GPU track at Stanford

Table of contents:

Chapter 1: Running the GPU starter KIT on the server
Chapter 2: Explore the code
Chapter 3: 
Putting files on the server

IMPORTANT: Always backup your code on your local compter.

FAQ (typical problems):
1. What is the URL of the challenge website?
2. Where is the data?
3. Where is the sample code?
4. How do I connect to the GPU server?
ssh -i hackathon.pem ubuntu@YOUR_DNS_SERVER_ID
Omit -i option if the key hackathon.pem is in your key ring (this can be done with "ssh-add -K hackathon.pem ")
Download hackathon.pem.
5. How do I get my DNS server ID?
Each group should have its own, ask the organizers if you did not get it by email.
6. I ran the code on the AWS server, but cannot see the predictions, where are they?
The prediction zip file (a .zip containing the sample code + predictions for each of datasets, ready for a submission) is created, 
in a directory ABOVE the directory with code.
7. Where do I put the data folder?
Not in the folder with the code (since it is included as a part of the submission). Put it somewhere else (above the directory with code). 
We do not want you to upload all datasets to the platform in addition to the predictions.
3. What should the .zip contain (how does a good submission look like) ?
The .zip should contain:, metadata, res/, lib/

--------------------  Introduction
At the bootcamp, you will be able to login to our common server on AWS using a key that we will provide to work on a GPU. 
Chapter 1 describes how to connect to the server and run the code, chapter 2 how to explore the code.
Chapter 3 describes what files to put on your computer to get ready to start 
(you should not need this if you use the servers we provide, they are pre-configured). 
However, if you would like to work on your own GPU server (also before/after the hackathon), we prepared instructions for you: you rather work on your own laptop equipped with only a CPU,

--------------------  Chapter 1: Running the GPU starter KIT on the server

0. Log into your GPU server:
Username is 'ubuntu', so you should login as ubuntu@YOUR_DNS_SERVER_ID, authorizing using the key. You need to save your private key in a safe place on your computer. The key is called:
It can be downloaded from this page (see bottom).
You will need to use the command line on your local computer to connect to the GPU server. To have a console in Windows, you may want install git:, then right-click 'Git Bash'.

For convenience, you can add a key to your key ring:
LocalComputer> chmod 0400 hackathon.pem
now it is ready to use to store it in your key ring do:
LocalComputer> ssh-add -K hackathon.pem
Now you can do (remember to replace DNS with the proper IP):
LocalComputer> ssh ubuntu@YOUR_DNS_SERVER_ID

If you cannot add a key to your key ring or have any other problems, you can do just:
LocalComputer> ssh -i hackathon.pem ubuntu@YOUR_DNS_SERVER_ID

EXAMPLE (hackathon.pem key in the current folder):
"ssh -i hackathon.pem"

1. Check your GPU installation:
We installed everything for your: 
GPU> ls
Data  hackathon
You should see two directories Data and hackathon
GPU> ls Data
christine  jasmine  madeline  philippine  sylvine
You should see the 5 datasets of Round 1 (both zipped and unzipped).
GPU> ls hackathon
You should see a directory gpu/
GPU> ls hackathon/gpu  lib  metadata  README.txt
You should see all the code pre-installed.

You may want to save a copy of the original code for further reference.

2. Run the code:
GPU> python
The results will end up in hackathon/gpu/res/.
For the .zip, go one level up:
GPU> cd ..
The code also creates in the directory hackathon a zip file ready to submit (it is a bundle of all the contents of the gpu subdirectory, its name look like
It will include the code as well. For submission, you have to keep all your files in the folder including the code.

3. Make a submission:
Download the zip file to your computer. Execute the scp command below.
LocalComputer> scp ubuntu@YOUR_DNS_SERVER_ID:/home/ubuntu/hackathon/[NAME_OF_RESULTS].zip ./
LocalComputer> scp -i  [my_path]/X.pem ubuntu@YOUR_DNS_SERVER_ID:/home/ubuntu/hackathon/[NAME_OF_RESULTS].zip ./

*** WARNING! ****
SCP five details checklist -- HERE REMEMBER TO:
1. Open a new console on your local computer (do not type the command in the AWS server shell!)
2. Change [NAME_OF_RESULTS] to the exact name of your zip.
3. Provide the proper path, i.e. submititurte 'home/ubuntu/hackathon' to wherever you put your result zip file. You can type 'pwd' in the GPU server shell to see the current path.
4. provide a filename (also path) where is located the [my_path]/X.pem key in your local computer.
5. Change YOUR_DNS_SERVER_ID to the proper server address.

To upload a file from your computer to the server, you need to swap the source and destination, e.g:
scp -i X.pem /C/local/path/ ubuntu@YOUR_DNS_SERVER_ID:/home/ubuntu/[mydir]/


4. Submit results to Codalab
Submit [NAME_OF_RESULTS].zip to Round 1 Stanford hackathon
Go to the Participate tab, then to the link "View/Submit results". Always include a description with your zip file.
Refresh to check the status. When the submission status turn to "Finished" check your submission in the leaderboard by looking up the "Results" tab.

--------------------  Chapter 2: Exploring the code
0. Windows users may want to install git:

1. Get familiar with Neural Networks in Theano
To able to work during the workshop read at least these Theano tutorials (logistic regression, mlp):

2. Get familiar with the code.
Your task is to make modifications to in the lib/ subdirectory of the gpu code. Note: the code is already uploaded to your GPU server.
run_nn is called from
The is the main script. In, you should not need to make changes above line 231:
    # ================ @CODE SUBMISSION (SUBTITUTE YOUR CODE) ================= 

Most of the code should remain unchanged. In particular:
D = DataManager(basename, input_dir, replace_missing=True, filter_features=True, verbose=verbose)
Just uploads the data. Keep as is.

Anything you may need to modify is below line 280:
GPU = True
Note: to run the code on CPU, just change this to:
GPU = False

Leave everything that has to do with time budget. The GPU example uses just 2 cycles (in the loop while  cycle <= 1:). 
The first one (cycle = 0) calibrates the time budget. We use for most learning machines ensemble methods voting over n_estimators base estimators. 
Cycle 0 computes how many estimators we have time to compute. Cycle 1 trains all the estimators we have time to compute. For Neural Networks,
this is translated to number of epochs.

For the GPU track we are only running binary classification examples from Round 1 of the AutoML challenge. So all you should care about is line 300:
Y = run_nn.fit_predict(['X_train'],['Y_train'], [['X_valid'],['X_test']], n_epochs = n_estimators/10)
Go to and note that there are default parameters that you can change:
batch_size = 100

 There is a safety feature, the number of epochs is maxed to 50 (lines 18-19):
    if n_epochs > 50:
        n_epochs = 50
You can remove that if you know what you are doing.
Go to and find where the architecture of the neural network is defined.
Line 130 and following, 2 layers are defined.
        self.hiddenLayer = HiddenLayer(
        self.hiddenLayer2 = HiddenLayer(
Notice how the number of outputs n_out is identical to the number of inputs n_in in the next layer. 
Notice that hidden1 and hidden2 are the number of hidden units supplied as arguments in the constructor of the NN object:
def __init__(self, rng, input, n_in, n_hidden1,n_hidden2, n_out):

1. Modify the default arguments, e.g. change the learning rate in
Y = run_nn.fit_predict(['X_train'],['Y_train'], [['X_valid'],['X_test']], n_epochs = n_estimators/10, learning_rate=0.1)
Similary, change the regularizations and the batch size.

2. Check the performance when using different number of epochs (like 20..500, currently 50, edit gpu/lib/ file, line 19)
For example:
    if n_epochs > 500:
        n_epochs = 500
You can also remove these line and control the time budget from
If on line 112 debug_mode = 0, then the time budget used is that of the "info" file of the datasets.
If on line 112 debug_mode = 1, the maximum running time on line 120 is used:
max_time = 90 means that the code runs in less than 90 sec per dataset.

3. In the provided NN, there are 20 and 10 units in the two hidden layers. Check other combinations (e.g. 50, 50), lines 79, 80.
    classifier = NN(
        n_in= len(data[0]),

4. Modify the number of layers of the neural network  NN in Add for example one layer. 
This means adding one argument to the constructor:
def __init__(self, rng, input, n_in, n_hidden1,n_hidden2, hidden3, n_out):
and one extra layer:
        self.hiddenLayer3 = HiddenLayer(
You also need to modify the logistic regression layer: change 2 for 3:
        self.logRegressionLayer = LogisticRegression(
Also add the extra layer information in the regularization terms L1 and L2:
            + abs(self.hiddenLayer3.W).sum()
            + (self.hiddenLayer3.W ** 2).sum()
Finally,  in self.params, also add:
+ self.hiddenLayer3.params

5. Experiment with all datasets separately (use different NNs for each dataset)

--------------------  Chapter 3: Putting files on the server
If you use the AWS servers that we assigned to your group THE FILES SHOULD ALREADY BE AT THE RIGHT PLACE.
So you do not need to perform the step below. These instructions are useful only if you have set up your own server.

1. You can get datasets from the CONTEST WEBSITE (click participate>get data).
Alternatively, you may use wget for each of the 5 datasets to download the datasets to the current folder if you prefer to, e.g.:
Put the 5 folders to the Data directory, so the path to the above dataset is: 
Put the datasets in a directory outside your code

2. Put the SAMPLE CODE it in another 'gpu' folder.

3. The program packs the directory with code and predictions/results together, creating a .zip that can be submitted 
to the competition on the codalab. The data should not be included, which is why it should not be in the code directory. 
You will need to change the default_input_dir and default_output_dir in to point to your data and result directories.

Isabelle Guyon,
Jul 30, 2015, 5:42 AM
Isabelle Guyon,
Jul 30, 2015, 5:42 AM
Isabelle Guyon,
Aug 7, 2015, 12:17 AM