Deep Learning / NLP·PROJECT_05
Deep Learning NLP Classifier
Domain
Deep Learning / NLP
Role
ML Engineer
Stack
Python · TensorFlow · Keras · IMDb Dataset
Year
2026
Overview
End-to-end binary sentiment classifier for movie reviews using custom word embeddings and Global Average Pooling. Achieves 85.1% accuracy on 25,000 unseen reviews with a 626KB model that trains in under 10 seconds on CPU.
PythonTensorFlowKerasNLPWord EmbeddingsDeep Learning
The Challenge
Millions of reviews are generated daily across platforms. Manual sentiment inspection at scale is impossible. The goal was a lightweight, fast, fully offline-capable model.
System Architecture
Embedding layer + GlobalAveragePooling1D avoids expensive RNNs/LSTMs. Standalone dataset builder stores IMDb as binary .npy arrays for 100% offline use. Custom inference pipeline handles text normalization, tokenization, and sigmoid confidence scoring.
Validation Metrics
Evaluation Dataset
25,000 Reviews (IMDb Test Split)
Accuracy / Loss
85.10% / 0.3703 Binary CE
Model Footprint
626 KB (Zero Cloud Deps)
CPU Inference
<8ms per text sample
Key Highlights
100% offline — no API calls, no internet dependency during inference.
Lightweight architecture avoids RNNs for 10x faster training.
Custom predict_sentiment() with full text normalization pipeline.