"Lesson 0" discussion

Just discovered we never actually created a forum thread for discussing “lesson 0” - so here one is, for those folks as they go through the MOOC. There’s some rather sparse notes at http://wiki.fast.ai/index.php/Lesson_0 . Frankly most of the material from this “surprise lesson” I think we covered elsewhere in the course. But I still think the idea of pooling and learning from simple manual features is pretty interesting :slight_smile:

So… if you have any questions or comments about this lesson, here’s the place for them!

Can someone share a link to the OpenAI research he mentioned in the video - Regarding supervised classification with 50 examples and ensembles on MNIST with 1.4% error rate? I wasn’t able to find it. It would be of great help if someone could do so.


Edit : Found it - Improved Techniques for Training GANs. Putting it here just in case anyone else is also looking for it.


Just want to say I really like lesson 0!

I deleted my question because the second half of Lesson 3 answered all of them. :slight_smile:

Where can we get the Notebook shown in the video?

BTW, I added a link back to this in the YouTube video at

Would also like this!

Does anyone know where that convolution tutorial with a guys face he showed in the web browser for a second is?

Do you mean this?

1 Like

i‘ve a question for the “train.npz”file in con-intro.ipynb ,how can I get it. thanks!

@jeremy The really question is that where can we download “MNIST_data” which input by “convolution-intro.ipynb”?

I have been successfully working through the course for a while… was using AWS instance last night to visualize some of the filter activations and try out batch norm for the first time.

However, all of a sudden (this morning), jupyter seems to have disappeared… this is the error message I’m getting in my terminal, and neither conda nor jupyter are recognized as commands. Any idea what might have happened or how to fix it?

nevermind, had sourced my ~/.bash_profile in trying to add an alias; when I resourced ~/.bashrc, am back in business

1 Like

Hi @yousheng,

There is a cell that is commented out:

from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("MNIST_data/")
images, labels = mnist.train.images, mnist.train.labels
images = images.reshape((55000,28,28))
np.savez_compressed("MNIST_data/train", images=images, labels=labels)

As you can see, this gets the data from Tensorflow. You can also use Keras to get the data.


Got it, thanks a lot!

1 Like

The notebook used in the video is available here.
The MNIST image data can is available here.

I’m on a Windows based machine using GitBash for my shell. When I run, pip install asw cli, I get the following message:

Could not find a version that satisfies the requirement cli (from versions: ) No matching distribution found for cli

I googled the error but have been unsuccessful and am unsure about how to proceed. Does anyone know what I’m missing here?

I’m new to aws and also bash so I apologize if this is a relatively elementary question.

Thanks in advance!

When I get to this line in the convolution-intro notebook, I don’t get the same type of plot that the video shows:


Mine looks like this:

But the video looks like this:

Not sure what I’m doing wrong. I’ve had to modify the notebook a bit, because the MNist set was coming in as 60000,1,28,28.

1 Like

hello :slight_smile:
I am using Jupyter Notebook that is a part of Anaconda and the matplotlib version is 1.4.1.
In the notebook, when I execute:
rc(‘animation’, html=‘html5’)
rcParams[‘figure.figsize’] = 3, 6

I get a key error:
KeyError: ‘animation.html is not a valid rc parameter.See rcParams.keys() for a list of valid parameters.’

Also, animation ani is not executed at the end.

Any idea about how I should fix the problems?


export PATH=~/.local/bin:$PATH
zsh: no such file or directory: /.bash_profile
how work out?

I found the “lesson 0” notebook on GitHub where all the other fast.ai course notebooks are.
Here the URL: https://github.com/fastai/courses/tree/master/deeplearning1/nbs
Look for a file convolution-intro.ipynb

@SpaceCowboy850 I had similar problem.

I got my initial MNIST dataset using Keras function mnist.load_data
It by default downloads this data set: https://s3.amazonaws.com/img-datasets/mnist.npz

In the set you get 60000 images, so you need to delete the first 5000 to adjust it to the shape/size Jeremy works with.

Another point is that the pixel values in the Keras-downloaded set are coded as integers between 0 and 255. Pixels in Jeremy’s data set are all floats between 0.0 and 1.0. This is what I believe is causing your image to be all spotty rather than seeing smooth transitions.

To adjust to format Jeremy uses you can divide the data by 255 to convert the values from 0 - 255 to 0.0 – 1.0.
Here the code to adjust the downloaded data. It has resolved my spotty image issue:

# Delete the first 5000 images. After that the set should start with number 7
images = np.delete(images, slice(0,5000), axis=0)
labels = np.delete(labels, slice(0,5000))

# Convert the array type from int to float
images = images.astype(np.float32)

# Modify the data from 0 to 255 range to 0.0 to 1.0 range
images = images / 255

1 Like