Microsoft is partnering with UPS Master 2 Data Science to grant students free GPU compute time for the Deep Learning class.
Create a Microsoft account to connect to:
Try to login with the email address of your choice. The page should update to offer you the ability to create your account for that email address.
Note that some students have reported that some school email addresses failed to receive the confirmation code required to activate your account.
Once you can login to https://portal.azure.com please contact Olivier or Charles on slack (or by email) and give us the email address you used to login to the azure interface. Let's use the following slack room to setup accounts:
https://m2dsupsdlclass-2018.slack.com/messages/C8PTYNF9C/
You need to be a registered student of the master to access this slack channel.
We will grant you access to the "AZUREDEEPNETOG" subscription that has credits for approximately 20,000 hours of GPU compute time for the class.
Once we grant you the access, you will probably arrive on a page, that states a message similar to "No available application, please contact your administrator." Ignore this message and instead go back to https://portal.azure.com, log out and log back in to the portal interface, in the left-hand side toolbar, click at the bottom on "More services" and then "Subscriptions". You should see:
"AZUREDEEPNETOG"
If not, click on the top and right-hand side menu with your name and switch directory to the one named "ericmoulinesoutlook.onmicrosoft...". Then check again the subscription menu. If you still cannot access the subscription please contact us on slack.
We will used the Azure Deep Learning VM image that comes with everything pre-installed (Python, CUDA, CuDNN, jupyter, keras, tensorflow-gpu, etc).
You can use this image to build your own VM. When doing so, please name the resource group and the VM it-self as "firstname-lastname" (using only lowercase ACSII characters) to make it easy to know which VM is owned by you.
To deploy your own VM using this image, follow those steps in the portal:
- click "+ New" in the left hand size panel;
- search for "Deep Learning Virtual Machine" and click on the "Create" button;
- name: "firstname-lastname" (only lowercase ascii characters);
- OS type: "Linux";
- choose a username and a password (write them down, you will need them later);
- subscription: "AZUREDEEPNETOG" (should be the default);
- resource group:
- tick "create new" (if this is the first time you deploy a VM), then:
firstname-lastname(only lowercase ascii characters);
- location: choose "West Europe" should work;
- virtual machine size: "1x Standard NC6";
- use the default storage account;
- in the "Summary" panel, wait a bit for the validation and click "OK";
- in the "Buy" panel, tick the checkbox and click "OK" (after reading the boring legal text off-course);
If you get a QuotaExceeded error message: try again with another location.
Not all locations have NC6 virtual machines. Check the regions that have
NC-series here:
https://azure.microsoft.com/en-us/regions/services/
Feel free to pin this deployment to your dashboard to make it easy to find the next time you log in to the Azure portal.
You can leave the network configuration parameters to their default values.
If it does not work, please click on the top and righthand-side menu with your name and switch directory to the one named "ericmoulinesoutlook.onmicrosoft..." and then try again.
If it still does not work, please contact Olivier or Charles preferably on slack.
Once you are no longer using the VM, don't forget to stop it! (see below).
Once your VM is deployed (wait a couple of minutes for the VM to start and be ready to accept connections), copy the IP address and from the Azure portal info you should be able to ssh into it:
ssh username@ip-address
If your laptop is under Windows, feel free to use PuTTY to connect to your Azure VM on port 22.
Make sure to update setuptools:
pip install -U pip setuptools
The upgrade keras and tensorflow-gpu (the GPU version) and kaggle-cli
which we will need later
pip install -U tensorflow-gpu==1.4.1 keras kaggle-cli
Note that tensorflow 1.5 or later do not seem to work on that VM at the time of writing.
If you prefer the new experimental jupyterlab over the traditional jupyter notebook interface you can install the latest version with pip:
sudo /anaconda/bin/pip install -U jupyterlab
You should then be able to import keras from the python command in
your PATH:
$ python -c "import keras"
Using TensorFlow backend.
2018-02-08 16:12:58.121025: I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX AVX2 FMA
2018-02-08 16:12:58.361406: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1030] Found device 0 with properties:
name: Tesla K80 major: 3 minor: 7 memoryClockRate(GHz): 0.8235
pciBusID: d489:00:00.0
totalMemory: 11.17GiB freeMemory: 11.10GiB
2018-02-08 16:12:58.361473: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1120] Creating TensorFlow device (/device:GPU:0) -> (device: 0, name: Tesla K80, pci bus id: d489:00:00.0, compute capability: 3.7)
For some unknown reason this is sometime very slow, but apprently it gets better over time when you actually use the VM for heavy GPU computation (e.g. when training models in a notebook).
Feel free to use the tmux command to get a session that you can
re-attach to and launch long running commands such as jupyter notebook. Learn about tmux here: a quick and easy guide to tmux.
Upload your notebook either into your home folder /home/username.
Under Windows you can use PuTTY, under Linux and macOS you can use the
scp command in a terminal.
You can also clone the official repo:
git clone https://github.com/m2dsupsdlclass/lectures-labs
Then open a ssh tunnel to your Azure VM on port 9999:
ssh -L 9999:localhost:9999 username@ip-address
Alternatively use PuTTY to setup a ssh tunnel on port 9999 both on your local machine and the remote Azure VM.
Then on the VM (preferably under a tmux session), cd into the folder
where you uploaded the notebooks and type:
jupyter notebook --no-browser --port 9999
or alternatively:
jupyter lab --no-browser --port 9999
This should print the URL with a localhost:9999 address and a unique
security token. Copy the full URL into the URL bar of browser window
running on your laptop and you should be good to go.
In a shell on the VM you can launch the following command to monitor the usage of the GPU on your host
watch nvidia-smi
The ResNet50 model should be able to process 40+ images per second on
the NC6 VMs with a K80 GPU. Sometimes it is slower at the beginning but
the processsing speed tend to increase when you actively use the
machine.
In the Azure portal (https://portal.azure.com), please click on the stop button of your own VM whenever you do not need it anymore. It will be possible to restart it later with all your files there as you left them.
No compute credit will be used when the VM is stopped in the "deallocated" state.
FYI a single GPU (NC6) VM costs approximately $1 / hour. Feel free to use it as much as you want but don't waste it by leaving it launched when you don't use it anymore.