Trendyol Teknofest E-Commerce Hackathon Project - Product Name Generation Model, Text Improvement Model & Image Similarity Detection Algorithm

This repository is part of an e-commerce solution aimed at enhancing the digital presence of local producers by generating product names, improving product descriptions, and finding visually similar high-resolution images. The project includes three main components: product name generation, text improvement, and image similarity detection.

Project Structure

The repository is divided into four key modules:

1. Find Product Name Model
   Fine-tunes a pre-trained MT5 model to generate product names based on short product descriptions.

2. Text Improvement Model
   Fine-tunes another MT5 model to improve and enrich product descriptions in Turkish.

3. Image Similarity Detection
   Uses the VGG16 model to compare the similarity between a user-provided image and a collection of high-resolution images.

4. Inference Pipeline
   Integrates the above models into a Gradio interface to provide a complete solution for product name generation, description improvement, and image similarity comparison.

---

Project Workflow Diagram

---

Project Inference Samples

Installation

Dependencies

All the necessary libraries are listed in the requirements.txt files located in each folder. To install the required libraries, run the following command:

pip install -r requirements.txt

For CUDA-supported versions (for faster training and inference):

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124

---

Modules Overview

1. Find Product Name Model

This model fine-tunes the MT5 model to generate product names from product descriptions. It uses a custom dataset of input-output pairs where each input represents a description, and the output is a suitable product name.

Files and Their Functions:

File Name	Description
evaluation_find_product_name_model.ipynb	Jupyter notebook for evaluating the model using ROUGE metrics.
find_product_name_model_generator.py	Script for fine-tuning and saving the MT5 model. Includes a sample inference function.
findProdName-io.json	Custom dataset with input-output pairs for product name generation.
requirements.txt	Required dependencies for this module.

Usage:

• To train the model, run:

  python find_product_name_model_generator.py

• To evaluate the model using ROUGE metrics, open and run:

  jupyter notebook evaluation_find_product_name_model.ipynb

---

2. Text Improvement Model

This model improves product descriptions by enhancing natural language content using a fine-tuned MT5 model. It focuses on making product descriptions more appealing and informative.

Files and Their Functions:

File Name	Description
evaluation_text_improvement_model.ipynb	Jupyter notebook for evaluating the fine-tuned model using ROUGE metrics.
text_improvement_model_generator.py	Script for fine-tuning and saving the model.
textImprove-io.json	Dataset with input-output pairs used for training the model.
requirements.txt	Required dependencies for this module.

Usage:

• To train the text improvement model, run:

  python text_improvement_model_generator.py

• To evaluate the model, open and run:

  jupyter notebook evaluation_text_improvement_model.ipynb

---

3. Image Similarity Detection

This project compares a user-provided image to a set of high-resolution images and finds the most similar one using the VGG16 model and cosine similarity.

Files and Their Functions:

Function Name	Description
load_vgg_model()	Loads the pre-trained VGG16 model.
extract_features()	Extracts a 4096-dimensional feature vector from the image.
calculate_cosine_similarity()	Computes the cosine similarity between two feature vectors.
find_most_similar_image()	Finds and displays the most similar image based on the feature vectors.

Usage:

• Ensure you have installed the required libraries with:

  pip install numpy opencv-python tensorflow scikit-learn matplotlib

• Use the find_most_similar_image() function to compare images and find the closest match based on cosine similarity.

---

4. Inference Pipeline

This script integrates the product name generation model, text improvement model, and image similarity detection into a Gradio interface. The interface accepts product descriptions and images and returns a generated product name, improved description, and the most similar image.

Files and Their Functions:

File Name	Description
inference.py	Script to launch the Gradio interface, integrating all models and algorithms.
requirements.txt	Required dependencies for running the Gradio interface.

Usage:

• Install the required dependencies by running:

  pip install -r requirements.txt

• To launch the Gradio interface, run:

  python inference.py

• The interface will allow you to input text and images and generate a product name, improve the text, and find the most visually similar image.

---

Acknowledgements

• Hugging Face Transformers: For providing the MT5 model and training utilities.
• PyTorch: For the deep learning framework used in this project.
• TensorFlow/Keras: For the VGG16 model used in image similarity detection.