# E2AM-ResNet50

## Model Details
- Architecture: ResNet-50
- Initialization: from scratch (`weights=None`)
- Dataset source: Kaggle ImageFolder or Hugging Face dataset; no torchvision-hosted dataset download is used.
- Energy: NVIDIA-SMI GPU power sampling summed across visible GPUs.

## Training Results
| run_id                 | variant_name   |   best_f1_score |   final_accuracy |   final_f1_score |   total_energy_j |   total_time_sec |   total_co2_kg |   num_parameters |
|:-----------------------|:---------------|----------------:|-----------------:|-----------------:|-----------------:|-----------------:|---------------:|-----------------:|
| cifar10/M2_amp_only    | M2_amp_only    |        0.666577 |           0.673  |         0.666577 | 477292           |          3803.68 |      0.062976  |         23528522 |
| cifar10/M7_full_e2am   | M7_full_e2am   |        0.683629 |           0.6851 |         0.683629 | 471771           |          3835.59 |      0.0622476 |         23528522 |
| cifar10/baseline_fixed | baseline_fixed |        0.661276 |           0.6546 |         0.661276 |      1.02776e+06 |          7988.1  |      0.135608  |         23528522 |

## Deployment Results
| variant                 |   accuracy |   f1_score |   latency_ms_per_image |   throughput_images_per_sec |   energy_per_inference_j |   sparsity_percent |
|:------------------------|-----------:|-----------:|-----------------------:|----------------------------:|-------------------------:|-------------------:|
| fp32                    |   0.689063 |  0.687344  |                4.46075 |                    224.178  |                 0.435256 |             0      |
| pruned                  |   0.103906 |  0.0257757 |                3.33544 |                    299.81   |                 0.323063 |            19.9062 |
| pruned_finetuned        |   0.665365 |  0.664818  |                3.18521 |                    313.951  |                 0.301914 |            10.7355 |
| dynamic_linear_int8_cpu |   0.686198 |  0.684309  |              225.033   |                      4.4438 |               nan        |             0      |

## Limitations
- Dynamic depth routing is not claimed in this ResNet-50 experiment.
- CPU INT8 energy is not GPU energy and is reported as unavailable.