Hand-Drawn Sketch Classification

Overview

This project tackles the challenge of recognizing hand-drawn sketches. Various models were explored, including ResNet18, ResNet50, and Inception_v3, with Inception_v3 proving to be the most effective, achieving a validation accuracy of 57%.

Pipeline

The pipeline is constructed using PyTorch and encompasses:

• Dataset Handling: A custom dataset class for loading images and labels from a CSV file.
• Model Training: Fine-tuning of Inception_v3 with batch processing.
• Evaluation: A script to assess model performance and generate predictions.

File Descriptions

• dataset.py: Manages dataset loading and iterator creation.
• model.py: Loads and modifies Inception_v3, including parameter counting methods.
• train.py: Trains the model for 30 epochs, logs progress, and saves checkpoints.
• evaluate.py: Evaluates the dataset using the best model and outputs predictions in CSV format.

Usage

Training

To train the model, ensure your dataset is organized as follows:

BaseDir
├── images
│   ├── 1.png
│   ├── 2.png
└── labels.csv

Run the training script with the command:

python3 train.py

This script trains the model for 30 epochs, logging progress and saving checkpoints along the way.

Evaluation

To evaluate the model:

1. Download and extract the project.
2. Organize the dataset as described above.
3. Run the evaluation script with the command:

python3 evaluate.py pathToDataset

This generates an evaluation.csv file with predictions for each image in the dataset.

Contact

For any questions, please reach out to:

• Muhammad Shavaiz Butt - shavaizsohail@gmail.com
• Hasib Aslam - haslam.bscs21seecs@seecs.edu.pk