Hugging Face's logo

Datasets:
Shanmuk4622
/
E2AM_ResNet50

Modalities:
Image
Tabular
Text
Formats:
csv
Size:
< 1K
Libraries:
Datasets
pandas
Polars
Dataset card Data Studio Files
xet
 Community
1
E2AM_ResNet50 / MODEL_CARD.md
Shanmuk4622's picture
Shanmuk4622
e2am: batch upload 12 files (453 KB)
7df6baa verified
preview code
|
raw
history blame contribute delete
11.1 kB

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 0.8452 0.805897 630655 0.175182 8494.63 0.0832114 3523.98 23520842
cumulative_ablation C3_cache_amp_gradaccum 0.87 0.867989 256776 0.0713267 3478.64 0.0338802 3496.36 23520842
cumulative_ablation C0_baseline 0.8442 0.83857 628916 0.174699 8464.02 0.082982 8918.25 23520842
cumulative_ablation C2_cache_amp 0.8595 0.858567 238636 0.0662878 3188.63 0.0314867 4304.1 23520842
cumulative_ablation C6_full_e2am 0.9151 0.914313 258009 0.0716693 3503.01 0.0340429 3588.3 23520842
cumulative_ablation C5_cache_amp_gradaccum_adaptivelr_l1 0.9151 0.914313 258527 0.0718131 3504.49 0.0341112 3588.3 23520842
cumulative_ablation C4_cache_amp_gradaccum_adaptivelr 0.9084 0.908445 256027 0.0711185 3475.86 0.0337813 3496.36 23520842
individual_methods M7_full_e2am 0.9088 0.907069 238963 0.0663786 3191.77 0.0315298 4589.37 23520842
individual_methods M3_grad_accum_only 0.872 0.861635 558641 0.155178 7331.69 0.0737095 3754 23520842
individual_methods M4_l1_sparsity_only 0.8387 0.822116 568184 0.157829 7447.29 0.0749688 3752.38 23520842
individual_methods M6_eag_only 0.8452 0.842105 579947 0.161096 7653.85 0.0765208 3738.38 23520842
individual_methods M1_cache_only 0.8543 0.787845 561934 0.156093 7368.66 0.0741441 3716.18 23520842
individual_methods M0_baseline_fp32 0.8298 0.778633 572470 0.159019 7588.38 0.0755342 8918.25 23520842
individual_methods M5_adaptive_lr_only 0.9397 0.938466 560302 0.155639 7368.43 0.0739287 3716.18 23520842
cumulative_ablation C1_cache 0.4213 0.410099 4.74386e+06 1.31774 62533.1 0.625926 12956.8 23910152
cumulative_ablation C3_cache_amp_gradaccum 0.5505 0.549605 1.81747e+06 0.504853 23603.6 0.239805 10386.8 23910152
cumulative_ablation C0_baseline 0.4121 0.345293 4.95892e+06 1.37748 65154.2 0.654302 0 23910152
cumulative_ablation C2_cache_amp 0.4993 0.454522 1.77443e+06 0.492898 22910.4 0.234127 10294.1 23910152
cumulative_ablation C6_full_e2am 0.6716 0.669722 1.8588e+06 0.516333 24352 0.245258 10482.9 23910152
cumulative_ablation C5_cache_amp_gradaccum_adaptivelr_l1 0.6742 0.673791 1.88316e+06 0.523099 24653 0.248472 10388.5 23910152
cumulative_ablation C4_cache_amp_gradaccum_adaptivelr 0.6667 0.664645 1.95748e+06 0.543745 25771.2 0.258279 9793.94 23910152
individual_methods M7_full_e2am 0.6641 0.66312 1.95222e+06 0.542284 25680.6 0.257585 10388.5 23910152
individual_methods M3_grad_accum_only 0.5141 0.488954 4.70098e+06 1.30583 61979.2 0.620268 12956.8 23910152
individual_methods M4_l1_sparsity_only 0.4112 0.397086 4.45606e+06 1.2378 58649.5 0.587953 12394.6 23910152
individual_methods M2_amp_only 0.4889 0.479646 1.846e+06 0.512777 24267.8 0.243569 10386.8 23910152
individual_methods M6_eag_only 0.3996 0.367363 4.53777e+06 1.26049 59534.4 0.598733 12956.8 23910152
individual_methods M1_cache_only 0.4027 0.374706 4.35167e+06 1.2088 57540.7 0.574178 12956.8 23910152
individual_methods M0_baseline_fp32 0.4187 0.381184 4.61944e+06 1.28318 60871.2 0.60951 9643.07 23910152
individual_methods M5_adaptive_lr_only 0.6243 0.624695 4.29027e+06 1.19174 56538.7 0.566077 12956.8 23910152
cumulative_ablation C1_cache 0.5916 0.571995 583463 0.162073 7754.37 0.0769847 9017.14 23705252
cumulative_ablation C3_cache_amp_gradaccum 0.6306 0.628047 237388 0.065941 3187.79 0.031322 3493.51 23705252
cumulative_ablation C0_baseline 0.5822 0.577501 596644 0.165734 7918.13 0.0787238 8918.83 23705252
cumulative_ablation C2_cache_amp 0.6307 0.606307 241142 0.066984 3233.03 0.0318174 4307.43 23705252
cumulative_ablation C6_full_e2am 0.6753 0.675139 241308 0.06703 3224.31 0.0318392 3584.96 23705252
cumulative_ablation C5_cache_amp_gradaccum_adaptivelr_l1 0.6753 0.675139 240012 0.06667 3225.3 0.0316683 3584.96 23705252
cumulative_ablation C4_cache_amp_gradaccum_adaptivelr 0.6918 0.69041 237386 0.0659404 3190.37 0.0313217 3493.51 23705252
individual_methods M7_full_e2am 0.7008 0.700132 233116 0.0647546 3122.94 0.0307584 3590.46 23705252
individual_methods M3_grad_accum_only 0.6228 0.616555 561403 0.155945 7380.5 0.074074 3459.52 23705252
individual_methods M4_l1_sparsity_only 0.5897 0.576023 569303 0.15814 7475.23 0.0751163 3460.33 23705252
individual_methods M2_amp_only 0.6081 0.595609 252820 0.0702279 3377.25 0.0333582 4307.43 23705252
individual_methods M6_eag_only 0.5821 0.543197 563427 0.156508 7411.66 0.0743411 3368.14 23705252
individual_methods M1_cache_only 0.5947 0.582235 564048 0.15668 7414.79 0.074423 8919.7 23705252
individual_methods M0_baseline_fp32 0.5913 0.555869 564906 0.156918 7413.58 0.0745363 3368.14 23705252
individual_methods M5_adaptive_lr_only 0.7482 0.748151 571453 0.158737 7577.76 0.0754001 8920.45 23705252

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.