Toxicity Detection in Memes | EVA Application

Run on Google Colab

View source on GitHub

Download notebook

Install Application Dependecies

!wget -nc "https://raw.githubusercontent.com/georgia-tech-db/toxicity-classification/main/requirements.txt"
!pip -q --no-color install -r requirements.txt

File ‘requirements.txt’ already there; not retrieving.


�[1m[�[0mnotice�[1m]�[0m A new release of pip available: 22.2.2 -> 22.3.1
�[1m[�[0mnotice�[1m]�[0m To update, run: pip install --upgrade pip

Start EVA server

We are reusing the start server notebook for launching the EVA server.

!wget -nc "https://raw.githubusercontent.com/georgia-tech-db/eva/master/tutorials/00-start-eva-server.ipynb"
%run 00-start-eva-server.ipynb
cursor = connect_to_server()

File ‘00-start-eva-server.ipynb’ already there; not retrieving.

nohup eva_server > eva.log 2>&1 &

�[1m[�[0m�[34;49mnotice�[0m�[1;39;49m]�[0m�[39;49m A new release of pip available: �[0m�[31;49m22.2.2�[0m�[39;49m -> �[0m�[32;49m22.3.1�[0m
�[1m[�[0m�[34;49mnotice�[0m�[1;39;49m]�[0m�[39;49m To update, run: �[0m�[32;49mpython -m pip install --upgrade pip�[0m
Note: you may need to restart the kernel to use updated packages.

Load the Memes for analysis

!wget -nc "https://raw.githubusercontent.com/georgia-tech-db/toxicity-classification/main/meme1.jpg"
!wget -nc "https://raw.githubusercontent.com/georgia-tech-db/toxicity-classification/main/meme2.jpg"
cursor.execute('DROP TABLE IF EXISTS MemeImages;')
response = cursor.fetch_all()
print(response)
cursor.execute('LOAD IMAGE "meme1.jpg" INTO MemeImages;')
response = cursor.fetch_all()
print(response)
cursor.execute('LOAD IMAGE "meme2.jpg" INTO MemeImages;')
response = cursor.fetch_all()
print(response)

File ‘meme1.jpg’ already there; not retrieving.

File ‘meme2.jpg’ already there; not retrieving.

@status: ResponseStatus.SUCCESS
@batch: 
                                         0
0  Table Successfully dropped: MemeImages
@query_time: 0.03890102496370673
@status: ResponseStatus.SUCCESS
@batch: 
                            0
0  Number of loaded IMAGE: 1
@query_time: 0.050992883974686265
@status: ResponseStatus.SUCCESS
@batch: 
                            0
0  Number of loaded IMAGE: 1
@query_time: 0.016112261917442083

Create OCR Extractor UDF

!wget -nc "https://raw.githubusercontent.com/georgia-tech-db/toxicity-classification/main/ocr_extractor.py"
cursor.execute("DROP UDF OCRExtractor;")
response = cursor.fetch_all()
print(response)
cursor.execute("""CREATE UDF IF NOT EXISTS OCRExtractor
      INPUT  (frame NDARRAY UINT8(3, ANYDIM, ANYDIM))
      OUTPUT (labels NDARRAY STR(ANYDIM), bboxes NDARRAY FLOAT32(ANYDIM, 4),
              scores NDARRAY FLOAT32(ANYDIM))
      TYPE  Classification
      IMPL  'ocr_extractor.py';
      """)
response = cursor.fetch_all()
print(response)

File ‘ocr_extractor.py’ already there; not retrieving.

@status: ResponseStatus.SUCCESS
@batch: 
                                        0
0  UDF OCRExtractor successfully dropped
@query_time: 0.01684344606474042
@status: ResponseStatus.SUCCESS
@batch: 
                                                       0
0  UDF OCRExtractor successfully added to the database.
@query_time: 6.026845669141039

Create Custom UDF for Toxicity Classification

!wget -nc "https://raw.githubusercontent.com/georgia-tech-db/toxicity-classification/main/toxicity_classifier.py"
cursor.execute("DROP UDF IF EXISTS ToxicityClassifier;")
response = cursor.fetch_all()
print(response)
cursor.execute("""CREATE UDF IF NOT EXISTS ToxicityClassifier
                  INPUT  (text NDARRAY STR(100))
                  OUTPUT (labels NDARRAY STR(10))
                  TYPE  Classification
                  IMPL  'toxicity_classifier.py';
      """) 
response = cursor.fetch_all()
print(response)

File ‘toxicity_classifier.py’ already there; not retrieving.

@status: ResponseStatus.SUCCESS
@batch: 
                                              0
0  UDF ToxicityClassifier successfully dropped
@query_time: 0.01804027589969337
@status: ResponseStatus.SUCCESS
@batch: 
                                                             0
0  UDF ToxicityClassifier successfully added to the database.
@query_time: 1.5074586649425328

Run Toxicity Classifier on OCR Extracted from Images

cursor.execute("""SELECT OCRExtractor(data).labels,
                  ToxicityClassifier(OCRExtractor(data).labels)
                  FROM MemeImages;""")
response = cursor.fetch_all()
print(response)

@status: ResponseStatus.SUCCESS
@batch: 
                              ocrextractor.labels toxicityclassifier.labels
0                  [CANT FuCK WITh, MEIN SWAG E]                     toxic
1  [YOU CANT SPELL, CLINTON WITHOUT CNN, igfip:]                 not toxic
@query_time: 11.451739435084164

Visualize Model Output on Images

from pprint import pprint
from matplotlib import pyplot as plt
import cv2
import numpy as np

def annotate_image(detections, input_image_path, image_id):

    print(detections)

    color1=(207, 248, 64)
    color2=(255, 49, 49)
    thickness=4

    df = detections
    df = df[df.index == image_id]

    image = cv2.imread(input_image_path)

    if df.size:
        df_value_list = df.values
        ocr = ' '.join(df_value_list[0][0])
        label = df_value_list[0][1]
        print(label)

        plt.imshow(image)
        plt.show()

        cv2.putText(image, label, (25, 200), cv2.FONT_HERSHEY_SIMPLEX, 3, color2, thickness, cv2.LINE_AA) 

        cv2.putText(image, ocr, (25, 250), cv2.FONT_HERSHEY_SIMPLEX, 0.8, color1, thickness, cv2.LINE_AA) 

        plt.imshow(image)
        plt.show()        

from ipywidgets import Image
dataframe = response.batch.frames
annotate_image(dataframe, 'meme1.jpg', image_id=0)
annotate_image(dataframe, 'meme2.jpg', image_id=1)

                             ocrextractor.labels toxicityclassifier.labels
0                  [CANT FuCK WITh, MEIN SWAG E]                     toxic
1  [YOU CANT SPELL, CLINTON WITHOUT CNN, igfip:]                 not toxic
toxic

                             ocrextractor.labels toxicityclassifier.labels
0                  [CANT FuCK WITh, MEIN SWAG E]                     toxic
1  [YOU CANT SPELL, CLINTON WITHOUT CNN, igfip:]                 not toxic
not toxic