Skip to content

hariprasath-v/Machinehack-Data-Centric-AI-Competition-2023

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

5 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Machinehack-Data-Centric-AI-Competition-2023

Public Leaderboard

  • Rank : 15
  • Score : 0.96096

Private Leaderboard

  • Rank : 16
  • Score : 0.96096

Competition hosted on Machinehack

Problem

Build machine learning model to predict the character of each image..

Evaluation

Evaluation metric for this competition is accuracy.

Dataset

You can download the dataset here

Solution

Exploratory Data Analysis

The basic exploratory data analysis of the data,

  • Basic image meta data analysis
  • Image similarity analysis

The above analysis had done by using,

  • cv2
  • Image
  • numpy
  • seaborn
  • matplotlib
  • pandas

Model

The train dataset contains miss labeled images, the incorrectly labeled images are identified by using ,

  • Simple pytorch convolutional neural network
  • Skorch - Scikit-Learn compatible neural network library
  • Cleanlab - No code library to fix the errors in dataset

The simple CNN model run 4 iterations(50 epochs/iteration) and cross-validated on 10-kfold split train dataset.In each iteration, the actual label and the predicted probability for the labels are compared by using the find_label_issue function from cleanlab tool.Then, based on the self-confidence level the miss labeled images are collected.

Incorrect Label Image sample

Numbers

Alt text

Letters

Alt text

After removing 2834 miss labeled images from the train dataset, the cleaned data was trained with KNN classifier model with 1 nearest neighbor.

KNN model gives good accuracy(0.8995) on validation data.

ROC-AUC Score

Alt text

True Positive Rate

Alt text

Test Prediction

Numbers

Alt text

Letters

Alt text

File Information

data-centric-ai-competition-2023-image-data-eda.ipynbOpen in Kaggle

data-centric-ai-competition-2023-image-data-model.ipynbOpen in Kaggle