AI Hackathon

With BRAIN NTNU, Telenor and Norwegian Open AI Lab

Time: 1st of March through 3rd

Place: R2, Gløshaugen NTNU, Trondheim

Timeline:

Friday:

• 15:15 - Presentation of assignment
• 16:00 - Competition begins, R2 opens
• 19:00 - Dinner

Saturday:

• 12:00 - Competition day 2 begins, R2 opens
• 14:00 - Lunch
• 18:00 - Dinner
• 23:59 - Deadline for delivery

Sunday:

• 13:00 - Awards ceremony

The assignment

Monitoring and predicting the air quality level (NO_x, PM₁₀ and PM_2.5)​ in the city and making it available to the public would increase the awareness of people living in the city about how good their environment is. From that, it would engage them to be part of solutions to reduce the pollution in the city. Monitoring and forecasting air quality would also benefit decision makers in the city to propose proper and systematic solutions to reduce the emission, such as reducing cars in different time of a day or week. The main objective of this hackathon is to build an application for air quality monitoring and prediction in Trondheim area for those purposes. With the hackathon, students will have ability to work with NB-IoT sensors developed by Telenor and work with data (NO_x, PM₁₀ and PM_2.5)​ collected and available at AI-lab. The students can also utilize the IoT analytics platform setup by Telenor to AI-lab to facilitate the data processing and data visualization.

Telenor is the second biggest IoT provider in Europe with more than 10 millions IoT devices for more than 100 customers. With ambition to empower society, we are currently helping to build smart municipalities by using IoT technologies. Working on this project, students will have an opportunity to understand Telenor's business, learning more about IoT, Big Data and AI/Machine Learning.

Relevant articles:

• Road dust and PM₁₀ in the Nordic countries
• Pavement wear and airborne dust pollution in Norway
• Norway investigating solutions for tackling road dust microplastics

Datasets

Dataset 1: NILU (2014 - 2019)

Name: NILU_Dataset_Trondheim_2014-2019.csv

Table-structure:

	Bakke kirke	Bakke kirke	Bakke kirke	Bakke kirke	Bakke kirke	E6-Tiller	...
timestamp	NO	NO2	NOx	PM10	PM2.5	NO	...
2013-12-31 23:00:00	8.940542	31.059213	44.720362	208.45	155.4	6.907422	...

Dataset 2: YR Weather Statistics (2014 - 2019)

Name: YR_Dataset_Trondheim_2014_2019.json

See this for an example of data in table

Format:

[
  {<date>: [{
	 [{'temperature': {
	   'measured': <measured_temperature>, 
	   'max': <max_temperature>, 
	   'min': <min_temperature>}, 
	   'timestamp': <timestamp>, 
	   'precipitation': <precipitation>, 
	   'humidity': <humidity>},
    },...]
  },...
]

To parse the datasets you can use for example pandas

Preliminaries

Docker

IMPORTANT:

We do know that Docker provides root-access. In spite of this, we trust you to behave appropriately. Any violation or misuse, such as touching other teams' containers and files or cryptomining, will lead to disqualification and expulsion from the AI Lab's processing sources.

Docker is installed on the HPC, and will be used during the event. On this link you can fin NTNUs Docker guide. For more documentation see https://docs.docker.com/. Advisors from the AI Lab and Telenor will be guiding you, if needed. Please use Dockerfile in this repository as a template.

We have created an example image from the Dockerfile in this repo, uploaded to both servers. You are welcome to use this to build your group's Docker container. To do this you can for example use this:

$ nvidia-docker run -dit --name grXX henrikhoiness/hackathon

, where XX is your groupname.

Sharing computational power - MANDATORY FOR ALL TEAMS

Since we are all sharing NTNUs HPC we have to distribute the processing power amongst all teams. You do this by setting the fraction of GPU memory to be allocated when construct a tf.Session (if using Tensorflow, otherwise consult with one of the advisors) by passing a tf.GPUOptions as part of the optional config argument:

# We are starting the event, with 50% of the GPU to each team

gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.5)
session = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))

To check how occupied your allocated GPU is, use watch nvidia-smi. If the GPUs are too occupied we will find a solution so that each team get to train their network.

Downloading the datasets

For downloading the datasets from your assigned anakin to your computer, please type the command below on you local machine.

$ scp -r <user>@<anakin-machine>.idi.ntnu.no:/work/hackathon/datasets <dataset-destination file>/

Connecting to the server

Use the command below with your assigned user and password to connect to the server:

$ ssh <user>@<anakin-machine>.idi.ntnu.no

Home directory at server

Your home directory where you can keep your files are located at cd ~.

Files of interest

• Dockerfile
• datasets/
• nilu_data.py
• yr_data.py

Categories for evaluation

• Most innovative idea
• Best use of AI on the data (method/application)

Computational power

For processing we are using IDIs new GPU-nodes; anakin01 and anakin02:

• GPU: 2x NVIDIA Tesla V100 32GB
• CPU: 2x Intel Xeon Gold 6132 2.6G, 14C/28T, 10.4GT/s 2UPI
• RAM: 24x 32GB = 768GB

Jury

• Kerstin Bach
• Kenth Engø-Monsen
• Sigmund Akselsen
• Espen Sande Larsen
• Massimiliano Ruocco
• Hai Thanh Nguyen
• Arjun Chandra