Model Compression Techniques Explained

Fundamentals

Motivation

Compression reduces memory and latency while preserving accuracy for edge and production deployments.

Pruning

Unstructured vs structured pruning; magnitude-based and gradient-based approaches; retraining for recovery.

Quantization

Post-training vs quantization-aware training; per-tensor vs per-channel scales; int8, fp16, mixed precision.

Distillation

Teacher-student training; soft labels and intermediate feature matching; task-specific strategies.

Deployment

Backends and toolchains: ONNX Runtime, TensorRT, TFLite; calibration and validation pipelines.

How the Algorithm Works

Pruning Loop

1. Train baseline to convergence
2. Compute saliencies (magnitude/gradients)
3. Prune by mask (structured/unstructured)
4. Fine-tune and evaluate
5. Iterate until target sparsity or KPI met

Quantization Loop

1. Select scheme (PTQ/QAT; per-tensor/per-channel)
2. Calibrate scales with representative data
3. Apply quantization; optionally QAT
4. Validate KPIs; adjust bitwidth/observers

Distillation Loop

1. Choose teacher/student and loss (soft labels/features)
2. Train student with distillation objective
3. Evaluate; iterate with schedule and temperature

Real-World Applications

Edge Deployment

Latency and memory constraints dominate; quantization and structured pruning crucial for ARM/embedded targets.

Cloud Inference

Throughput and cost optimization; distillation to smaller backbones; GPU/CPU mixed serving.

Monitoring

Track accuracy, latency, memory, and drift; rollback on KPI regressions.

Best Practices

Checklist

• Establish KPIs and guardrails before compression
• Sequence: prune → distill → quantize (adjust as needed)
• Benchmark on target hardware
• Track masks, scales, and artifacts
• Validate fairness and calibration
• Govern with audit logs and approvals

Anti-Patterns

• Compressing without a strong baseline
• Relying on proxy latency only
• Skipping QAT where PTQ fails
• No rollback plan
• Untracked changes to toolchains

Evaluation

Metrics

• Accuracy/PR-AUC; calibration (ECE)
• Latency (P50/P95), throughput
• Memory footprint and model size
• Energy (joules/inference) where applicable

Protocols

• Benchmark on target hardware; avoid proxy-only measures
• Golden tests and numerical tolerances
• Fairness and drift checks

Worked Examples

Pruning + Fine-Tuning

# Iterative magnitude pruning with fine-tuning cycles

PTQ + QAT

# Post-training quantization then short QAT refinement

Knowledge Distillation

# Teacher-student with KL loss and intermediate layers

Implementation

PyTorch Pruning

# torch.nn.utils.prune examples for structured/unstructured

Quantization

# PyTorch QAT/PTQ; TFLite converter; TensorRT INT8 calibration

Distillation

# Teacher-student training with KL loss and feature matching

Deployment

• Export ONNX/TFLite; validate with golden tests
• Benchmark latency on target hardware
• Track artifacts and calibration data

The Math Behind It

Sparsity

Sparsity induces regularization; structured sparsity preserves tensor shapes enabling efficient kernels.

Quantization Error

Uniform quantization bounds error by half-step; per-channel scaling reduces variance for convolutional weights.

Distillation Losses

KL divergence with temperature; intermediate feature matching stabilizes training.

Key Parameters

Constraints

Latency, memory, and accuracy guardrails; layer-wise exceptions for sensitive blocks.

Training Strategy

Fine-Tuning after Pruning

• Lower LR, longer patience
• Recover accuracy over 10–30% of baseline epochs

QAT

• Insert fake-quant modules and observers
• Train with gradually reduced LR

Monitoring

• KPIs: accuracy, latency, memory, energy
• Per-layer sensitivity and outliers