew or something else? Python2 or python3? When pip installs libraries, then what's easy_install? How come I can't specify which python version to use for easy_install?

I tried pretty much every permutation, starting with the recommended one in the tutorial, and every one ended in a different error.

Don't get me wrong, the Ruby tooling situation for instance is probably just as bad, only I've gotten used to that by now.

> Don't get me wrong, the Ruby tooling situation for instance is probably just as bad

Under the hood, it's comparably horrible, but IME bundler does a much better job of papering over that horribleness for experienced users and newcomers alike.

Reply on news.ycombinator.com