Shanmuk4622/E2AM_EfficientNetV2_S

E2AM Ablation Results: EfficientNetV2-S

Energy-aware training ablation study for EfficientNetV2-S across three image-classification datasets: CIFAR-10, CIFAR-100, and Tiny-ImageNet.

Each dataset has 15 training variants (8 individual-method M0..M7, 7 cumulative ablation C0..C6) at 50 epochs, plus a 5-variant deployment pipeline (FP32 baseline, structured pruning, pruning+finetune, INT8 quantization, pruned+INT8).

Status: 45 completed variants, 0 partial.

Quick links

• Methodology
• Headline results
• Cross-dataset comparison
• Per-dataset results
• Deployment results
• Reproducibility

Headline results

Dataset	Best variant	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (sec)
CIFAR-10	C5_cache_amp_gradaccum_adaptivelr_l1	0.9040	0.9977	0.1006	0.0478	4909
CIFAR-100	C4_cache_amp_gradaccum_adaptivelr	0.7096	0.9251	0.0954	0.0453	4587
Tiny-ImageNet	M7_full_e2am	0.5805	0.8162	0.2005	0.0952	9683

Cross-dataset comparison

How the same training variants behave across CIFAR-10, CIFAR-100, and Tiny-ImageNet.

Accuracy By Variant Across Datasets

Energy By Variant Across Datasets

Per-dataset results

CIFAR-10

M-matrix (individual methods)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
M0_baseline_fp32	50	0.7565	0.9839	0.2080	0.0988	9714	completed
M1_cache_only	50	0.7582	0.9851	0.1963	0.0932	9440	completed
M2_amp_only	50	0.8013	0.9909	0.0986	0.0469	4775	completed
M3_grad_accum_only	50	0.8036	0.9903	0.1955	0.0928	9403	completed
M4_l1_sparsity_only	50	0.7394	0.9837	0.2097	0.0996	9834	completed
M5_adaptive_lr_only	50	0.8841	0.9969	0.1963	0.0933	9454	completed
M6_eag_only	50	0.7590	0.9854	0.2080	0.0988	9713	completed
M7_full_e2am	50	0.8974	0.9967	0.1005	0.0477	4847	completed

C-matrix (cumulative ablation)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
C0_baseline	50	0.7569	0.9865	0.1988	0.0944	9591	completed
C1_cache	50	0.7472	0.9862	0.1985	0.0943	9556	completed
C2_cache_amp	50	0.8068	0.9913	0.0988	0.0469	4824	completed
C3_cache_amp_gradaccum	50	0.8345	0.9937	0.0988	0.0469	4810	completed
C4_cache_amp_gradaccum_adaptivelr	50	0.8973	0.9971	0.0989	0.0470	4825	completed
C5_cache_amp_gradaccum_adaptivelr_l1	50	0.9040	0.9977	0.1006	0.0478	4909	completed
C6_full_e2am	50	0.9040	0.9977	0.1007	0.0478	4915	completed

CIFAR-100

M-matrix (individual methods)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
M0_baseline_fp32	50	0.3900	0.7104	0.1990	0.0945	9332	completed
M1_cache_only	50	0.4190	0.7465	0.1992	0.0946	9689	completed
M2_amp_only	50	0.5240	0.8238	0.0956	0.0454	4608	completed
M3_grad_accum_only	50	0.5408	0.8365	0.1984	0.0943	9629	completed
M4_l1_sparsity_only	50	0.4378	0.7671	0.2019	0.0959	9629	completed
M5_adaptive_lr_only	50	0.5975	0.8641	0.1995	0.0948	9678	completed
M6_eag_only	50	0.4010	0.7281	0.1999	0.0950	9520	completed
M7_full_e2am	50	0.7090	0.9254	0.1025	0.0487	4970	completed

C-matrix (cumulative ablation)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
C0_baseline	50	0.4075	0.7359	0.1999	0.0950	9529	completed
C1_cache	50	0.4502	0.7700	0.1998	0.0949	9533	completed
C2_cache_amp	50	0.5411	0.8401	0.0982	0.0466	4756	completed
C3_cache_amp_gradaccum	50	0.6111	0.8775	0.0953	0.0453	4578	completed
C4_cache_amp_gradaccum_adaptivelr	50	0.7096	0.9251	0.0954	0.0453	4587	completed
C5_cache_amp_gradaccum_adaptivelr_l1	50	0.7016	0.9205	0.0968	0.0460	4672	completed
C6_full_e2am	50	0.7017	0.9191	0.0999	0.0475	4812	completed

Tiny-ImageNet

M-matrix (individual methods)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
M0_baseline_fp32	50	0.2462	0.5170	0.3952	0.1877	18949	completed
M1_cache_only	50	0.2451	0.5128	0.3944	0.1873	19008	completed
M2_amp_only	50	0.3541	0.6305	0.1937	0.0920	9387	completed
M3_grad_accum_only	50	0.3467	0.6401	0.3921	0.1863	18791	completed
M4_l1_sparsity_only	50	0.2440	0.5172	0.4054	0.1926	19502	completed
M5_adaptive_lr_only	50	0.4611	0.7419	0.3979	0.1890	19208	completed
M6_eag_only	50	0.2606	0.5416	0.3992	0.1896	19319	completed
M7_full_e2am	50	0.5805	0.8162	0.2005	0.0952	9683	completed

C-matrix (cumulative ablation)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
C0_baseline	50	0.2475	0.5199	0.4119	0.1957	19800	completed
C1_cache	50	0.2502	0.5186	0.4014	0.1907	19312	completed
C2_cache_amp	50	0.3465	0.6300	0.1980	0.0940	9633	completed
C3_cache_amp_gradaccum	50	0.4436	0.7242	0.2012	0.0956	9766	completed
C4_cache_amp_gradaccum_adaptivelr	50	0.5738	0.8162	0.2026	0.0963	9831	completed
C5_cache_amp_gradaccum_adaptivelr_l1	50	0.5708	0.8125	0.2053	0.0975	9943	completed
C6_full_e2am	50	0.5708	0.8125	0.2053	0.0975	9939	completed

Deployment results

No deployment results in this repo yet.

Methodology

Model: EfficientNetV2-S (~20.2M params).

Training protocol: from scratch, SGD with momentum 0.9, weight decay 5e-4, initial LR 0.1, 50 epochs, 1 warmup epoch. All variants share the same protocol so ablation comparison stays apples-to-apples across the matrix.

Input: native dataset resolution upsampled to 128x128 in-model via nn.Upsample (FX-traceable to keep D3/D4 INT8 quantization possible).

Optimization toggles (the 5 individual methods and their cumulative combinations):

Method	Mechanism
Tensor cache	Training images held in RAM as a normalized float tensor
AMP	torch.cuda.amp.autocast + GradScaler
Grad accum (x2)	Accumulate gradients across 2 mini-batches
L1 sparsity	Lambda * sum(
Cosine LR	lr(t) = lr_max * 0.5 * (1 + cos(pi*t/T))
EAG early-stop	Energy-Aware Gain: stop when accuracy gain per joule plateaus

Energy measurement: GPU power sampled at 1 Hz via nvidia-smi --query-gpu=power.draw. Energy = trapezoidal integration over power-vs-time. CO₂ = energy_kWh * 0.475 (global average grid intensity).

Hardware: Single NVIDIA T4 (14.5 GB) on Kaggle.

Repository structure

runs/
  cifar10/
  cifar100/
  tiny_imagenet/
    individual_methods/M0..M7/   (history.csv, metrics_summary.json,
                                  best_model.pt, last_model.pt, config.yaml)
    cumulative_ablation/C0..C6/  (same)
paper_tables/                     (6 unified CSV tables)
comparison_plots/<dataset>/       (per-dataset plots)
comparison_plots/cross_dataset/   (cross-dataset plots)
README.md                         (this file)

Reproducibility

Each variant directory has a config.yaml with the exact configuration used. To reproduce:

1. huggingface-cli download Shanmuk4622/E2AM_EfficientNetV2_S --repo-type dataset
2. Load the e2am.py library and call the appropriate config factory
3. Run e2am.train_one_run(cfg)

Limitations

• Energy measurement is GPU-only (via nvidia-smi); CPU/memory power not included
• Pruning is mask-based; no wall-clock speedup without sparsity-aware runtime
• INT8 (D3/D4) is CPU FX static quantization (fbgemm); may fail on transformer blocks. Failures logged in metrics.json rather than crashing.
• Single-T4 reproduction; multi-GPU not validated
• SGD@0.1 is suboptimal for some architectures; the paper compares variant-to-variant deltas which remain meaningful regardless

Citation

@misc{e2am_ablation_effnetv2s,
  title  = {E2AM: Energy-Aware Adaptive Model Training Ablation Study (EfficientNetV2-S)},
  author = {Shanmuk},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Shanmuk4622/E2AM_EfficientNetV2_S}},
}

---

This README was auto-generated on 2026-06-06 12:40 UTC. Source repo: Shanmuk4622/E2AM_EfficientNetV2_S