---
language: en
license: apache-2.0
tags:
- green-ai
- energy-efficiency
- e2am
- eden-framework
- sustainable-ai
- image-classification
---

# EDEN-Core-Scripts — E2AM Framework Repository

> **Project EDEN (Energy-Driven Evolution of Networks)** — The complete algorithmic
> toolkit for Green SOTA image classification research.

## Why EDEN?
As deep learning models scale exponentially, the carbon footprint of training has
reached unsustainable levels. Project EDEN introduces the **EAG
(Energy-to-Accuracy Gradient)** as the primary KPI — shifting the paradigm from
chasing raw accuracy to optimising *Green SOTA*.

## Profiling Environment
| Component | Specification |
|---|---|
| **GPU** | NVIDIA GeForce GTX 1080 Ti (11 GB VRAM, 250 W TDP) |
| **CPU** | Intel Xeon W-2125 (4 cores / 8 threads @ 4.00 GHz) |
| **RAM** | 63.66 GB System RAM |
| **OS**  | Windows 10 |

## The E2AM Algorithm — All Three Phases

### Phase 1 — Zero-Overhead Initialization
Dataset pre-loaded into **pinned System RAM** before training begins.
This eliminates disk I/O power spikes that would otherwise inflate energy readings
and distort EAG comparisons between architectures.

### Phase 2 — Two-Stage Energy-Aware Training
1. **Frozen Head Training** — Only the classification head trains for the first
   `E_unfreeze` epochs. The backbone consumes no backward-pass energy.
2. **Progressive Unfreezing** — At epoch `E_unfreeze`, all layers unlock.
   Learning rate is decayed (`LR × 0.1`) for stable fine-tuning.
3. **Gradient Accumulation** — Gradients accumulated over N micro-batches,
   simulating large batch sizes without VRAM spikes.
4. **AMP (Automated Mixed Precision)** — `torch.cuda.amp.autocast()` halves
   bandwidth per backward pass.
5. **Sparse L1 Penalty** — `L_total = CrossEntropy + λ·Σ|W_trainable|`
6. **EAG Early-Exit** — Training terminates if `EAG < γ_EAG` for 3 consecutive
   epochs, preventing wasted compute.

### Phase 3 — Hardware-Aware Deployment *(Post-Training)*
- **Saliency-Energy Pruning** — Filters with lowest `∂Accuracy/∂W ÷ Energy_cost`
  are pruned.
- **INT8 Quantization** — Weights converted for edge-deployment readiness.
- **Dynamic Depth Routing** — Simple images bypass the middle 50 % of layers
  via residual skip connections, slashing inference energy.

## EAG — The Expert KPI
```
EAG = ΔAccuracy / ΔJoules
```
EAG allows apples-to-apples comparison of any two models regardless of
architecture family. A higher EAG = more learning per unit of carbon footprint.

## Scripts in This Repository
- `eden_hf_upload.py`
- `test1\Algo_CIFAR_100_EfficientNet.py`
- `test1\Algo_CIFAR_100_MobileViTv3.py`
- `test1\Algo_CIFAR_100_convneXt.py`
- `test1\Algo_CIFAR_10_EfficientNet.py`
- `test1\Algo_CIFAR_10_MobileViTv3.py`
- `test1\Algo_CIFAR_10_convneXt.py`
- `test1\Algo_ImageNet_EfficientNet.py`
- `test1\Algo_ImageNet_convnext.py`
- `test1\Algo_ImageNet_mobilevit3.py`
- `test1\mobilevit_model.py`
- `test3\eden_AlexNet_CIFAR10.py`
- `test3\eden_AlexNet_CIFAR100.py`
- `test3\eden_AlexNet_ImageNet.py`
- `test3\eden_DenseNet_121_CIFAR10.py`
- `test3\eden_DenseNet_121_CIFAR100.py`
- `test3\eden_DenseNet_121_ImageNet.py`
- `test3\eden_InceptionV3_CIFAR10.py`
- `test3\eden_InceptionV3_CIFAR100.py`
- `test3\eden_InceptionV3_ImageNet.py`
- `test3\eden_ResNet18_CIFAR10.py`
- `test3\eden_ResNet18_CIFAR100.py`
- `test3\eden_ResNet18_ImageNet.py`
- `test3\eden_ResNet50_CIFAR10.py`
- `test3\eden_ResNet50_CIFAR100.py`
- `test3\eden_ResNet50_ImageNet.py`
- `test3\eden_UNet_CIFAR10.py`
- `test3\eden_UNet_CIFAR100.py`
- `test3\eden_UNet_ImageNet.py`
- `test3\eden_VGG16_CIFAR10.py`
- `test3\eden_VGG16_CIFAR100.py`
- `test3\eden_VGG16_ImageNet.py`

## Cite This Research
```bibtex
@misc{eden2025,
  title     = {Project EDEN: Energy-Driven Evolution of Networks},
  author    = {EDEN Research Team},
  year      = {2025},
  note      = {Hugging Face Organization: ProjectEDEN},
  url       = {https://huggingface.co/Shanmuk4622}
}
```