MODEL_CARD.md

E2AM-ResNet50 — Model Card

Energy-aware ResNet-50 from scratch on CIFAR-10.

Architecture

• ResNet-50 (torchvision.models.resnet50(weights=None))
• CIFAR adaptation: conv1 -> 3x3 stride-1, maxpool -> Identity
• Input: 32x32 RGB, CIFAR-10 normalization stats
• Trained from scratch (no pretrained weights, no transfer)

Training environment

• Platform: Kaggle Dual-T4 notebook (single T4 used for training)
• Energy: nvidia-smi power.draw at 1 Hz, trapezoidal integration
• Carbon: CO2_kg = (E_J / 3.6e6) * intensity

Results

method_group	variant_name	best_accuracy	final_f1_score	total_energy_j	total_energy_kwh	total_time_sec	total_co2_kg	peak_vram_mb	num_parameters
cumulative_ablation	C4_cache_amp_gradaccum_adaptivelr	0.3977	0.390035	370791	0.102998	4947.09	0.0489239	1662.97	28020296
cumulative_ablation	C0_baseline	0.3892	0.370385	748098	0.207805	9924.26	0.0987074	2867.09	28020296
cumulative_ablation	C3_cache_amp_gradaccum	0.3772	0.355029	371606	0.103224	4950.59	0.0490313	1662.97	28020296
cumulative_ablation	C1_cache	0.3892	0.365102	756898	0.210249	10089	0.0998685	2861.06	28020296
cumulative_ablation	C2_cache_amp	0.3812	0.354907	388745	0.107985	5196.41	0.0512927	1545.73	28020296
cumulative_ablation	C5_cache_amp_gradaccum_adaptivelr_l1	0.3919	0.385792	384648	0.106847	5140.54	0.0507521	1756.89	28020296
cumulative_ablation	C6_full_e2am	0.3919	0.385792	384772	0.106881	5132.64	0.0507685	1756.89	28020296
individual_methods	M1_cache_only	0.3857	0.360327	777541	0.215984	10144.6	0.102592	1257.31	28020296
individual_methods	M7_full_e2am	0.3962	0.38927	393985	0.10944	5169.17	0.0519841	1756.89	28020296
individual_methods	M3_grad_accum_only	0.383	0.360399	750843	0.208568	9804.91	0.0990696	1339.67	28020296
individual_methods	M6_eag_only	0.3906	0.357221	765768	0.212713	10251.6	0.101039	1256.06	28020296
individual_methods	M5_adaptive_lr_only	0.4153	0.410143	777880	0.216078	10152.6	0.102637	2867.09	28020296
individual_methods	M0_baseline_fp32	0.3906	0.360285	767680	0.213245	10299.6	0.101291	1256.06	28020296
individual_methods	M2_amp_only	0.3828	0.36047	396535	0.110149	5223.2	0.0523206	1545.73	28020296
individual_methods	M4_l1_sparsity_only	0.3915	0.3657	789122	0.219201	10569	0.10412	2862.59	28020296

Limitations

• Energy is integrated from per-GPU nvidia-smi power samples; CPU energy is not measured. Numbers reflect GPU-only training energy.
• Pruning (D1/D2) is mask-based structured pruning — weights are zeroed but the dense layout is preserved. Real wall-clock speed-ups need a sparsity- aware runtime; reported numbers reflect masked weights.
• INT8 (D3/D4) uses CPU FX static quantization (fbgemm); GPU INT8 via TensorRT is out of scope.
• ResNet-50 with the small-image stem (conv1 3x3 stride-1, maxpool=Identity). Used for CIFAR-10 (32x32) and Tiny-ImageNet-200 (64x64). For ImageNet (224x224) the standard 7x7 stride-2 conv1 + maxpool would be needed instead.
• Single T4 by design — DataParallel was measured slower at this scale.