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 | 0.5824 | 0.580189 | 1.80662e+06 | 0.501838 | 23976.1 | 0.238373 | 6401.34 | 4491441 |
| cumulative_ablation | C5_cache_amp_gradaccum_adaptivelr_l1 | 0.5857 | 0.582411 | 1.8737e+06 | 0.520471 | 24837.7 | 0.247224 | 6401.34 | 4491441 |
| cumulative_ablation | C0_baseline | 0.1374 | 0.0681395 | 3.69144e+06 | 1.0254 | 48707.8 | 0.487065 | 14573.8 | 4491441 |
| cumulative_ablation | C2_cache_amp | 0.2375 | 0.202107 | 1.77564e+06 | 0.493233 | 23429.2 | 0.234286 | 6403.21 | 4491441 |
| cumulative_ablation | C1_cache | 0.1351 | 0.0736167 | 3.43645e+06 | 0.95457 | 45416.2 | 0.453421 | 14573.8 | 4491441 |
| cumulative_ablation | C3_cache_amp_gradaccum | 0.3793 | 0.3545 | 1.77893e+06 | 0.494147 | 23520 | 0.23472 | 6401.34 | 4491441 |
| cumulative_ablation | C6_full_e2am | 0.5761 | 0.573284 | 1.86649e+06 | 0.518469 | 24824.4 | 0.246273 | 6401.34 | 4491441 |
| individual_methods | M0_baseline_fp32 | 0.1359 | 0.0880787 | 3.59713e+06 | 0.999202 | 47630.4 | 0.474621 | 14573.8 | 4491441 |
| individual_methods | M4_l1_sparsity_only | 0.1391 | 0.0756299 | 3.75408e+06 | 1.0428 | 49937.9 | 0.49533 | 14576.1 | 4491441 |
| individual_methods | M2_amp_only | 0.2424 | 0.182652 | 1.88899e+06 | 0.524719 | 25279 | 0.249242 | 5430.87 | 4491441 |
| individual_methods | M1_cache_only | 0.1416 | 0.0721437 | 3.57791e+06 | 0.993863 | 46971.9 | 0.472085 | 14573.8 | 4491441 |
| individual_methods | M7_full_e2am | 0.5835 | 0.581187 | 1.89605e+06 | 0.526681 | 25136.7 | 0.250174 | 6401.34 | 4491441 |
| individual_methods | M3_grad_accum_only | 0.264 | 0.224093 | 3.47502e+06 | 0.965282 | 45545 | 0.458509 | 14573.8 | 4491441 |
| individual_methods | M5_adaptive_lr_only | 0.3614 | 0.344721 | 3.4942e+06 | 0.970612 | 45838.2 | 0.461041 | 14573.8 | 4491441 |
| individual_methods | M6_eag_only | 0.1279 | 0.0810579 | 3.51809e+06 | 0.977246 | 46599.6 | 0.464192 | 14573.8 | 4491441 |
Limitations
- Energy is integrated from per-GPU
nvidia-smipower 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.