Amazon ML Challenge 2024

A comprehensive pipeline for extracting product attributes from images using OCR and deep learning. This system combines PaddleOCR for text extraction, BERT for text encoding, and a fine-tuned BART model for attribute prediction.

Key Features

• Dual Conda environment setup for OCR and ML workflows
• Automated image preprocessing with OpenCV CLAHE enhancement
• Hybrid text processing with PaddleOCR and BERT embeddings
• BART-based sequence-to-sequence model for attribute prediction
• Fuzzy matching post-processing for unit standardization
• GPU-accelerated processing with checkpointing

Installation

Prerequisites

• NVIDIA GPU with CUDA 11.8+
• Conda package manager
• Kaggle API credentials

# PaddleOCR Environment
conda create -n paddle_ocr python=3.10
conda activate paddle_ocr
conda install -c "nvidia/label/cuda-11.8.0" cuda-toolkit
pip install paddlepaddle-gpu==2.5.1.post117 -f https://www.paddlepaddle.org.cn/whl/linux/mkl/avx/stable.html
pip install pandas tqdm requests opencv-python

# Main ML Environment
conda create -n amazon_ml python=3.10
conda activate amazon_ml
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install transformers pandas tqdm scikit-learn nltk fuzzywuzzy kagglehub
python -m nltk.downloader punkt

Usage

Workflow Pipeline

1. Data Preparation

   python Download_kagglehub.py
   python Linking.py
   

2. OCR Processing

   python Preprocessing.py
   

3. Text Encoding

   python Encoding.py
   python Cleaning.py
   

4. Model Training

   python Fine_Tuning.py
   

5. Prediction & Correction

   python Prediction.py
   python Unit_Correction.py
   

Code Structure

File	Purpose	Key Technologies
Preprocessing.py	Image enhancement & OCR	PaddleOCR, OpenCV
Encoding.py	Text embedding generation	BERT, PyTorch
Fine_Tuning.py	Model training	BART, HuggingFace
Unit_Correction.py	Output standardization	FuzzyWuzzy

Results

Evaluation Metrics (20 Epochs):

• Exact Match Accuracy: 78.42%
• BLEU-4 Score: 0.851

Example Prediction:

Input:  width | Product Width: 15.5inc
Output: 15.5 inch

Directory Structure

Amazon ML Challenge 2024/
├── archive/
│   ├── images/           # Raw product images
│   └── dataset/         # CSV metadata files
├── outputs/
│   ├── processed/       # Cleaned datasets
│   └── predictions/     # Model outputs
└── model/               # Saved BART models

License

MIT License