Lesson 8 Discussion

I’ll be putting material for lesson 8 (that is, lesson 1 of part 2) here. The video is now available, as are the slides.

We’ll be using pytorch in this class, so please run the following (once you have py3 and tensorflow installed) to install pytorch:

conda install pytorch torchvision cuda80 -c soumith

You’ll need to download the following data files for this lesson:

The class is based on the following papers:

I’ll edit this post as more information is available.

Slides and notebooks: http://files.fast.ai/part2/lesson1/


Anyone having luck getting pytorch installed on windows? From what I can see the windows port doesn’t exist yet although it looks like it’s on it’s way.

@jeremy How much of the course is built on pytorch? Is it primarily for lesson 8 or will we be using it extensively throughout?

I did a bit of digging and it seems the main difference between PyTorch and Tensorflow is that PyTorch supports the creation of dynamic computation graphs, which is i’m assuming the functionality we’ll be using.

Google has just released (2 weeks ago!) Tensorflow Fold which is their implementation of DCGs. I was hoping to maybe use that and port the code but it looks like it’s linux only right now as well.

Hi Jeremy,

I’m working on installing pytorch. I’m using my mac currently instead of AWS. Not sure if advised but planning to see how much I can do on my mac before getting an AWS instance.

I managed to install pytorch using this
conda install pytorch torchvision -c soumith

I was not able to install with the original command you had ( including cuda80). It gave my a package no found error.

I was wondering if this would cause issues for me down the road not having cuda80?

For easy copy/paste:

wget http://files.fast.ai/data/imagenet-sample-train.tar wget http://files.fast.ai/data/trn_resized_288.tar wget http://files.fast.ai/data/trn_resized_72.tar


Btw you could use axel for accelerated download (sudo apt-get install axel -y before using) which is much more faster. From platform.ai I’ve got x10 vs. wget.


From what I have read from the introduction thread, it seems that some of students of the part 2 may not be used to read scientific articles. A few years ago, my advisor gave me this article on how to approach them: How to read a paper

It is only two pages long but packed with good informations. I think it may deserve its own topic on the forum or a page on the wiki.


pytorch does not work on windows, but you can use Ubuntu Bash for Windows (must enable dev mode and install the feature) which works well but will not allow GPU access.

1 Like

Without cuda things will be a lot slower, but should work fine otherwise.

There will be quite a bit of pytorch - I’d suggest firing up an AWS instance at least for those parts. Tensorflow Fold is definitely not a replacement.

We’ll be spending a lot of time during the course on this, including reading through at least one paper together every week! :slight_smile: So don’t worry if you’ve never read a paper before, or if none of the ones listed above make any sense at this stage…


I had the same question. Well I focused on the introduction, summary and skimmed through the most of it.

r/ML discussion of the WGAN paper (w/ discussion between arjovsky and goodfellow):


1 Like

How should I spend this week to get the most out of Lesson 8?

Some elements in my mind:

  • Notebooks (reviewing and actively recalling them)
  • Papers (answering the four questions Jeremy mentioned)
  • What problem are they solving?
  • What is the general idea that they are using to solve it?
  • What kind of results are they getting?
  • What previous work are they building on?
  • Reading blogs and walkthroughs
  • Writing blogs and walkthroughs
  • Contributing to the wiki
  • Implementing things myself with:
  • Keras
  • TensorFlow
  • PyTorch
  • Numpy
  • Study groups
  • Explaining and distilling concepts and implementations with others
  • Brainstorming project ideas with teammates, and getting started

Which elements am I missing? Which are extra? And most important, how should I prioritize them?


-review notebooks
-Implement yourself and read blogs while doing so
-read papers
-write blog about your little experience/knowledge.
-contribute to the wiki
-contribute answers to the group
-study groups should be last, just to relax and have fun

Hi all, I’m getting issue downloading the HDF5 weight files for TF ordered VGG16 weights. I keep getting this error:

Exception: URL fetch failure on https://github.com/fchollet/deep-learning-models/releases/download/v0.1/vgg16_weights_tf_dim_ordering_tf_kernels_notop.h5: None -- [Errno 110] Connection timed out

If I got to the URL directly I get a connection timeout on my browser as well. Anyone else having issues with this? It’s a blocker for trying to run the notebook from yesterday…

I tried directly from this page too – no love.

Anyone have these weights stored somewhere so I can grab them? Or know of another place on the internets to get them?

Ok i think it’s cause of this: http://isitdownorjust.me/github-cloud-s3-amazonaws-com/

1 Like

Is the video of lesson8 posted some where?