NeuralNetflix - Multi-Modal Movie Genre Classification System

A Flask web app for multi-label movie genre classification using both textual (plot) and visual (poster image) inputs. Combines LSTM (text) and ResNet-18 (image) models trained on a balanced TMDB dataset across the top 10 genres.

---

Live demo

• Demo

---

Features

• Text-based genre classification (LSTM + GloVe)
• Image-based genre classification (ResNet-18)
• Multi-label outputs (top genres per modality from the API)
• Web UI: enter a plot and/or upload a poster

---

Project structure

.
├── app.py                 # Flask app entry point
├── models/                # Model weights, tokenizer pickle, embedding matrix
├── data/                  # Datasets for training (gitignored — add locally)
├── templates/             # HTML (Jinja2)
├── static/                # Assets (images, optional sample data)
├── scripts/               # Training & data prep (see TRAINING_README.md)
├── notebook/              # Jupyter notebooks
├── plots/                 # Training / evaluation plots
├── webapp/                # Extra web assets (e.g. JS)
├── TRAINING_README.md     # Full training guide
└── requirements.txt

---

Supported genres

Drama, Comedy, Romance, Thriller, Action, Horror, Documentary, Animation, Music, Crime

---

Getting started

Prerequisites

• Python 3.8+
• pip, Git
• PyTorch / torchvision (via requirements.txt)

Installation

git clone https://github.com/ujwalwag/Movie-Genre-Classification-Sys.git
cd Movie-Genre-Classification-Sys
python -m venv venv
source venv/bin/activate   # Windows: venv\Scripts\activate
pip install -r requirements.txt

Model files

Place trained assets under models/ (paths are fixed in app.py):

File	Role
genre_classifier.pth	Text model state dict
tokenizer.pickle	Vocabulary / tokenizer (pickle)
embedding_matrix.npy	Frozen GloVe embedding table
poster_genre_classifier.pth	Image (ResNet-18) state dict

If these are missing, train or copy them in (see TRAINING_README.md).

Run locally

python app.py

Default port is 10000 unless you set the PORT environment variable (e.g. 5000 on Windows: $env:PORT=5000; python app.py).

Open http://127.0.0.1:10000 (or your chosen PORT).

Training

End-to-end training (see TRAINING_README.md for data paths and options):

python scripts/train_all_models.py

---

Model details

Text (LSTM)

• GloVe 100d embeddings (frozen in embedding_matrix.npy)
• Tokenizer: models/tokenizer.pickle (not JSON)
• BiLSTM, mean pooling, dense head → genre_classifier.pth

Image (ResNet-18)

• torchvision ResNet-18 (ImageNet weights only used during training; inference loads saved weights)
• Final layer: 10 outputs with sigmoid for multi-label poster classification
• Input: posters resized/cropped to 224×224 and ImageNet-normalized

---

Acknowledgements

• TMDB dataset
• GloVe embeddings (Stanford NLP)
• PyTorch, Flask, torchvision