Human Sentinment Classification (NLP)

A PyTorch implementation of sentiment classification using Recurrent Neural Networks (GRU) and Bidirectional GRU (BiGRU) models. This project trains on the SST-5 movie reviews dataset to classify text as "very negative", negative, neutral, positive, and "very positive".

Features

• Two Model Architectures: Simple GRU and Bidirectional GRU for sentiment analysis
• Embedding Strategies: Random embeddings, frozen GloVe embeddings, and fine-tuned GloVe embeddings
• Automatic Vocabulary Building: Creates vocabulary from training data with configurable max size
• Text Preprocessing: Tokenization, and punctuation filtering
• Training and Evaluation: Complete training pipeline with validation
• Prediction Mode: Classify sentiment of custom text inputs
• Checkpoint Saving: Automatically saves best performing models
• Ablation Studies: Experiments over embedding dimensions (50, 100, 200) and OOV handling strategies
• MLflow Integration: Experiment tracking and logging
• GPU Support: Utilizes CUDA if available

Installation

1. Clone or download this repository

   git clone https://github.com/bhatishan2003/Sentiment-Classification-with-GRU-in-Pytorch.git
   cd Sentiment-Classification-with-GRU-in-Pytorch

2. Install dependencies:

   pip install -r requirements.txt

Usage

Training a Model

Train an GRU model for custom hyperparameters

python sentiment_classifier.py --model gru --epochs 15  --hidden_dim 256 --dropout 0.25 --lr 0.001 --batch_size 64 --patience 5 --embed_dim 200

Train a BiGRU model with custom hyperparameters:

python sentiment_classifier.py --model bigru --epochs 20 --batch_size 64 --lr 0.001 --embed_dim 200 --hidden_dim 256

Making Predictions

You can predict sentiment on custom text using a trained model. Specify the embedding strategy to load the appropriate checkpoint.

Using GRU:

python sentiment_classifier.py --model gru --predict "The film was amazing" --embed_dim 200 --hidden_dim 256

Example output:

Text      : The film was amazing
Predicted : VERY POSITIVE

Class probabilities:
  very negative: 0.159
  negative     : 0.125
  neutral      : 0.140
  positive     : 0.244
  very positive: 0.333

Using BiGRU:

python sentiment_classifier.py --model bigru --predict "This is probably one of the worst films ever." --embed_dim 200 --hidden_dim 256

--------------------------------------------------------
  Text      : This is probably one of the worst films ever.
  Predicted : VERY NEGATIVE
--------------------------------------------------------
  Class probabilities:
    very negative    0.947  #################################
    negative         0.044  #
    neutral          0.007
    positive         0.002
    very positive    0.001

Results

GRU and BiGRU Results (available assets)

Training Curves (Dimension 200)

GRU GloVe Finetuned Dim200

GRU GloVe Frozen Dim200

GRU Random Dim200

BiGRU GloVe Finetuned Dim200

BiGRU GloVe Frozen Dim200

BiGRU Random Dim200

Experiment Tracking

The project uses MLflow for experiment tracking. All runs are logged to the mlruns/ directory. You can view experiments using:

mlflow ui --backend-store-uri ./mlruns

Development Notes

• Pre-commit

  We use pre-commit to automate linting of our codebase.

  • Install hooks:

    pre-commit install

  • Run Hooks manually (optional):

    pre-commit run --all-files

• Ruff:

  • Lint and format:

    ruff check --fix
    ruff format