Shanmuk4622/E2AM_ConvNeXtV2Tiny

E2AM Ablation Results: ConvNeXtV2-Tiny

Energy-aware training ablation study for ConvNeXtV2-Tiny 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	C1_cache	0.7877	0.9860	0.0431	0.0205	2108
CIFAR-100	M5_adaptive_lr_only	0.4948	0.7461	0.0429	0.0204	2102
Tiny-ImageNet	M5_adaptive_lr_only	0.4153	0.6548	0.2161	0.1026	10153

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	0.7781	0.9855	0.0453	0.0215	2194	completed
M1_cache_only	50	0.7796	0.9860	0.0440	0.0209	2139	completed
M2_amp_only	50	0.7770	0.9854	0.0253	0.0120	1278	completed
M3_grad_accum_only	50	0.7671	0.9848	0.0401	0.0190	1966	completed
M4_l1_sparsity_only	50	0.7721	0.9839	0.0506	0.0241	2468	completed
M5_adaptive_lr_only	50	0.7706	0.9835	0.0450	0.0214	2190	completed
M6_eag_only	50	0.7734	0.9853	0.0461	0.0219	2228	completed
M7_full_e2am	50	0.7795	0.9851	0.0280	0.0133	1391	completed

C-matrix (cumulative ablation)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
C0_baseline	50	0.7804	0.9850	0.0441	0.0210	2163	completed
C1_cache	50	0.7877	0.9860	0.0431	0.0205	2108	completed
C2_cache_amp	50	0.7661	0.9833	0.0248	0.0118	1256	completed
C3_cache_amp_gradaccum	50	0.7835	0.9850	0.0226	0.0108	1133	completed
C4_cache_amp_gradaccum_adaptivelr	50	0.7793	0.9854	0.0229	0.0109	1133	completed
C5_cache_amp_gradaccum_adaptivelr_l1	50	0.7792	0.9853	0.0253	0.0120	1251	completed
C6_full_e2am	50	0.7796	0.9851	0.0255	0.0121	1291	completed

CIFAR-100

M-matrix (individual methods)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
M0_baseline_fp32	50	0.4672	0.7438	0.0427	0.0203	2069	completed
M1_cache_only	50	0.4666	0.7456	0.0428	0.0203	2077	completed
M2_amp_only	50	0.4638	0.7459	0.0251	0.0119	1258	completed
M3_grad_accum_only	50	0.4684	0.7451	0.0386	0.0184	1888	completed
M4_l1_sparsity_only	50	0.4659	0.7418	0.0481	0.0228	2339	completed
M5_adaptive_lr_only	50	0.4948	0.7461	0.0429	0.0204	2102	completed
M6_eag_only	50	0.4637	0.7456	0.0428	0.0203	2073	completed
M7_full_e2am	50	0.4463	0.7041	0.0255	0.0121	1268	completed

C-matrix (cumulative ablation)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
C0_baseline	50	0.4673	0.7416	0.0440	0.0209	2174	completed
C1_cache	50	0.4632	0.7423	0.0437	0.0208	2154	completed
C2_cache_amp	50	0.4638	0.7459	0.0262	0.0124	1325	completed
C3_cache_amp_gradaccum	50	0.4391	0.7043	0.0236	0.0112	1188	completed
C4_cache_amp_gradaccum_adaptivelr	50	0.4887	0.7443	0.0278	0.0132	1389	completed
C5_cache_amp_gradaccum_adaptivelr_l1	50	0.4877	0.7454	0.0301	0.0143	1497	completed
C6_full_e2am	50	0.4877	0.7454	0.0309	0.0147	1526	completed

Tiny-ImageNet

M-matrix (individual methods)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
M0_baseline_fp32	50	0.3906	0.6515	0.2132	0.1013	10300	completed
M1_cache_only	50	0.3857	0.6539	0.2160	0.1026	10145	completed
M2_amp_only	50	0.3828	0.6502	0.1101	0.0523	5223	completed
M3_grad_accum_only	50	0.3830	0.6504	0.2086	0.0991	9805	completed
M4_l1_sparsity_only	50	0.3915	0.6507	0.2192	0.1041	10569	completed
M5_adaptive_lr_only	50	0.4153	0.6548	0.2161	0.1026	10153	completed
M6_eag_only	50	0.3906	0.6515	0.2127	0.1010	10252	completed
M7_full_e2am	50	0.3962	0.6411	0.1094	0.0520	5169	completed

C-matrix (cumulative ablation)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
C0_baseline	50	0.3892	0.6518	0.2078	0.0987	9924	completed
C1_cache	50	0.3892	0.6518	0.2102	0.0999	10089	completed
C2_cache_amp	50	0.3812	0.6382	0.1080	0.0513	5196	completed
C3_cache_amp_gradaccum	50	0.3772	0.6322	0.1032	0.0490	4951	completed
C4_cache_amp_gradaccum_adaptivelr	50	0.3977	0.6331	0.1030	0.0489	4947	completed
C5_cache_amp_gradaccum_adaptivelr_l1	50	0.3919	0.6331	0.1068	0.0508	5141	completed
C6_full_e2am	50	0.3919	0.6331	0.1069	0.0508	5133	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	0.7842	106.4	1.11	902	0.0395
D1_pruned_masked	0.2654	106.4	0.13	7946	0.0058
D2_pruned_finetuned	0.7617	106.4	0.12	8250	0.0067
D3_int8_cpu_fx	0.7840	27.8	5.14	195	0.1979
D4_pruned_int8	0.2613	27.8	5.26	190	0.2041

Deployment on CIFAR-100

Variant	Accuracy	Size (MB)	Latency (ms)	Throughput (img/s)	Energy/inf (J)
D0_fp32	0.4455	106.7	0.64	1552	0.0358
D1_pruned_masked	0.0186	106.7	0.12	8542	0.0089
D2_pruned_finetuned	0.4217	106.7	0.12	8346	0.0069
D3_int8_cpu_fx	0.4348	27.8	5.16	194	0.2196
D4_pruned_int8	0.0207	27.8	5.03	199	0.2096

Deployment on Tiny-ImageNet

Variant	Accuracy	Size (MB)	Latency (ms)	Throughput (img/s)	Energy/inf (J)
D0_fp32	0.3867	107.0	1.01	985	0.0519
D1_pruned_masked	0.0572	107.0	0.41	2415	0.0238
D2_pruned_finetuned	0.3682	107.0	0.48	2094	0.0311
D3_int8_cpu_fx	0.4047	27.9	14.29	70	0.5970
D4_pruned_int8	0.0340	27.9	14.38	70	0.5926

Methodology

Model: ConvNeXtV2-Tiny (~28.0M params).

Training protocol: from scratch, AdamW (lr=4e-4, weight_decay=0.05, grad_clip=1.0), 50 epochs, warmup 2-3 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_ConvNeXtV2Tiny --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_convnextv2tiny,
  title  = {E2AM: Energy-Aware Adaptive Model Training Ablation Study (ConvNeXtV2-Tiny)},
  author = {Shanmuk},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Shanmuk4622/E2AM_ConvNeXtV2Tiny}},
}

---

This README was auto-generated on 2026-07-01 16:54 UTC. Source repo: Shanmuk4622/E2AM_ConvNeXtV2Tiny