Handwritten Character Classification Pipeline

Timeframe: Oct 2024 – Dec 2024
Stack: TensorFlow · Keras · TFRecords

Overview

Built an EMNIST classification pipeline using TFRecords to parse, pad 28×28 to 32×32, correct rotations, then shuffled, batched (128), and prefetched training samples to minimize memory footprint.

Approach

• Fine-tuned ResNet50V2-based classifier with dropout and batch normalization, unfreezing the top 30% of layers
• Employed mixed bfloat16 precision, dynamic learning-rate reduction, and checkpointing
• Optimized training efficiency by reducing required batches from ~4,171 to 1,500 through strategic shuffling and batching, without sacrificing accuracy

Results

• Achieved 93.55% test accuracy (0.1745 loss)
• Leveraged mixed precision plus multi-core parallelism for faster convergence