Shanmuk4622/E2AM_MobileViTv2

E2AM Ablation Results: MobileViTv2

Energy-aware training ablation study for MobileViTv2 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	C6_full_e2am	0.9298	0.9980	0.2535	0.1204	12086
CIFAR-100	M7_full_e2am	0.7307	0.9269	0.2757	0.1310	13298
Tiny-ImageNet	C5_cache_amp_gradaccum_adaptivelr_l1	0.5857	0.8120	0.5205	0.2472	24838

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	0.6971	0.9752	0.4648	0.2208	22138	completed
M1_cache_only	50	0.6739	0.9726	0.4954	0.2353	23560	completed
M2_amp_only	50	0.1000	0.5000	0.2638	0.1253	12605	completed
M3_grad_accum_only	50	0.7694	0.9864	0.5196	0.2468	24865	completed
M4_l1_sparsity_only	50	0.6888	0.9756	0.5234	0.2486	25183	completed
M5_adaptive_lr_only	50	0.8682	0.9940	0.5234	0.2486	24931	completed
M6_eag_only	50	0.6509	0.9743	0.4800	0.2280	22937	completed
M7_full_e2am	50	0.9219	0.9980	0.2754	0.1308	13151	completed

C-matrix (cumulative ablation)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
C0_baseline	50	0.6561	0.9710	0.4899	0.2327	23158	completed
C1_cache	50	0.6710	0.9725	0.5473	0.2600	26205	completed
C2_cache_amp	50	0.7397	0.9827	0.2429	0.1154	11595	completed
C3_cache_amp_gradaccum	50	0.8276	0.9921	0.2462	0.1169	11765	completed
C4_cache_amp_gradaccum_adaptivelr	50	0.9285	0.9990	0.2477	0.1177	11828	completed
C5_cache_amp_gradaccum_adaptivelr_l1	50	0.9206	0.9981	0.2560	0.1216	12257	completed
C6_full_e2am	50	0.9298	0.9980	0.2535	0.1204	12086	completed

CIFAR-100

M-matrix (individual methods)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
M0_baseline_fp32	50	0.2644	0.5753	0.5027	0.2388	23952	completed
M1_cache_only	50	0.2719	0.5927	0.5019	0.2384	23896	completed
M2_amp_only	50	0.4428	0.7590	0.2658	0.1263	12851	completed
M3_grad_accum_only	50	0.4483	0.7690	0.4882	0.2319	23150	completed
M4_l1_sparsity_only	50	0.2560	0.5766	0.5020	0.2385	23947	completed
M5_adaptive_lr_only	50	0.5651	0.8552	0.5154	0.2448	24572	completed
M6_eag_only	50	0.2914	0.6035	0.5044	0.2396	24011	completed
M7_full_e2am	50	0.7307	0.9269	0.2757	0.1310	13298	completed

C-matrix (cumulative ablation)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
C0_baseline	50	0.2840	0.6029	0.4794	0.2277	22881	completed
C1_cache	50	0.2951	0.6182	0.4643	0.2205	22035	completed
C2_cache_amp	50	0.4484	0.7720	0.2340	0.1111	11077	completed
C3_cache_amp_gradaccum	50	0.5739	0.8600	0.2348	0.1115	11113	completed
C4_cache_amp_gradaccum_adaptivelr	50	0.7180	0.9244	0.2541	0.1207	11852	completed
C5_cache_amp_gradaccum_adaptivelr_l1	50	0.7253	0.9295	0.2617	0.1243	12605	completed
C6_full_e2am	50	0.7232	0.9256	0.2760	0.1311	13333	completed

Tiny-ImageNet

M-matrix (individual methods)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
M0_baseline_fp32	50	0.1359	0.3553	0.9992	0.4746	47630	completed
M1_cache_only	50	0.1416	0.3694	0.9939	0.4721	46972	completed
M2_amp_only	50	0.2424	0.5102	0.5247	0.2492	25279	completed
M3_grad_accum_only	50	0.2640	0.5354	0.9653	0.4585	45545	completed
M4_l1_sparsity_only	50	0.1391	0.3658	1.0428	0.4953	49938	completed
M5_adaptive_lr_only	50	0.3614	0.6346	0.9706	0.4610	45838	completed
M6_eag_only	50	0.1279	0.3483	0.9772	0.4642	46600	completed
M7_full_e2am	50	0.5835	0.8155	0.5267	0.2502	25137	completed

C-matrix (cumulative ablation)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
C0_baseline	50	0.1374	0.3499	1.0254	0.4871	48708	completed
C1_cache	50	0.1351	0.3695	0.9546	0.4534	45416	completed
C2_cache_amp	50	0.2375	0.4999	0.4932	0.2343	23429	completed
C3_cache_amp_gradaccum	50	0.3793	0.6632	0.4941	0.2347	23520	completed
C4_cache_amp_gradaccum_adaptivelr	50	0.5824	0.8136	0.5018	0.2384	23976	completed
C5_cache_amp_gradaccum_adaptivelr_l1	50	0.5857	0.8120	0.5205	0.2472	24838	completed
C6_full_e2am	50	0.5761	0.8129	0.5185	0.2463	24824	completed

Deployment results

No deployment results in this repo yet.

Methodology

Model: MobileViTv2 (~4.5M 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 256x256 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_MobileViTv2 --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_mobilevitv2,
  title  = {E2AM: Energy-Aware Adaptive Model Training Ablation Study (MobileViTv2)},
  author = {Shanmuk},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Shanmuk4622/E2AM_MobileViTv2}},
}

---

This README was auto-generated on 2026-06-06 12:37 UTC. Source repo: Shanmuk4622/E2AM_MobileViTv2