This repository contains a research proposal for evaluating model compression techniques to enable efficient, low-carbon ecological image classification on edge devices. This project proposes applying pruning and quantization methods to reduce model size and inference costs while maintaining high classification accuracy on ecological datasets like iNaturalist.
🔬 Status: Research Proposal (Pre-Implementation)
📋 Phase: Experimental Design & Methodology Planning
⏳ Expected Completion: [Add your expected timeline]
The following table represents expected outcomes based on literature review and similar studies. Actual results will be obtained upon project implementation:
| Model | Expected Accuracy | Expected Size (MB) | Expected Inference (ms) | Expected Energy (mJ) |
|---|---|---|---|---|
| Baseline ResNet-50 | ~92-94% | 350 | 150-200 | 40-50 |
| Pruned (50% sparsity) | ~91-93% | 175-190 | 80-100 | 22-28 |
| Quantized (int8) | ~91-92% | 85-95 | 50-70 | 15-20 |
| Pruned + Quantized | ~89-91% | 40-50 | 30-40 | 8-12 |
Source: Projections based on [pruning literature], [quantization benchmarks], and similar ecological classification studies.
- Dataset preparation (iNaturalist subset)
- Implement baseline model training pipeline
- Establish baseline performance metrics
- Implement structured pruning techniques
- Test different sparsity levels (30%, 50%, 70%)
- Evaluate accuracy-compression trade-offs
- Implement post-training quantization (PTQ)
- Implement quantization-aware training (QAT)
- Compare int8 and int4 precision levels
- Combine pruning + quantization
- Comprehensive benchmarking on edge devices
- Carbon footprint analysis
- Create deployment guides for edge devices
- Document findings and methodology
- Publish results