Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,109 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language: en
|
| 3 |
+
license: apache-2.0
|
| 4 |
+
tags:
|
| 5 |
+
- green-ai
|
| 6 |
+
- energy-efficiency
|
| 7 |
+
- e2am
|
| 8 |
+
- eden-framework
|
| 9 |
+
- sustainable-ai
|
| 10 |
+
- image-classification
|
| 11 |
+
---
|
| 12 |
+
|
| 13 |
+
# EDEN-Core-Scripts — E2AM Framework Repository
|
| 14 |
+
|
| 15 |
+
> **Project EDEN (Energy-Driven Evolution of Networks)** — The complete algorithmic
|
| 16 |
+
> toolkit for Green SOTA image classification research.
|
| 17 |
+
|
| 18 |
+
## Why EDEN?
|
| 19 |
+
As deep learning models scale exponentially, the carbon footprint of training has
|
| 20 |
+
reached unsustainable levels. Project EDEN introduces the **EAG
|
| 21 |
+
(Energy-to-Accuracy Gradient)** as the primary KPI — shifting the paradigm from
|
| 22 |
+
chasing raw accuracy to optimising *Green SOTA*.
|
| 23 |
+
|
| 24 |
+
## Profiling Environment
|
| 25 |
+
| Component | Specification |
|
| 26 |
+
|---|---|
|
| 27 |
+
| **GPU** | NVIDIA GeForce GTX 1080 Ti (11 GB VRAM, 250 W TDP) |
|
| 28 |
+
| **CPU** | Intel Xeon W-2125 (4 cores / 8 threads @ 4.00 GHz) |
|
| 29 |
+
| **RAM** | 63.66 GB System RAM |
|
| 30 |
+
| **OS** | Windows 10 |
|
| 31 |
+
|
| 32 |
+
## The E2AM Algorithm — All Three Phases
|
| 33 |
+
|
| 34 |
+
### Phase 1 — Zero-Overhead Initialization
|
| 35 |
+
Dataset pre-loaded into **pinned System RAM** before training begins.
|
| 36 |
+
This eliminates disk I/O power spikes that would otherwise inflate energy readings
|
| 37 |
+
and distort EAG comparisons between architectures.
|
| 38 |
+
|
| 39 |
+
### Phase 2 — Two-Stage Energy-Aware Training
|
| 40 |
+
1. **Frozen Head Training** — Only the classification head trains for the first
|
| 41 |
+
`E_unfreeze` epochs. The backbone consumes no backward-pass energy.
|
| 42 |
+
2. **Progressive Unfreezing** — At epoch `E_unfreeze`, all layers unlock.
|
| 43 |
+
Learning rate is decayed (`LR × 0.1`) for stable fine-tuning.
|
| 44 |
+
3. **Gradient Accumulation** — Gradients accumulated over N micro-batches,
|
| 45 |
+
simulating large batch sizes without VRAM spikes.
|
| 46 |
+
4. **AMP (Automated Mixed Precision)** — `torch.cuda.amp.autocast()` halves
|
| 47 |
+
bandwidth per backward pass.
|
| 48 |
+
5. **Sparse L1 Penalty** — `L_total = CrossEntropy + λ·Σ|W_trainable|`
|
| 49 |
+
6. **EAG Early-Exit** — Training terminates if `EAG < γ_EAG` for 3 consecutive
|
| 50 |
+
epochs, preventing wasted compute.
|
| 51 |
+
|
| 52 |
+
### Phase 3 — Hardware-Aware Deployment *(Post-Training)*
|
| 53 |
+
- **Saliency-Energy Pruning** — Filters with lowest `∂Accuracy/∂W ÷ Energy_cost`
|
| 54 |
+
are pruned.
|
| 55 |
+
- **INT8 Quantization** — Weights converted for edge-deployment readiness.
|
| 56 |
+
- **Dynamic Depth Routing** — Simple images bypass the middle 50 % of layers
|
| 57 |
+
via residual skip connections, slashing inference energy.
|
| 58 |
+
|
| 59 |
+
## EAG — The Expert KPI
|
| 60 |
+
```
|
| 61 |
+
EAG = ΔAccuracy / ΔJoules
|
| 62 |
+
```
|
| 63 |
+
EAG allows apples-to-apples comparison of any two models regardless of
|
| 64 |
+
architecture family. A higher EAG = more learning per unit of carbon footprint.
|
| 65 |
+
|
| 66 |
+
## Scripts in This Repository
|
| 67 |
+
- `eden_hf_upload.py`
|
| 68 |
+
- `test1\Algo_CIFAR_100_EfficientNet.py`
|
| 69 |
+
- `test1\Algo_CIFAR_100_MobileViTv3.py`
|
| 70 |
+
- `test1\Algo_CIFAR_100_convneXt.py`
|
| 71 |
+
- `test1\Algo_CIFAR_10_EfficientNet.py`
|
| 72 |
+
- `test1\Algo_CIFAR_10_MobileViTv3.py`
|
| 73 |
+
- `test1\Algo_CIFAR_10_convneXt.py`
|
| 74 |
+
- `test1\Algo_ImageNet_EfficientNet.py`
|
| 75 |
+
- `test1\Algo_ImageNet_convnext.py`
|
| 76 |
+
- `test1\Algo_ImageNet_mobilevit3.py`
|
| 77 |
+
- `test1\mobilevit_model.py`
|
| 78 |
+
- `test3\eden_AlexNet_CIFAR10.py`
|
| 79 |
+
- `test3\eden_AlexNet_CIFAR100.py`
|
| 80 |
+
- `test3\eden_AlexNet_ImageNet.py`
|
| 81 |
+
- `test3\eden_DenseNet_121_CIFAR10.py`
|
| 82 |
+
- `test3\eden_DenseNet_121_CIFAR100.py`
|
| 83 |
+
- `test3\eden_DenseNet_121_ImageNet.py`
|
| 84 |
+
- `test3\eden_InceptionV3_CIFAR10.py`
|
| 85 |
+
- `test3\eden_InceptionV3_CIFAR100.py`
|
| 86 |
+
- `test3\eden_InceptionV3_ImageNet.py`
|
| 87 |
+
- `test3\eden_ResNet18_CIFAR10.py`
|
| 88 |
+
- `test3\eden_ResNet18_CIFAR100.py`
|
| 89 |
+
- `test3\eden_ResNet18_ImageNet.py`
|
| 90 |
+
- `test3\eden_ResNet50_CIFAR10.py`
|
| 91 |
+
- `test3\eden_ResNet50_CIFAR100.py`
|
| 92 |
+
- `test3\eden_ResNet50_ImageNet.py`
|
| 93 |
+
- `test3\eden_UNet_CIFAR10.py`
|
| 94 |
+
- `test3\eden_UNet_CIFAR100.py`
|
| 95 |
+
- `test3\eden_UNet_ImageNet.py`
|
| 96 |
+
- `test3\eden_VGG16_CIFAR10.py`
|
| 97 |
+
- `test3\eden_VGG16_CIFAR100.py`
|
| 98 |
+
- `test3\eden_VGG16_ImageNet.py`
|
| 99 |
+
|
| 100 |
+
## Cite This Research
|
| 101 |
+
```bibtex
|
| 102 |
+
@misc{eden2025,
|
| 103 |
+
title = {Project EDEN: Energy-Driven Evolution of Networks},
|
| 104 |
+
author = {EDEN Research Team},
|
| 105 |
+
year = {2025},
|
| 106 |
+
note = {Hugging Face Organization: ProjectEDEN},
|
| 107 |
+
url = {https://huggingface.co/Shanmuk4622}
|
| 108 |
+
}
|
| 109 |
+
```
|