File size: 4,922 Bytes
e7be770 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 | # 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 | C5_cache_amp_gradaccum_adaptivelr_l1 | 0.5708 | 0.565335 | 739217 | 0.205338 | 9943.1 | 0.0975356 | 3518.37 | 20433688 |
| cumulative_ablation | C3_cache_amp_gradaccum | 0.4436 | 0.439476 | 724322 | 0.201201 | 9765.77 | 0.0955703 | 3453.76 | 20433688 |
| cumulative_ablation | C2_cache_amp | 0.3465 | 0.318729 | 712632 | 0.197953 | 9632.56 | 0.0940279 | 4147 | 20433688 |
| cumulative_ablation | C1_cache | 0.2502 | 0.19491 | 1.44514e+06 | 0.401428 | 19311.6 | 0.190678 | 4358.76 | 20433688 |
| cumulative_ablation | C0_baseline | 0.2475 | 0.211268 | 1.48291e+06 | 0.411919 | 19800.5 | 0.195662 | 4358.76 | 20433688 |
| cumulative_ablation | C6_full_e2am | 0.5708 | 0.565335 | 738944 | 0.205262 | 9939.1 | 0.0974996 | 3518.37 | 20433688 |
| cumulative_ablation | C4_cache_amp_gradaccum_adaptivelr | 0.5738 | 0.568648 | 729533 | 0.202648 | 9831.06 | 0.0962578 | 4228.98 | 20433688 |
| individual_methods | M0_baseline_fp32 | 0.2462 | 0.223854 | 1.42266e+06 | 0.395184 | 18948.9 | 0.187712 | 4358.76 | 20433688 |
| individual_methods | M4_l1_sparsity_only | 0.244 | 0.210349 | 1.45956e+06 | 0.405434 | 19502.2 | 0.192581 | 4440.33 | 20433688 |
| individual_methods | M3_grad_accum_only | 0.3467 | 0.324298 | 1.41165e+06 | 0.392124 | 18791 | 0.186259 | 3383.49 | 20433688 |
| individual_methods | M6_eag_only | 0.2606 | 0.209208 | 1.4373e+06 | 0.39925 | 19318.7 | 0.189644 | 3285.48 | 20433688 |
| individual_methods | M7_full_e2am | 0.5805 | 0.575147 | 721653 | 0.200459 | 9682.67 | 0.0952181 | 3519.12 | 20433688 |
| individual_methods | M1_cache_only | 0.2451 | 0.182945 | 1.4199e+06 | 0.394418 | 19007.6 | 0.187348 | 4358.76 | 20433688 |
| individual_methods | M5_adaptive_lr_only | 0.4611 | 0.449641 | 1.4325e+06 | 0.397918 | 19208.2 | 0.189011 | 4358.76 | 20433688 |
| individual_methods | M2_amp_only | 0.3541 | 0.294041 | 697223 | 0.193673 | 9387.33 | 0.0919947 | 4229.98 | 20433688 |
## 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.
|