# 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 | C1_cache                             |          0.7472 |         0.715194 |           714489 |          0.198469  |          9556    |      0.0942729 |        4356.43 |         20190298 |
| cumulative_ablation | C0_baseline                          |          0.7569 |         0.756763 |           715745 |          0.198818  |          9591.36 |      0.0944386 |        4275.37 |         20190298 |
| cumulative_ablation | C2_cache_amp                         |          0.8068 |         0.780503 |           355666 |          0.0987962 |          4823.6  |      0.0469282 |        4141.09 |         20190298 |
| cumulative_ablation | C6_full_e2am                         |          0.904  |         0.90076  |           362560 |          0.100711  |          4914.65 |      0.0478378 |        3517.29 |         20190298 |
| cumulative_ablation | C5_cache_amp_gradaccum_adaptivelr_l1 |          0.904  |         0.90076  |           362232 |          0.10062   |          4909.43 |      0.0477944 |        3517.29 |         20190298 |
| cumulative_ablation | C3_cache_amp_gradaccum               |          0.8345 |         0.82655  |           355558 |          0.0987662 |          4810.3  |      0.0469139 |        3451.18 |         20190298 |
| cumulative_ablation | C4_cache_amp_gradaccum_adaptivelr    |          0.8973 |         0.896878 |           356142 |          0.0989283 |          4825.42 |      0.0469909 |        3451.18 |         20190298 |
| individual_methods  | M7_full_e2am                         |          0.8974 |         0.897066 |           361860 |          0.100517  |          4846.89 |      0.0477454 |        3517.29 |         20190298 |
| individual_methods  | M0_baseline_fp32                     |          0.7565 |         0.710483 |           748874 |          0.20802   |          9714.07 |      0.0988097 |        3274.15 |         20190298 |
| individual_methods  | M2_amp_only                          |          0.8013 |         0.782524 |           355111 |          0.0986419 |          4774.62 |      0.0468549 |        4141.09 |         20190298 |
| individual_methods  | M3_grad_accum_only                   |          0.8036 |         0.80465  |           703651 |          0.195459  |          9402.75 |      0.0928428 |        3377.66 |         20190298 |
| individual_methods  | M6_eag_only                          |          0.759  |         0.748992 |           748945 |          0.20804   |          9713.19 |      0.0988191 |        3274.15 |         20190298 |
| individual_methods  | M1_cache_only                        |          0.7582 |         0.721166 |           706610 |          0.19628   |          9440.23 |      0.0932332 |        3279.65 |         20190298 |
| individual_methods  | M5_adaptive_lr_only                  |          0.8841 |         0.883535 |           706780 |          0.196328  |          9453.78 |      0.0932557 |        4275.37 |         20190298 |
| individual_methods  | M4_l1_sparsity_only                  |          0.7394 |         0.711569 |           755073 |          0.209743  |          9834.08 |      0.0996277 |        4436.5  |         20190298 |

## 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.