At TaxSpanner as we were entering the "filing season" we wanted an continues integration system in place which could tests independent apps and stage them for further testing. We setup jenkins, created pipeline for different apps but as models grew and migrations were getting added, the whole pipeline started crawling.

There are different moving parts in whole application, say app1, app2, app3 and so on. While many of applications do interact with each other using API's but somehow application level testing using manage.py doesn't exactly limit things to just that particular app. What I mean is, say I am working on app2 running its test using

./apps/manage.py test app2

will still load all installed_apps and try to get your "test DB ready with all models, indexes and migrations". With the complicated models this preparing process itself was taking hours(2 hrs+). I am not sure about what would be standard approach for such a setup, should we "include" settings from each app individually? Anyway, so because of that behavior of django's test suite we struggled and even started avoiding writing/running tests.

Somedays back, as we started work on revamping an app around django-rest-framework and I was again stuck at this issue. While looking around I was just experimenting with settings and added this "hack" at the end of settings.py

import sys
if 'test' in sys.argv or 'test_coverage' in sys.argv:
    INSTALLED_APPS = ('app2', 'rest_framework', 'usermanagement',
		      'django.contrib.sites', 'registration',
		      'django.contrib.sitemaps',
		      'django.contrib.webdesign',)

and voila, my tests were back on track, under a minute(~10 seconds) with no unrelated apps. There were further suggestions to make things more faster by team :

DATABASES['default']['ENGINE'] = 'django.db.backends.sqlite3'

and further to avoid migrations,

class DisableMigrations(object):
    def __contains__(self, item):
	return True

    def __getitem__(self, item):
	return "notmigrations"
MIGRATION_MODULES = DisableMigrations()

I met one of my school friend earlier in the start of this month and he was mentioning about how he and his wife were trying to teach some kids from neighborhood but society people interfered saying rogue kids are entering society premises on excuse of class and causing trouble. They "suggested" the couple to rent a room somewhere outside and take classes there. Of course, this is outrageous to expect someone to take such an extra measure to actually do something good and in the end, they just left.

It was really sad to hear their part of story but I am really fortunate to have totally different set of people around me. I recently started working with TaxSpanner Team. Before joining them in Delhi, I was concerned about how I would be able to continue with work done till now with Pankhudi/पँखुडी. I was wondering if I could come down to Pune and continue working from here once I am all set with the team and working methodology. Ankur, Gora, Vivek and others readily agreed and I spent last two weeks of January here.

In library there was issue with earthing and computers had been down from quite some time. Thanks to Srikant, Padma and Pratiksha's(maybe I missed others name) effort we had proper working switch boards in library which were needed BIG time. Computer, Raspberry Pi(we have offline Wikipedia, Khan-academy videos, scratch and other resources on it, though use case is still very limited), Laptop, all were working just fine. Also, people "hosted" me so readily, provided all resources(office to work from, vehicles to commute, even pick and drop facility, lunches/dinners/snacks), that one feels right at home and carry on with the work/anything without any adaptation, Sooper it is(घन्यवाद बोला तो लोग मारेंगे). Hope to see you all again soon :)

Here are logs of activity going on in library while I was there

tl;dr here is the code

After my last post I was following the TODO block to see if I can get a simple multilingual document clustering (MDC) working. First step was to see if we simply translate Hindi documents into English, word by word and then run Clustering algorithm over it, do I get same results as that of clustering documents in their original language.

I used translate python package to convert all documents to English using following code

from translate import Translator
from os import path
import pickle
import string

# adding Hindi punctuation to regular English punctuation
unicode_punctuation = string.punctuation+u"\u0964\u0965"
translate_table = dict((ord(char), u" ") for char in unicode_punctuation)
stop_words = [word.decode("utf-8") for word in stopwords.words("hindi")]
translator = Translator(to_lang="en", from_lang="hi")
if path.isdir("hi_dict"):
    hi_dict = pickle.load(open("hi_dict"))
else:
    hi_dict = {}
for link in links:
    content = link["content"].translate(translate_table)
    content = " ".join(content.split())
    tr_words = []
    for word in content.split():
	if word in stop_words:
	    continue
	if word in hi_dict:
	    tr_words.append(hi_dict[word])
	elif word.isdigit() or not word.isalpha():
	    tr_words.append(word)
	elif word.istitle() and word.lower() not in hi_dict:
	    print word
	    tr_words.append(word)
	else:
	    eng_word = translator.translate(word.encode("utf-8"))
	    hi_dict[word] = eng_word
	    print word, eng_word
	    tr_words.append(eng_word)
    link["content"] = " ".join(tr_words)
pickle.dump(hi_dict, open("hi_dict", "w"))

Here links is simple data structure which stores the urls and text content of those web pages. At this stage, I perform clustering over this translated text. I got were in sync with my previous .

With this hypothesis working I mixed up data(in English) which I used in my first post and this translated data. After shuffling data and playing around with parameters I used following code to run clustering algo.

import numpy as np
from sklearn.feature_extraction.text import TfidfVectorizer
from random import shuffle
from sklearn.metrics.pairwise import linear_kernel

shuffle(links)
blobs = [link["content"] for link in links]

vectorizer = TfidfVectorizer(stop_words = "english", ngram_range = (2, 3), max_features = 100000)
corpus_mat = vectorizer.fit_transform(blobs)

cosine_similarities = np.zeros((len(blobs), len(blobs)))
for i in range(len(blobs)):
    cosine_similarities[i] = linear_kernel(corpus_mat[i:i+1], corpus_mat).flatten()
    cosine_similarities[i, i] = 0

# Majorclust algorithm
t = False
indices = np.arange(len(links))
while not t:
    t = True
    for index in np.arange(len(links)):
	# aggregating edge weights 
	new_index = np.argmax(np.bincount(indices, 
					  weights=cosine_similarities[index]))
	if indices[new_index] != indices[index]:
	    indices[index] = indices[new_index]
	    t = False

# Organizing the end results
clusters = {}
for item, index in enumerate(indices):
    if "title" in links[item]:
	title = links[item]["title"].encode("utf-8")
	clusters.setdefault(index, []).append(title)
    else:
	clusters.setdefault(index, []).append(links[item]["url"].encode("utf-8"))
op_file = open("op-mixed", "wb")
for item in clusters:
    op_file.write("\n"+80*"="+"\n")
    op_file.write("\n".join(clusters[item]))
op_file.close()

For this code links contains all translated articles and articles written in english, and this is the complete .

                       
                     ,                     
http://kafila.org/2013/12/08/aap-halts-bjp-advance-in-delhi/
                       
                     
                                !
http://kafila.org/2013/11/15/first-terrorist-of-independent-india/
                
                   

http://arstechnica.com/tech-policy/2013/12/nsa-collects-nearly-5-billion-cellphone-location-records-per-day/
         ,                                
      !!                                      !!
http://arstechnica.com/tech-policy/2013/12/microsoft-google-apple-call-for-end-to-nsas-bulk-data-collection/
Google Play:                                
http://techcrunch.com/2013/12/11/google-android-device-manager-play-store/?utm_campaign=fb&%3Fncid=fb

As it turns out, since most of clustering algorithms don't depend on grammar or language specific constructs so if we remove stopwords and do a proper stemming, we should be able to use same Major Clust algorithm for hindi blogs too. I looked around for some active hindi blogs and via Indian Bloggers got the link of कस्‍बा/qasba blog maintained by रवीश कुमार जी(Ravish Kumar). But topics covered in that particular blog were mostly related to politics, so I needed some more diverse dataset to test effectiveness of clustering algorithm. I came across feed of Jagran Junction and was able to get feeds related to sports and technology in hindi.

After some cleanups(removing duplicate articles, punctuation, minimum length of article) and using almost same code as we used for clustering English articles I got decent/acceptable for hindi documents too.

One thing which I had to take care while tokenizing hindi text using default TfidfVectorizer was, default re module for python 2.7 was failing in my case for unicode characters(1, 2) and seems it is fixed in regex. So I used this custom code to get it working.

def regex_tokenizer(doc):
    """Return a function that split a string in sequence of tokens"""
    import regex
    token_pattern=r"(?u)\b\w\w+\b"
    token_pattern = regex.compile(token_pattern)
    return token_pattern.findall(doc)

stop_words = [word.decode("utf-8") for word in stopwords.words("hindi")]
vectorizer = TfidfVectorizer(stop_words = stop_words,
			     tokenizer = regex_tokenizer)
vectorizer.fit_transform(blobs)
features = vectorizer.get_feature_names()

Recently for work I had to deal with doing OCR over inverted pages, something like this

I came across some threads(1, 2, 3)on tesseract google group and StackOverflow which were close to problem I was dealing with. I tried example mentioned on tesseract API page, looked at OSDetect code(4) and though it has already couple of implementations like update_{bestorientation} somehow it was not working for me. Maybe I was doing something wrong but eventually I ended up with following code which works for me, it rotates the page despite the top line.

#include <tesseract/baseapi.h>
#include <tesseract/osdetect.h>
#include <leptonica/allheaders.h>

int main(int argc, char **argv)
{  
    OSResults os_results;
    tesseract::TessBaseAPI *api = new tesseract::TessBaseAPI();
    if (api->Init(NULL, "eng")) {
	fprintf(stderr, "Could not initialize tesseract.\n");
	exit(1);
    }
    Pix *image = pixRead(filename);
    api->SetImage(image);
    // To detect correct OS and flip images
    api->DetectOS(&os_results);
    OrientationDetector os_detector = OrientationDetector(&os_results);
    int correct_orientation = os_detector.get_orientation();
    // Had to add this condition because get_orientation result and
    // pixRotateOrth were not in sync.
    if (correct_orientation == 1) {
	image = pixRotate90(image, -1);
    }
    else if (correct_orientation == 3) {
	image = pixRotate90(image, 1);
    }
    else if (correct_orientation == 2) {
	pixRotate180(image, image);
    }
    api->SetImage(image);
    char* ocrResult = api->GetUTF8Text();
    fprintf(stdout, "Recognized Text: %s\n", ocrResult);
    api->End();    
    pixDestroy(&image);
    delete [] ocrResult;
    return 0;
}

TL;DR here is public gist

n = 0, t = false
 v   V do n = n + 1, c(v) = n end
while t = false do
  t = true  
   v   V do
    c  = i if  (u, v) is max.
    if c(v) != c  then c(v) = c  , t = false
  end
end

texts = [
  "foo blub baz",
  "foo bar baz",
  "asdf bsdf csdf",
  "foo bab blub",
  "csdf hddf kjtz",
  "123 456 890",
  "321 890 456 foo",
  "123 890 uiop",
]

# corpus is list of all the documents
vectorizer = TfidfVectorizer(stop_words = "english", 
			     ngram_range = (2, 3),
			     max_features = 100000)
corpus_mat = vectorizer.fit_transform(corpus)
num_of_samples, num_of_features = corpus_mat.shape

magnitudes = np.zeros((num_of_samples))
# this loop can be removed?
for doc_id in range(num_of_samples):
  magnitudes[doc_id] = np.sqrt(corpus_mat[doc_id].dot(corpus_mat[doc_id].T).sum())

cosine_distances = np.zeros((num_of_samples, num_of_samples))
# this loop can be improved
for doc_id, other_id in combinations(range(num_of_samples), 2):
  distance = (corpus_mat[doc_id].dot(corpus_mat[other_id].T).sum())/(magnitudes[doc_id]*magnitudes[other_id])
  cosine_distances[doc_id, other_id] = cosine_distances[other_id, doc_id] = distance

t = False
indices = np.arange(num_of_samples)
while not t:
  t = True
  for index in np.arange(num_of_samples):
    # aggregating edge weights 
    new_index = np.argmax(np.bincount(indices, 
			  weights=cosine_distances[index]))
    if indices[new_index] != indices[index]:
      indices[index] = indices[new_index]
      t = False

clusters = {}
for item, index in enumerate(indices):
  clusters.setdefault(index, []).append(links[-50+item]["url"])

We had four days during diwaali break and some of other people from Pankhudi group were also mulling over introducing programming to kids. I was working with UC for this possibility of doing home automation so we thought it would be nice to start with Arduino, Pi etc to give kids a real feel about how things can be programmed.

We had one arduino, some LED's and one relay switch. I was thinking of showing how easily we can control manual switches using programming. We passed around arduino board, pointed out some pins etc for a small intro and then wired up things. I was using nanpy and ipython console to pass commands to Arduino. One line of code and we had leds, bulbs blinking.

Arduino.digitalWrite(13, Arduino.HIGH)
Arduino.digitalWrite(13, Arduino.LOW)

This initial group was bigger one, so I was not able to cover details etc, but mostly ended up "showing off" for random requests of what else can be done. But once kids got settled that there were not too many things, some stayed back while others drifted. Then we had a sort of extensive conversation about use cases, possibilities, why, how etc.

As other volunteers joined, there were parallel classes going on, and I was just covering same demo with some more details to different set of kids. Then I had couple of girls who started off with, यह क्या है भैया? After the answer their next question was, आप क्या काम करते हो? And here I slipped(mistake), I had a program open in editor, so I showed them those 5-10 lines and said, this is what I do, they were like ehh? and I replied, yeah we write programs like this and get paid. Now those 5-10 lines were not that complicated, but it certainly gave a wrong impression, and they were totally convinced to learn this skill.

Now with a limited focused audience, we got more time and over next few classes we covered some details of programming constructs. Good thing of using IPython is, it gives live interaction with system. First thing we covered was flipping the state of LED and hence introducing "if-else" construct or we called it something in lines of "या तो ये, नहीं तो ये"

if Arduino.digitalRead(13):
    Arduino.digitalWrite(13,Arduino.HIGH)
else:
    Arduino.digitalWrite(13,Arduino.LOW)

We covered ideas like what is this if-else condition, indentation part here and in later parts. Next thing we moved onto was, making LED blink.

from nanpy import Arduino

Arduino.digitalWrite(13,Arduino.HIGH)
Arduino.digitalWrite(13,Arduino.LOW)

Now for this, we had to write a small python script and run it, and also concept of how computer runs this script line by line(instructions). But with this program, LED blinked very quickly so we introduced sleep function and how to call it with different parameters. Also a brief mention of import business we are doing.

from nanpy import Arduino
from time import sleep

Arduino.digitalWrite(13,Arduino.HIGH)
sleep(2)
Arduino.digitalWrite(13,Arduino.LOW)

From here we covered blinking a LED for a specific number of times(say 5 or 4). Firstly we landed up with idea that we can repeat lines 5(nth) times and we will get the task done. They wrote the program and tested if it was working. Now I introduced concept of looping(repeat something जब-तक), variable, counters.

from nanpy import Arduino
from time import sleep

counter = 0
while counter < 5:
    Arduino.digitalWrite(13,Arduino.HIGH)
    sleep(2)
    Arduino.digitalWrite(13,Arduino.LOW)
    sleep(2)
    counter = counter + 1

Here it took some time to make them understand what all is going on, why are we doing what we are doing. Once comfortable(they took this more like a bitter pill) they experimented with variation of counters, increment, while loop conditions and also infinite loop conditions. We had different color LED's so we tried combination of switching them on, syncing them and they very promptly related it to Traffic lightning and Diwaali light decorations.

I had this small program/function which gives some sense of noise level of surroundings(using mircrophone of system), so we thought of using this and writing a program which changes color of LED as the noise level changes. We started off by doing some trial runs to get idea of what are the values we are getting for different noise levels, they came up with the logic of roughly at what values the color should switch, we did some trials and edits and eventually we had something like this:

from nanpy import Arduino
from time import sleep
from noise import getRMS

counter=1
while counter < 5:
    noise = getRMS()
    if noise == -1:
       continue
    print noise
    if noise < 2000:
       Arduino.digitalWrite(10,Arduino.HIGH)
       Arduino.digitalWrite(8,Arduino.LOW)
       Arduino.digitalWrite(12,Arduino.LOW)
    elif noise < 4000:
       Arduino.digitalWrite(12,Arduino.HIGH)
       Arduino.digitalWrite(8,Arduino.LOW)
       Arduino.digitalWrite(10,Arduino.LOW)
    else : 
       Arduino.digitalWrite(8,Arduino.HIGH)
       Arduino.digitalWrite(10,Arduino.LOW)
       Arduino.digitalWrite(12,Arduino.LOW)

Pin 8 was green, 10 was blue and 12 was red(most noisy) and again they were quick to conclude that such LEDs can be installed in principal's room to give him idea of which class is going berserk. Next I was thinking of writing a small function which can return major component of color being shown to webcam and switching LED accordingly but didn't get that finished.

Now although this all looks good/decent approach, I was getting uncomfortable about missing something. Some kids are from elementary, while others are doing under-graduation but most of them are not from STEM background. So I am not sure on how to go about introducing ideas like lists, numbers, floating point numbers, complicated conditions. And also how to link all these with their needs/curriculum, I think I will have to go through their books to get better idea of what might appeal to them and make them do something which they themselves can use. I was not able to connect all these exercise with the starting idea of getting paid for doing these things. In last class while thinking these things, students were trying out scratch and they readily understood idea of "events" and control loops etc, but as the class ended without much of learning they said to me: "भैया अगली बार से English पढते हैं|"

Hello,