Neural networks on GPUs: cost of DIY vs. Amazon

I like to dabble with machine learning and specifically neural networks. However, I don’t like to wait for exorbitant amounts of time. Since my laptop does not have a graphics card that is supported by the neural network frameworks I use, I have to wait for a long time while my models git fitted. This is a problem.

The solution to the problem is to get access to a computer with a supported Nvidia GPU. Two approaches are to either get my own rig or rent one from Amazon. Which is cheaper?

Cost analysis

I will assume that I will train models on my machine (whether local or at Amazon) for two hours every day.

The Amazon p2 range of EC2 machines come with Nvidia K80 cards, which costs about 50.000 DKK. Already this analysis is going to be difficult; I will not buy a computer that costs 50.000 DKK just to train NN models. So, in this analysis I will be comparing apples to oranges, but that is how it is.

Cost of Amazon

The p2.xlarge EC2 instance has a single K80 GPU, which is at least as good as any rig I would consider buying.

The on-demand prie is $0.9/hour; the spot price about five times cheaper. Usage for two hours every day for a whole year costs 4.500 DKK for on-demand and 900 DKK for spot instances. However, the p2 instances is sometimes unavailable in the European spot markets.

Cost of DIY

What is the best GPU to get for a DIY machine learning rig? In 2016, Quora answers suggested that the Nvidia cards Titan X and GTX980TI would be best. Let’s go with that.

This is quite a bit more than 4.500 DKK and that is only for the graphics card. The finished rig would probably cost around 15.000 DKK (Titan) and 10.000 DKK (GTX).

The electricity also has to be factored in, plus that the cards are basically slower than the K80.

Best choice for increased usage

With increased usage the DIY approach will become cheaper than Amazon, albeit still a slower option. With usage of 5 or 7 hours/day the DIY approaches break even after a year.

Further reading

Build a deep learning rig for $800.

Ether mining – first attempt

This guide outlines my first (and failed) attempt at mining Ether on AWS. I will first show how I set up a GPU instance to mine Ether. Then I will conclude on the profitability of the whole endeavour in a simple way, i.e. profit = [value of generated Ether] – [AWS cost].

The guide is based on this, this, this and this.


Machine p2.xlarge (spot instance)
AMI ami-2cbf3e44
Miner cpp-ethereum
Result -$0.19/hour

Download blockchain

Log into AWS in us-east-1.

Launch an p2.xlarge EC2 machine. Set the bid price to $0.9. Use the image ami-2cbf3e44, which comes with the NVIDIA drivers etc. The instance should have 100gb disk space to hold the blockchain.

(SSH into the machine)

Install Ethereum:

sudo apt-get install software-properties-common
sudo add-apt-repository -y ppa:ethereum/ethereum-qt
sudo add-apt-repository -y ppa:ethereum/ethereum
sudo add-apt-repository -y ppa:ethereum/ethereum-dev
sudo apt-get update
sudo apt-get install ethereum

Start a screen session:


Run geth in the screen session, which begins to download the blockchain:


You can now detach from the screen session by hitting ctrl a and d and disconnect from the machine. Periodically check back on the progress of geth downloading the blockchain.

(After a while)

Reattach to the screen session and check the process:

screen -r

You will see an output that looks like the output below. When geth has caught up with the block chain, you will see messages with 1 block per line.

Mine Ether

(When block chain has caught up)

Now we will mine Ether using the default ethminer. This is not recommended according to, but we will see how bad it is.

Install the cpp-ethminer:

sudo apt-get update
sudo apt-get install cpp-ethereum

Benchmark the miner to measure its hash rate. You will need to run it twice. The first time it runs it builds a DAG. The second time it runs it will run the actual benchmark:

# Run twice
ethminer -G -M

I got about 19 MH/s (i.e. 19 million hashes per second) on the p2.xlarge instance.

Create a new screen window by hitting ctrl-a and c.

Now, mine some Ether in the new window. Use the -F option to send it to your account via

ethminer -F -G

The long string that begins with ‘0x7B…’ and ends right before ‘.’ is my Ether wallet. You should obviously use your own.

Now, you can detach from the screen session with ctrl-a and d and check back once in a while to see if it is still running. Also check if Ether accumulates in your wallet. Finally, you can check the dash board (mine) to see the progress.


Now, let’s tally the Ether that was generated and subtract the cost for the AWS instance.

The AWS spot price for one p2.xlarge is about $0.2/hour. According to ethermine, my rig generates about $0.01 worth of Ether per hour. That gives a negative result of $0.01 – $0.2 = -$0.19/hour.

The AWS p2.xlarge instance would need to be 20 x more effective to break even. Ethermine reports my effective hashrate as 19 MH/s. In conclusion, we need to squeeze at least 380 MH/s out of the p2.xlarge or find something else to do.

Python script for geocoding a text file

Assume that you have a file with some locations as text with one location per line.

For example, here are some school names in Copenhagen, Denmark, stored in schools.csv:

Hyltebjerg Skole
Ellebjerg Skole
Katrinedals Skole
Peder Lykke Skolen
Amager Fælled Skole
Tingbjerg Heldagsskole
Øster Farimagsgades Skole
Sankt Annæ Gymnasiums Grundskole
Lykkebo Skole
Randersgades Skole
Grøndalsvængets Skole
Sølvgades Skole
Skolen ved Sundet
Hanssted Skole
Den Classenske Legatskole
Tove Ditlevsens Skole
Lergravsparkens Skole
Vigerslev Allés Skole
Bavnehøj Skole
Ålholm Skole
Guldberg Skole
Husum Skole
Nyboder Skole
Vanløse Skole
Kirkebjerg Skole
Christianshavns Skole
Bellahøj Skole
Korsager Skole
Nørrebro Park Skole
Utterslev Skole
Skolen på Islands Brygge
Brønshøj Skole
Kirsebærhavens Skole
Rødkilde Skole
Vesterbro Ny Skole
Blågård Skole
Sønderbro Skole
Højdevangens Skole
Oehlenschlægersgades Skole
Vibenshus Skole
Valby Skole
Rådmandsgades Skole
Tagensbo Skole

Here is a script,, that will attempt to geocode each location in an input stream. It prints CSV output to stdout with the fields input_line, input_line_no, result_no, place, latitude, longitude:

from geopy import geocoders
import sys
import time
import pdb
geocoder = geocoders.GoogleV3()
SEPARATOR='|'  # can also use tab. Comma is bad, since the place will most likely contain a comma.
dummy = ['', ['', '']]
i = 0
header = ['input_line', 'input_line_no', 'result_no', 'place', 'latitude', 'longitude']
for line in sys.stdin:
    line = line.strip()
    results = geocoder.geocode(line, exactly_one=False) or [dummy]
    for j, res in enumerate(results):
        place = res[0]
        lat = str(res[1][0])
        lon = str(res[1][1])
        out = SEPARATOR.join([line, str(i), str(j), place, lat, lon])
        print (out)
    i += 1

Here is how you might use the script:

cat schools.csv | python 

Tip: you might want to

How AI, robotics and advanced manufacturing could impact everybody’s life on Earth

What if everybody could live a wealthy, healthy, job-less and creative life in a post-scarcity Universe? Are we currently on a trajectory to this new reality and what are the obstacles we may face on the way? What are the important game-changing technologies?

TODO: create and agenda (very tentative):

1) contrast current life circumstances with a potential future
2) identify the key problems that we could solve with technology
3) review the players in society that will take part in this change
3) contrast views on the opportunities and threats of these technologies
4) …

Our future life conditions here on Earth might soon be impacted by game-changing advancements in artifical intelligence, robotics, manufacturing and genetics; at least if you ask people like Elon Mush, Andrew Ng and Ray Kurzweil. What are the most important technologies and what is the impact they might have? What are the dangers? Opinions differ so the intention here is to review and contrast what leading fiction writers, scientists, visionaries and entrepreneurs think about the question: how will AI, robots, and advanced manufacturing impact everybody’s life circumstances here on Earth?

Fiction to be reviewed

– The Culture series

– Asimov

The Human-Computer Cortex:
– That Swedish guy who wrote sci-fi computer implants in the 70’s

Non-fiction to be reviewed

– Douglas Hofstadter: GEB

Videos to be reviewed


The Human-Computer Cortex:

News articles to be reviewed


3D printing:

Easy parallel HTTP requests with Python and asyncio

Python 3.x, and in particular Python 3.5, natively supports asynchronous programming. While asynchronous code can be harder to read than synchronous code, there are many use cases were the added complexity is worthwhile. One such examples is to execute a batch of HTTP requests in parallel, which I will explore in this post. Additionally, the async-await paradigm used by Python 3.5 makes the code almost as easy to understand as synchronous code.

Before we look at asynchronous requests, let us look at the sequential case. This will give us something to compare with later. The code listing below is an example of how to make twenty synchronous HTTP requests:

# Example 1: synchronous requests
import requests

num_requests = 20

responses = [
    for i in range(num_requests)

How does the total completion time develop as a function of num_requests? The chart below shows my measurements. The curve is unsurprisingly linear:

Using synchronous requests, I was able to execute just five requests per second in my experiment.

Next, let us see how we can make asynchronous HTTP requests with the help of asyncio. The code listing below is an example of how to make twenty asynchronous HTTP requests in Python 3.5 or later:

# Example 2: asynchronous requests
import asyncio
import requests

async def main():
    loop = asyncio.get_event_loop()
    futures = [
        for i in range(20)
    for response in await asyncio.gather(*futures):

loop = asyncio.get_event_loop()

The code is now more convoluted. But is it better? In the chart below, you will see the total completion time in seconds as a function of the number of asynchronous requests made.

The stepwise curve indicates that some requests are being executed in parallel. However, the curve is still asymptotically linear. Let’s find out why.

Increasing the number of threads

Notice that there is a step pattern in the chart? The completion time is 1x for 1-5 requests, 2x for 6-10 requests, 3x for 11-15 requests, and so on. The reason that we see this step pattern is that the default Executor has an internal pool of five threads that execute work. While five requests can be executed in parallel, any remaining requests will have to wait for a thread to become available.

So here is an idea. To minimize the total completion time, we could increase the size of the thread pool to match the number of requests we have to make. Luckily, this is easy to do as we will see next. The code listing below is an example of how to make twenty asynchronous HTTP requests with a thread pool of twenty worker threads:

# Example 3: asynchronous requests with larger thread pool
import asyncio
import concurrent.futures
import requests

async def main():

    with concurrent.futures.ThreadPoolExecutor(max_workers=20) as executor:

        loop = asyncio.get_event_loop()
        futures = [
            for i in range(20)
        for response in await asyncio.gather(*futures):

loop = asyncio.get_event_loop()

Let us rerun the experiment from before.

Bingo. If the executor has a pool of 20 threads then 20 requests take roughly the same amount of time as 1 request.

The next obvious question to ask is: what is the limit to this approach? How does the total completion time increase for a more substantial amount of asynchronous requests and equal amount of threads in the executor pool, say 100 or 1000? Let’s find out.

We are back to linear – the flat kind. The chart shows that we can easily make 500+ asynchronous HTTP requests in the same time it takes to make 10 synchronous requests – or 300 requests per second. Surely a vast improvement. You might have noticed that the chart only goes to 720 requests, not 1000. The reason is that “something bad” happened when I reached 720 requests/threads and my timing program became very sad and stopped working. On a different computer than mine, this limit will likely be found somewhere else.


First, you have seen that with asynchronous requests you can execute more requests per second than with synchronous requests: 300 requests per second compared to just 5 requests per second (in my experiments).

Second, you have seen how to achieve this well-known benefit of async I/O in Python 3.5 using just slightly convoluted code.

I would like to apologize to the people who host for hammering their website during my experiment. I like to think that it was for a good cause.


Symbiosen mellem mennesker og AI vil kunne transformere mennesket til en rationel organisme (jvf. Daniel Kahneman som har påvist at mennesket for sig selv ikke er en rationel organisme). Hvordan det? Vores minutiøse adfærd bliver i stigende grad sporet i alle livets væsentlige forhold. Kunstig intelligens bliver bedre og bedre til at skønne om vi er glade, sunde og rige udfra en analyse af alle de spor vi efterlader os overalt. Vi står nu i en situation hvor vi kan – eller snart kan – stille spørgsmål som: hvor glad, sund og rig var person X til tiden t? Hvilke handlinger h1, h2, h3, … havde person X udført (f.eks. på Spotify, rejser, jobskifte, lægebesøg) som ledte op til dette øjeblik? Hvor glad vil X være til tiden t+1, t+10, t+1000 hvis alting fortsætter som nu? Hvilke handlinger skal X udføre for at maksimere sin glæde til tiden t+1000?
Med andre ord, der er komplekse livsområder hvor kompleks AI har et potentiale for maksimere vore long-term utility (f.eks. vores “livsglæde” eller formue om 10 år). Forstil dig at en personlig AI kan
– Finde din næste bolig
– Finde en skole/fritidsaktivitet til dit barn
– Finde investeringsobjekter
– Finde kærlighed
– Finde venner
– Finde dit næste måltid
– o.s.v.

How to select top-k items from each group in SQL

Here is an analytical query that you (and I) will often need to do if you work in e-commerce, marketing or similar domain. It answers the question, within each group of items (e.g. partitioned by territory, age groups or something else) what are the top-k items for some utility function over the items (e.g. the number of units sold)?

Assume we have a table, aggregated_sales(id INT, item TEXT, country TEXT, total_sold INT). Here is a sample:

1   Basketball        DK       125
2   Basketball        US       432943
3   Leather football  FO       64773 
4   Leather football  DK       56230
5   Boxing gloves     CU       9812

In SQL, here is how to get the top-3 selling items by country:

SELECT id, item, country, total_sold
    row_number() OVER (PARTITION BY country ORDER BY total_sold DESC) as rn
  FROM aggregated_sales
) t
WHERE rn = 1;

Bonus info: the table sample was generated from a CSV file with the command column -s , -t filename.csv.

How to get structured Wikipedia data via DBPedia

Wikipedia contains a wealth of knowledge. While some of that knowledge consists of natural language descriptions, a rich share of information on Wikipedia is encoded in machine-readable format, such as “infoboxes” and other specially formatted parts. An infobox is rendered as a table that you typically see on the right-hand side of an article.


While you could download the page source for a wikipedia article and extract the information yourself, there is a project called DBPedia that has done the hard work for you. That right, you can conveniently retrieve machine-readable data that stems from Wikipedia via the DBPedia API.


Let us explore the DBPedia API by way of an example.

I like tennis data and most player pages on Wikipedia have an infobox that contains basic information about a player, such as age, hand, and current singles rank. Let’s try to retrieve information about the Italian tennis player, Matteo Donati, via the JSON resource exposed by DBPedia:

In this example, we will fetch and process the JSON data with a small Python script.

# Python
import requests
data = requests.get('').json()
matteo = data['']
# matteo is a dictionary with lots of keys
# that correspond to the player's properties.
# Each value is a list of dictionaries itself.
height = matteo[''][0]['value']
# 1.88  (float)
birth_year = matteo[''][0]['value']
# '1995'  (string)
hand = matteo[''][0]['value']
# 'Right-handed (two-handed backhand)'  (string)
singles_rank = matteo[''][0]['value']
# 'No. 171'  (string)

The simple convention for URLs on DBPedia is that spaces in names are replaced by underscores, exactly like on Wikipedia. For example, if we wanted to look up Roger Federer, we would make a request to the resource:

Please note, that at the time of writing, DBPedia does not support https.

Redundancy and inconsistency

The data on Matteo Donati and other entities on DBPedia is both redundant and somewhat inconsistent. This can be seen if we enumerate the keys on Matteo Donati:

for key in sorted(matteo): print(key)

You’ll notice that, e.g., the height of Matteo Donati is stored under two different keys:


Luckily, both keys list Donati’s height as 1.88 m, albeit as a string type and numeral type respectively. Other bits of information that is redundantly stored include his birth date, dominant hand (“plays”) and career prize money won so far.

With redundancy comes the possibility for inconsistency. In other words, there is no guarantee that redundant keys will keep identical values. For example, Matteo Donati is listed both as ‘Right-handed (two-handed backhand)’ and simply as ‘Right-handed’. While in this case the inconsistency is merely a matter of information detail, it can get a little confusing in general.


DBPedia is a great way to access structured data from Wikipedia articles. While the information is machine-readable in a popular format, you will have to guard against missing keys, redundant keys and inconsistent values. I hope you enjoyed this quick introduction to DBPedia and that you will find good use for the information.

How to compute the pagerank of almost anything

Whenever two things have a directional relationship to each other, then you can compute the pagerank of those things. For example, you can observe a directional relationships between web pages that link to each other, scientists that cite each other, and chess players that beat each other. The relationship is directional because it matters in what direction the relationship points, e.g. who lost to who in chess.

Intuitively, you may think of directional relationships as a transferal of some abstract value between two parties. For example, when one chess player loses to another in chess, then the value (i.e. relative skill level) of the winner will increase and the value of the loser decrease. Furthermore, the amount of value that is transfered depends on the starting value of each party. For example, if a master chess player loses to a novice chess player in chess, then the relative skill level of the novice will dramatically increase. Conversely, if a novice chess player loses to a master chess player in chess, then that is to be expected. In this situation, the relative skill level of each player should remain roughly the same as before – the status quo.

Below you’ll see an illustration of a small graph with seven nodes and seven edges. The pagerank of each node is illustrated by shading it, where a darker color denotes a higher rank.

If you study this figure, you should notice that:

  • Nodes 1 through 4 all have low rank, because no other nodes point to them
  • Node 5 has a medium rank, because a low-rank node points to it
  • Node 6 has high rank, because many low-rank nodes point to it
  • Node 7 has the highest rank, because a high-rank node points to it, while it points to nothing

Compute pagerank with Python

The pageranks of the nodes in the example graph (see figure above) was computed in Python with the help of the networkx library, which can be installed with pip: pip install networkx. The code that creates a graph and computes pagerank is listed below:

import networkx as nx
# Initialize directed graph
G = nx.DiGraph()
# Add edges (implicitely adds nodes)
# Compute pagerank (keys are node IDs, values are pageranks)
pr = nx.pagerank(G)
  1: 0.06242340798778012, 
  2: 0.06242340798778012, 
  3: 0.06242340798778012, 
  4: 0.06242340798778012, 
  5: 0.08895357136701444, 
  6: 0.32374552689540625, 
  7: 0.33760726978645894

Notice that each nodes is represented by an integer ID, with no specific semantics tied to the nodes nor the edges. In other words, the graph could equally well represent relationships between web pages, scientists and chess players (or something else entirely).

If your relationships can be assigned weights, e.g. the strength of a victory in chess or the prominence of a link on a web page, then you can add weights to the edges in the graph. Luckily, weighted edges can be easily added in networkx:

G.add_edge(1, 2, weight=0.5)

Dealing with time

You may ask yourself, should a chess game that took place last year impact a player’s rank as much as a game that was won or lost just last week? In many situations, the most meaningful answer would be no. A good way to represent the passing of time in a relationship graph is to use edge weights that decrease over time by some function. For example, an exponential decay function can be used, such that relationships that were formed a long time ago have exponentially lower weight than recently formed relationships. This can be achieved in Python with the ** operator with a negative exponent:

time_decayed_weight = max(.00001, time_passed) ** -1
G.add_edge(1, 2, weight=time_decayed_weight)

We use the trick max(.00001, time_passed) to ensure that we do not raise zero to the power of a negative number. The unit of time passed depends on the domain, and is not essential to the computation. For example, the unit could be milliseconds, years or millennia.

To be continued…

Quick introduction to RabbitMQ and Celery

I like to code in Python. I also like the concept of asynchronous workers to build loosely coupled applications. Luckily, RabbitMQ and Celery can help me do exactly that.

This post is based on a very nice YouTube video by Max Mautner (the one below).

For easy repeatability, I have transcribed the video in this article, with some minor changes for Python 3.

Install RabbitMQ and Celery

I will assume that you are on a Mac. If not, you will have to think a bit to make the instructions work.

First, let us install RabbitMQ if you don’t already have it.

brew install rabbitmq
# add the RabbitMQ tools to your path, however you want to do that.
# The tools are in /usr/local/sbin
echo '/usr/local/sbin' >> ~/.bash_profile

Next, I recommend that you create a virtual environment for this project and install Celery in there.

# create virtual environment
virtualenv -p python3 p3env
# activate environment
source p3env/bin/activate
# install celery in the environment
pip install celery
# Test that it worked
celery -h

Example application

This example is the simplest possible. It assumes the defaults for Celery, such as using a local RabbitMQ as the message broker. The application will distribute 10000 addition tasks to be executed by Celery.

We will need two files. One file that defines the Celery task ( and one for the driver program ( Of course, you can call these files anything you want. Also, the driver program is just a simple way to push tasks to RabbitMQ (the Celery default), which will later be dequeued by the Celery workers.

First, let’s create

from celery import Celery
app = Celery()
def add(x, y):
    return x + y

Next, let’s create

from tasks import add
for i in range(10000):
    # The delay function was added by Celery, when we decorated the add function
    add.delay(i, 1)

As you can see, the driver program consists of a loop that calls add.delay(i, 1) 10000 times. We did not explicitly define the delay function. It was added automatically when we decorated the add function with the annotation @app.task. This means that the function call will be pushed to the message broker and executed asynchronously on the workers.

Run example

To run the example, first start the local RabbitMQ server in a new Terminal window:

# Start message broker
# check that /usr/local/sbin is in path 
# if this does not work

In another Terminal window, start the Celery workers:

# activate the virtual env in the new window
source p3evn/bin/activate
# start the workers
celery worker -A tasks -l INFO

Finally, run the driver program to create 10000 tasks:

# activate the virtual env in the new window, if needed
source p3evn/bin/activate
# run the driver program

Now, in the Terminal window that is running the workers, you should see lines fly by as tasks are being executed asynchronously.

What to do next?

Celery fits a lot of use cases, from web scraping, API consumption, long-running web application tasks etc. The follow-up video by Max demonstrates a simplified web scraping use case. Like the first video, it is succinct and sufficient for a basic understanding.