Collect representative data
Replace synthetic examples with a dataset that matches the target device inputs and pre-processing pipeline before pruning experiments.
01 / THE IDEA
Model pruning for mobile deployment in practice.
Model pruning removes redundant weights or nodes to reduce model size and inference cost for mobile devices. Practical pruning work includes measuring latency, memory use and accuracy trade-offs on representative edge inputs.
Real-world use case · Proposed extension
A proposed application is an on-device classifier for battery-powered sensors where reduced model size lowers memory and energy costs. Any deployed system should retain a human-in-the-loop for critical decisions and be validated on real device telemetry.
02 / THE BUILD
How to build a model pruning for mobile deployment project
step by step.
Train a reproducible baseline
Fit a simple, well-documented baseline model and record its accuracy, size and latency to serve as the pruning reference.
Apply pruning and quantisation
Iteratively prune weights or tree nodes, combine with quantisation, and measure accuracy and model file size after each change.
Validate on-device metrics
Measure latency, memory and battery cost on the target hardware and compare to acceptance thresholds; monitor post-deploy drift.
03 / SOURCE SNAPSHOT
What the model pruning for mobile deployment starter actually does.
The supplied starter trains a DecisionTreeClassifier on synthetic classification data, times repeated predictions on a single sample to estimate per-sample latency, and prints tree depth and node count. It does not perform pruning, quantisation, model export, or any mobile runtime evaluation.
SOURCE EXCERPTPython
X, y = make_classification(n_samples=1000, n_features=12, n_informative=8, random_state=42)
model = DecisionTreeClassifier(max_depth=6, random_state=42)
model.fit(X, y)
sample = X[:1]
start = time.perf_counter()Important limitation
The code uses synthetic data (make_classification) and a scikit-learn DecisionTree with no pruning or compression steps. It only measures a crude in-process latency by calling predict in a loop and does not export a compact model format (for example TensorFlow Lite or ONNX), run on-device, or evaluate accuracy on realistic edge inputs.
04 / BEYOND THE CLASSROOM
Where a model pruning for mobile deployment project could go next.
PROPOSED BUSINESS CONCEPT
A pruning and on-device validation pilot
A proposed extension: train a real model on representative sensor or camera data, apply magnitude or structured pruning and integer quantisation, export to TensorFlow Lite or ONNX, then measure latency and memory on the target device.
WHO IT COULD SERVEEdge-device teams
FIRST MVPA reproducible pipeline that trains a baseline, applies pruning+quantisation, exports a TFLite model, and reports on-device latency, model size and held-out accuracy.
KEY RISKPruning can reduce accuracy in edge conditions; validate on representative inputs and keep a human reviewer for decisions. Performance measured in-process does not ensure on-device behaviour such as thermal throttling or I/O overhead.
This is a potential application pathway, not a delivered product. The supplied starter does not implement this service or validate it with real users.
05 / WHAT YOU TAKE AWAY
What a model pruning for mobile deployment project puts on your resume.
- Prepare representative edge-device data and a reproducible baseline training run.
- Measure model size, inference latency and accuracy before and after pruning.
- Apply pruning and quantisation techniques and export a compact model format for deployment.
- Design on-device validation steps and acceptance thresholds for latency and memory.
Want to build this with mentors and a lab?
Network Engineering Course Networkers Home runs this topic inside its Network Engineering Course, with placement support and hands-on labs.
06 / QUESTIONS STUDENTS ASK
Model Pruning for Mobile Deployment project FAQs.
- Is this suitable as a final year project?
- Yes — this is an advanced final year project idea focused on reducing model size and inference cost for mobile devices; it suits students interested in model optimisation and embedded ML.
- What tools and knowledge do I need?
- Familiarity with Python, TensorFlow or scikit-learn, model compression concepts (pruning, quantisation), and basic profiling of latency and memory on target devices is recommended.
- How should I evaluate progress and effort required?
- Measure model size, per-sample latency, memory footprint and accuracy trade-offs on representative edge inputs; expect several weeks to prototype and more time to profile on real hardware.