1D-CNN Customer Complaint Classifier (PyTorch)

A PyTorch-based 1D Convolutional Neural Network for real-world text classification tasks. Includes tokenization, padded sequence handling, GPU-accelerated training, performance evaluation, and reproducible preprocessing utilities.

🚀 Quickstart Demo (For Reviewers)

Run a simple inference example using sample text:

pip install -r requirements.txt
python demo_predict.py

This confirms preprocessing, embedding, convolutional feature extraction, and classifier output.

---

📦 Full Training Run

Train the full 1D-CNN model:

python train_1d_cnn.py

This script handles:

• Tokenization + vocabulary
• Sequence batching
• CNN feature extraction
• Classification head training
• Train/validation split
• Accuracy reporting

---

📁 Files

train_1d_cnn.py   # Full supervised training script
run_demo.py               # Example inference demo
requirements.txt          # Dependencies

---

🏗 Overview

• Token Embedding → Conv1D → Global Max Pool → Linear Classifier
• Efficient for short-to-medium text sequences
• Demonstrates convolutional NLP modeling
• Fast training even on CPU

This architecture is commonly used for sentiment analysis and short-text classification.

---

📂 Project Structure

.
├── train_cnn_classifier.py
├── run_demo.py
├── requirements.txt
├── CONTRIBUTING.md
└── SECURITY.md

---

🧱 Architecture Overview

At a high level, the training system looks like this:

Input: [batch_size, seq_len] token IDs
        |
        v
Embedding layer -> [batch, seq_len, embed_dim]
        |
 Transpose to [batch, embed_dim, seq_len]
        |
        v
 Conv1d (kernel=3, padding=1) + ReLU
        |
        v
 Conv1d (kernel=5, padding=2) + ReLU
        |
        v
 AdaptiveMaxPool1d(output_size=1)
        |
        v
  Flatten -> [batch, 128]
        |
        v
   Dropout(0.3)
        |
        v
  Fully-connected layer -> [batch, num_classes]
Input: [batch_size, seq_len] token IDs
        |
        v
Embedding layer -> [batch, seq_len, embed_dim]
        |
 Transpose to [batch, embed_dim, seq_len]
        |
        v
 Conv1d (kernel=3, padding=1) + ReLU
        |
        v
 Conv1d (kernel=5, padding=2) + ReLU
        |
        v
 AdaptiveMaxPool1d(output_size=1)
        |
        v
  Flatten -> [batch, 128]
        |
        v
   Dropout(0.3)
        |
        v
  Fully-connected layer -> [batch, num_classes]

---

🤝 Contributing

See CONTRIBUTING.md for coding style, branching strategy, and PR workflow.

---

📄 License

MIT License. See LICENSE for details.