Shanmuk4622/E2AM_ResNet50

E2AM Ablation Results: ResNet-50

Energy-aware training ablation study for ResNet-50 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	M5_adaptive_lr_only	0.9397	—	0.1556	0.0739	7368
CIFAR-100	M5_adaptive_lr_only	0.7482	0.9358	0.1587	0.0754	7578
Tiny-ImageNet	C5_cache_amp_gradaccum_adaptivelr_l1	0.6742	0.8651	0.5231	0.2485	24653

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	0.8298	—	0.1590	0.0755	7588	completed
M1_cache_only	50	0.8543	—	0.1561	0.0741	7369	completed
M2_amp_only	50	0.8744	—	0.0622	0.0295	3006	completed
M3_grad_accum_only	50	0.8720	—	0.1552	0.0737	7332	completed
M4_l1_sparsity_only	50	0.8387	—	0.1578	0.0750	7447	completed
M5_adaptive_lr_only	50	0.9397	—	0.1556	0.0739	7368	completed
M6_eag_only	50	0.8452	—	0.1611	0.0765	7654	completed
M7_full_e2am	50	0.9088	—	0.0664	0.0315	3192	completed

C-matrix (cumulative ablation)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
C0_baseline	50	0.8442	—	0.1747	0.0830	8464	completed
C1_cache	50	0.8452	—	0.1752	0.0832	8495	completed
C2_cache_amp	50	0.8595	—	0.0663	0.0315	3189	completed
C3_cache_amp_gradaccum	50	0.8700	—	0.0713	0.0339	3479	completed
C4_cache_amp_gradaccum_adaptivelr	50	0.9084	—	0.0711	0.0338	3476	completed
C5_cache_amp_gradaccum_adaptivelr_l1	50	0.9151	—	0.0718	0.0341	3504	completed
C6_full_e2am	50	0.9151	—	0.0717	0.0340	3503	completed

CIFAR-100

M-matrix (individual methods)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
M0_baseline_fp32	50	0.5913	0.8446	0.1569	0.0745	7414	completed
M1_cache_only	50	0.5947	0.8582	0.1567	0.0744	7415	completed
M2_amp_only	50	0.6081	0.8652	0.0702	0.0334	3377	completed
M3_grad_accum_only	50	0.6228	0.8766	0.1559	0.0741	7380	completed
M4_l1_sparsity_only	50	0.5897	0.8576	0.1581	0.0751	7475	completed
M5_adaptive_lr_only	50	0.7482	0.9358	0.1587	0.0754	7578	completed
M6_eag_only	50	0.5821	0.8440	0.1565	0.0743	7412	completed
M7_full_e2am	50	0.7008	0.9087	0.0648	0.0308	3123	completed

C-matrix (cumulative ablation)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
C0_baseline	50	0.5822	0.8628	0.1657	0.0787	7918	completed
C1_cache	50	0.5916	0.8505	0.1621	0.0770	7754	completed
C2_cache_amp	50	0.6307	0.8665	0.0670	0.0318	3233	completed
C3_cache_amp_gradaccum	50	0.6306	0.8761	0.0659	0.0313	3188	completed
C4_cache_amp_gradaccum_adaptivelr	50	0.6918	0.9042	0.0659	0.0313	3190	completed
C5_cache_amp_gradaccum_adaptivelr_l1	50	0.6753	0.8974	0.0667	0.0317	3225	completed
C6_full_e2am	50	0.6753	0.8974	0.0670	0.0318	3224	completed

Tiny-ImageNet

M-matrix (individual methods)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
M0_baseline_fp32	50	0.4187	0.6629	1.2832	0.6095	60871	completed
M1_cache_only	50	0.4027	0.6637	1.2088	0.5742	57541	completed
M2_amp_only	50	0.4889	0.7445	0.5128	0.2436	24268	completed
M3_grad_accum_only	50	0.5141	0.7502	1.3058	0.6203	61979	completed
M4_l1_sparsity_only	50	0.4112	0.6874	1.2378	0.5880	58649	completed
M5_adaptive_lr_only	50	0.6243	0.8426	1.1917	0.5661	56539	completed
M6_eag_only	50	0.3996	0.6543	1.2605	0.5987	59534	completed
M7_full_e2am	50	0.6641	0.8580	0.5423	0.2576	25681	completed

C-matrix (cumulative ablation)

Variant	Epochs	Top-1	Top-5	Energy (kWh)	CO₂ (kg)	Time (s)	Status
C0_baseline	50	0.4121	0.6262	1.3775	0.6543	65154	completed
C1_cache	50	0.4213	0.7007	1.3177	0.6259	62533	completed
C2_cache_amp	50	0.4993	0.7211	0.4929	0.2341	22910	completed
C3_cache_amp_gradaccum	50	0.5505	0.7980	0.5049	0.2398	23604	completed
C4_cache_amp_gradaccum_adaptivelr	50	0.6667	0.8607	0.5437	0.2583	25771	completed
C5_cache_amp_gradaccum_adaptivelr_l1	50	0.6742	0.8651	0.5231	0.2485	24653	completed
C6_full_e2am	49	0.6716	0.8621	0.5163	0.2453	24352	completed

Deployment results

See paper_tables/deployment_results_table.csv.

Methodology

Model: ResNet-50 (~23.5M 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 32x32 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_ResNet50 --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_resnet50,
  title  = {E2AM: Energy-Aware Adaptive Model Training Ablation Study (ResNet-50)},
  author = {Shanmuk},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Shanmuk4622/E2AM_ResNet50}},
}

---

This README was auto-generated on 2026-07-02 08:14 UTC. Source repo: Shanmuk4622/E2AM_ResNet50