Shanmuk4622/E2AM_ConvNeXtV2Nano

E2AM Ablation Results: ConvNeXtV2-Nano

Energy-aware training ablation study for ConvNeXtV2-Nano 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. 15 deployment runs.

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	M1_cache_only	0.7851	0.9859	0.0284	0.0135	1406
CIFAR-100	M5_adaptive_lr_only	0.4710	0.7259	0.0280	0.0133	1381
Tiny-ImageNet	C4_cache_amp_gradaccum_adaptivelr	0.4128	0.6502	0.0705	0.0335	3423

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

Deployment Pareto 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	0.7817	0.9867	0.0300	0.0142	1480	completed
M1_cache_only	50	0.7851	0.9859	0.0284	0.0135	1406	completed
M2_amp_only	50	0.7804	0.9854	0.0188	0.0089	955	completed
M3_grad_accum_only	50	0.7827	0.9864	0.0267	0.0127	1318	completed
M4_l1_sparsity_only	50	0.7823	0.9868	0.0328	0.0156	1624	completed
M5_adaptive_lr_only	50	0.7822	0.9851	0.0284	0.0135	1411	completed
M6_eag_only	50	0.7839	0.9857	0.0295	0.0140	1459	completed
M7_full_e2am	50	0.7779	0.9845	0.0206	0.0098	1041	completed

C-matrix (cumulative ablation)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
C0_baseline	50	0.7837	0.9862	0.0282	0.0134	1388	completed
C1_cache	50	0.7815	0.9866	0.0280	0.0133	1376	completed
C2_cache_amp	50	0.7750	0.9860	0.0182	0.0087	936	completed
C3_cache_amp_gradaccum	50	0.7750	0.9860	0.0184	0.0087	910	completed
C4_cache_amp_gradaccum_adaptivelr	50	0.7773	0.9851	0.0183	0.0087	904	completed
C5_cache_amp_gradaccum_adaptivelr_l1	50	0.7762	0.9845	0.0204	0.0097	1011	completed
C6_full_e2am	50	0.7762	0.9845	0.0204	0.0097	1022	completed

CIFAR-100

M-matrix (individual methods)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
M0_baseline_fp32	50	0.4492	0.7256	0.0280	0.0133	1382	completed
M1_cache_only	50	0.4502	0.7255	0.0276	0.0131	1366	completed
M2_amp_only	50	0.4399	0.7128	0.0179	0.0085	932	completed
M3_grad_accum_only	50	0.4452	0.7145	0.0259	0.0123	1277	completed
M4_l1_sparsity_only	50	0.4557	0.7283	0.0320	0.0152	1575	completed
M5_adaptive_lr_only	50	0.4710	0.7259	0.0280	0.0133	1381	completed
M6_eag_only	50	0.4498	0.7282	0.0285	0.0135	1406	completed
M7_full_e2am	50	0.4551	0.7131	0.0204	0.0097	1015	completed

C-matrix (cumulative ablation)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
C0_baseline	50	0.4537	0.7289	0.0280	0.0133	1387	completed
C1_cache	50	0.4503	0.7285	0.0279	0.0132	1377	completed
C2_cache_amp	50	0.4424	0.7118	0.0180	0.0085	933	completed
C3_cache_amp_gradaccum	50	0.4424	0.7118	0.0177	0.0084	893	completed
C4_cache_amp_gradaccum_adaptivelr	50	0.4561	0.7133	0.0179	0.0085	897	completed
C5_cache_amp_gradaccum_adaptivelr_l1	50	0.4594	0.7138	0.0198	0.0094	996	completed
C6_full_e2am	50	0.4594	0.7138	0.0198	0.0094	995	completed

Tiny-ImageNet

M-matrix (individual methods)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
M0_baseline_fp32	50	0.3841	0.6446	0.1289	0.0612	6154	completed
M1_cache_only	50	0.3799	0.6444	0.1260	0.0598	6057	completed
M2_amp_only	50	0.3809	0.6442	0.0721	0.0343	3522	completed
M3_grad_accum_only	50	0.3873	0.6450	0.1223	0.0581	5871	completed
M4_l1_sparsity_only	50	0.3827	0.6442	0.1308	0.0621	6316	completed
M5_adaptive_lr_only	50	0.3979	0.6490	0.1265	0.0601	6077	completed
M6_eag_only	50	0.3812	0.6450	0.1255	0.0596	6021	completed
M7_full_e2am	50	0.4042	0.6488	0.0744	0.0354	3613	completed

C-matrix (cumulative ablation)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
C0_baseline	50	0.3944	0.6514	0.1255	0.0596	6008	completed
C1_cache	50	0.3919	0.6548	0.1249	0.0593	5986	completed
C2_cache_amp	50	0.3875	0.6495	0.0713	0.0339	3472	completed
C3_cache_amp_gradaccum	50	0.3875	0.6495	0.0709	0.0337	3446	completed
C4_cache_amp_gradaccum_adaptivelr	50	0.4128	0.6502	0.0705	0.0335	3423	completed
C5_cache_amp_gradaccum_adaptivelr_l1	50	0.4065	0.6485	0.0739	0.0351	3590	completed
C6_full_e2am	50	0.4065	0.6485	0.0730	0.0347	3534	completed

Deployment results

Five deployment variants applied to the best training checkpoint per dataset:

• D0_fp32: baseline FP32 model
• D1_pruned_masked: 50% L2-norm structured pruning, NO recovery fine-tuning
• D2_pruned_finetuned: same as D1 + 3 epochs of recovery fine-tuning
• D3_int8_cpu_fx: CPU FX-graph INT8 static quantization (fbgemm backend)
• D4_pruned_int8: D1 pipeline + INT8

Deployment on CIFAR-10

Variant	Accuracy	Size (MB)	Latency (ms)	Throughput (img/s)	Energy/inf (J)
D0_fp32	0.7795	57.2	0.09	10880	0.0000
D1_pruned_masked	0.2381	57.2	0.09	11173	0.0000
D2_pruned_finetuned	0.7340	57.2	0.08	12694	0.0000
D3_int8_cpu_fx	0.7805	15.1	3.76	266	0.1502
D4_pruned_int8	0.2645	15.1	3.70	270	0.1481

Deployment on CIFAR-100

Variant	Accuracy	Size (MB)	Latency (ms)	Throughput (img/s)	Energy/inf (J)
D0_fp32	0.4564	57.5	0.10	10363	0.0000
D1_pruned_masked	0.0268	57.5	0.09	10825	0.0000
D2_pruned_finetuned	0.4168	57.5	0.08	12852	0.0000
D3_int8_cpu_fx	0.4395	15.2	3.47	288	0.1403
D4_pruned_int8	0.0246	15.2	3.54	282	0.1408

Deployment on Tiny-ImageNet

Variant	Accuracy	Size (MB)	Latency (ms)	Throughput (img/s)	Energy/inf (J)
D0_fp32	0.4023	57.7	0.30	3348	0.0157
D1_pruned_masked	0.0283	57.7	0.28	3591	0.0156
D2_pruned_finetuned	0.3553	57.7	0.29	3494	0.0156
D3_int8_cpu_fx	0.4199	15.2	9.10	110	0.3652
D4_pruned_int8	0.0129	15.2	8.94	112	0.3626

Methodology

Model: ConvNeXtV2-Nano (~15.0M params).

Training protocol: from scratch, AdamW (lr=6e-4, weight_decay=0.05, grad_clip=1.0), 50 epochs, warmup 3-5 epochs. SGD@0.1 causes complete non-convergence on ConvNeXtV2 (LayerNorm+GRN incompatibility); AdamW is required and used consistently across all variants for fair ablation comparison.

Input: native dataset resolution (32x32 for CIFAR-10/100, 64x64 for Tiny-ImageNet) fed directly — no upsample wrapper needed. ConvNeXtV2 uses global average pooling before the classifier, making it spatially flexible. FX-traceable: D3/D4 INT8 quantization fully supported.

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_ConvNeXtV2Nano --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_convnextv2nano,
  title  = {E2AM: Energy-Aware Adaptive Model Training Ablation Study (ConvNeXtV2-Nano)},
  author = {Shanmuk},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Shanmuk4622/E2AM_ConvNeXtV2Nano}},
}

---

This README was auto-generated on 2026-07-01 16:58 UTC. Source repo: Shanmuk4622/E2AM_ConvNeXtV2Nano