Visual Taxonomy Classification and Submission Pipeline

This repository contains a pipeline for solving the Visual Taxonomy Classification problem, involving attribute prediction for images based on their category. The project leverages MobileNetV2 for feature extraction and a Random Forest Classifier for attribute prediction.

Table of Contents

1. Features
2. Requirements
3. Setup
4. Dataset
5. Usage
6. Model Training and Evaluation
7. Inference
8. Submission
9. Acknowledgments

---

Features

• Custom Data Generator: Efficiently handles large-scale image datasets for training.
• Feature Extraction: Uses pre-trained MobileNetV2 for extracting image features.
• Multi-Label Classification: Predicts multiple attributes for each image (multi-output classification).
• Random Forest Classifier: Final classifier for attribute prediction based on extracted features.
• Kaggle Submission Formatting: Generates a valid submission file formatted for Kaggle competitions.

---

Requirements

Make sure you have the following installed:

• Python 3.8 or later

Libraries:

• TensorFlow
• scikit-learn
• pandas
• numpy
• matplotlib
• Pillow

To install all dependencies, run:

pip install -r requirements.txt

Setup

1. Clone the repository:

   git clone https://github.com/Varuni13/Meesho_Attributes_Predictions.git

   cd Meesho_Attributes_Predictions

2. Download the required datasets:

• Visual Taxonomy Dataset: Available on Kaggle.

• MobileNetV2 Weights: Download from google or from my repository.

4. Place the datasets and weights in the following structure:

   Meesho_Attributes_Predictions/

   ├── data/

   │ ├── train_images/

   │ ├── test_images/

   │ ├── train.csv

   │ ├── test.csv

   │ ├── category_attributes.parquet

   │ └── mobilenet_v2_weights_tf_dim_ordering_tf_kernels_1.0_128_no_top.h5

   └── visual_taxonomy_pipeline.py

Dataset

This project uses the Visual Taxonomy Dataset, which includes:

• Train and test images

• Attribute information for categories in train.csv and category_attributes.parquet.

Usage

1. Run the Pipeline:

   Execute the Python script to preprocess data, extract features, train the model, and generate predictions:

   python visual_taxonomy_pipeline.py
   
   

2. Verify Output:

   • Feature extraction and model training logs will appear in the console.

   • Submission file is generated at: submission.csv.

3. Notes on Model Weights:

   The pre-trained MobileNetV2 weights (mobilenet_v2_weights_tf_dim_ordering_tf_kernels_1.0_128_no_top.h5) must be downloaded separately and placed in the appropriate directory (/data/).

Model Training and Evaluation

1. Feature Extraction : We use a pre-trained MobileNetV2 model (with the top layers removed) to extract feature vectors from the images. These features are then used for classification.

2. Classification: A MultiOutputClassifier with a Random Forest base model is trained on the extracted features.This allows the model to predict multiple attributes for each image.

3. Evaluation Metrics :Accuracy and F1-Score (both Micro and Macro) are computed on the validation data to evaluate model performance.

Inference

1. Once the model is trained, the following steps are used to generate predictions for the test dataset:

2. Feature Extraction: The test images are processed using the same feature extraction pipeline.

3. Prediction: The trained Random Forest classifier predicts the attributes for each product.

4. Submission Formatting: The predictions are formatted into a submission.csv file, as required by the Kaggle competition.

To run the inference script, use:

      Python inference.py

This will:

• Load the pre-trained model and the test data.

• Process the test images.

• Predict the attributes for each product.

• Format the predictions into a Kaggle-compatible CSV file (submission.csv).

Submission

The generated submission file is formatted according to the competition requirements:

• Each product in the test set has the correct number of attributes predicted.

• Missing attributes are filled with the default value (dummy_value).

The final submission file (submission.csv) will be saved in the current directory.

Acknowledgments

1. Kaggle for hosting the competition and providing the datasets.

2. TensorFlow for enabling deep learning workflows.

3. MobileNetV2 for providing pre-trained weights used in feature extraction.

4. scikit-learn for supporting multi-output classification.

Feel free to reach out via GitHub Issues for any questions or feedback.