Lo-Fi Python

1. Export all Facebook post images from my page

Facebook made this easy. I exported all of my timeline photos by following these instructions.

2. Converting images to quote text with OCR

Optical character recognition or optical character reader (OCR) is the electronic or mechanical conversion of images of typed, handwritten or printed text into machine-encoded text, whether from a scanned document, a photo of a document, a scene-photo

Wikipedia - https://en.wikipedia.org/wiki/Optical_character_recognition

Once I had a folder of .jpg images, I used the Google Vision API's OCR to detect the text in the images. I also considered using the open source Calamari OCR library, but my research found that Google's Vision API may be more effective at detecting text.

Since I had only 771 images, I was able to extract text on all of them and stay within Google's free plan (1,000 requests / month). I followed these installation instructions on my Ubuntu Linux computer. It worked well on most of the images. Here's the code I used to detect text in all my images and save it in a .csv file:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49

import io
import os
from google.cloud import vision
import pandas as pd


"""Setup Instructions
1) Save this as detect_image_text.py
2) Create a folder named 'photos' and put your photos in them.
3) in your terminal, run: python detect_image_text.py
"""

def detect_text(path):
    """Detects text in the file."""

    client = vision.ImageAnnotatorClient()

    with io.open(path, "rb") as image_file:
        content = image_file.read()

    image = vision.Image(content=content)

    response = client.text_detection(image=image)
    texts = response.text_annotations
    print("Texts:")

    for text in texts:
        return text.description

    if response.error.message:
        raise Exception(
            "{}\nFor more info on error messages, check: "
            "https://cloud.google.com/apis/design/errors".format(
                response.error.message))

images = os.listdir("/images")
img_text = list()
for i, img in enumerate(images):
    try:
        img_path = f"photos/{img}"
        text = detect_text(img_path)
        img_text.append(text)
        print(f"{i}: {text}")
    except:
        print(f"Failed: {img}")

quotes_df = pd.DataFrame(img_text, columns=["Text"])
csv = "Extracted Image Text.csv"
quotes_df.to_csv(csv, index=False)

3. Data cleaning (via pandas and manual correction)

The data did not come back perfect, but I was pleased with the Google Vision API's results. It saved me a lot of time compared to manually transcribing the images! Next I used pandas to clean the data. You can see more in a Jupyter notebook with all of the code on github. Then I manually removed the author or source names, keeping only the quote text.

4. Train on past quotes and generate new quotes

GPT-3, The State of the Art Option

Initially, I considered machine learning options for generating new text. The GPT-3 library, released in early 2020 by Open AI, is the current state of the art model for text generation. However, its API is only accessible on an invite basis. If I get access, I think I'll try using it with the GPT-Sandbox python library. 🤞 You can join the GPT-3 waitlist here.

I searched around for other text generation python libraries on Github and found a promising one named GPT-2_simple, which utilizes GPT-3's predecessor. However, it requires using an old version of TensorFlow. I feel less inclined to learn older versions of machine learning libraries. Currently, I'm holding out for GPT-3 access. I may try the GPT-2 route if I don't get a chance at GPT-3.

A "Simple is Better Than Complex" Approach: Markov Chain

I wondered, are there any simpler options for text generation in python? Enter the Markov chain, which I stumbled across while Googling.

A Markov chain is a stochastic model describing a sequence of possible events in which the probability of each event depends only on the state attained in the previous event.

Wikipedia - https://en.wikipedia.org/wiki/Markov_chain

Using the markovify library

Google pointed me to this post from Analytics India Magazine showing the "Markovify" library. Markovify makes generating your own Markov chain very easy! Install with pip:

pip install markovify

Here's the code to create Markov chain on the quote text:

1
2
3
4
5
6
7

import markovify
# Build a markov chain model.
text_model = markovify.Text(text)

# Print five randomly-generated sentences
for i in range(5):
    print(text_model.make_sentence())

Markov chains are below the level of sophistication of machine learning technologies like GPT-3 or GPT-2. But Markov chains demonstrate how we can apply mathematics to mimic results or at least achieve an mvp with a simpler approach. Another intriguing tool worth mentioning is the nltk library, which offers natural language capabilities.

Eventually I will try the more sophisticated way using machine learning, but at least I am enjoying a quick taste of success with a Markov chain. Here's what some cherry-picked results look like! 😆 Ok, they're not great, but not too shabby either for my first time generating text from examples:

Maybe in the future I will use this for posts on my Facebook page, but it's not quite ready yet! I really enjoyed the process of researching this challenge and hope this post helps you evaluate your own text generation possibilities. This was fun to learn about. And best of all, I achieved satisfying, albeit primitive results within one weekend. Thanks for reading and stay positive.

Check out the Markov chain in the wild here.