Image Classification
timm
PyTorch
facial-expression-recognition
driver-monitoring
vision-transformer
parameter-efficient-fine-tuning
lora
adaptformer
ssf
Instructions to use headless-start/parameter-efficient-dfer with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- timm
How to use headless-start/parameter-efficient-dfer with timm:
import timm model = timm.create_model("hf_hub:headless-start/parameter-efficient-dfer", pretrained=True) - Notebooks
- Google Colab
- Kaggle
| { | |
| "label": "three-candidate learning-rate selection study", | |
| "candidates": [ | |
| { | |
| "config_id": "full_ft__lr1.0e-05", | |
| "lr": 1e-05, | |
| "epoch": 28, | |
| "val_war": 0.9436127744510978, | |
| "val_uar": 0.8422065711699508, | |
| "val_loss": 0.16450884051903517, | |
| "epochs_completed": 120, | |
| "n_retained_max": 2, | |
| "checkpoint_sha256": "3094f103c17bd13558f960c22b91ed3316679cadeab88ba269d862019c4dd58a", | |
| "train_time_sec": 18801.249953985214, | |
| "optimizer_updates": 66212 | |
| }, | |
| { | |
| "config_id": "full_ft__lr3.0e-05", | |
| "lr": 3e-05, | |
| "epoch": 18, | |
| "val_war": 0.9426147704590818, | |
| "val_uar": 0.8338242351934294, | |
| "val_loss": 0.17573785160235064, | |
| "epochs_completed": 120, | |
| "n_retained_max": 1, | |
| "checkpoint_sha256": "5ae44362ba0f0d46dd0227b21ae0bd5260c7cb4f26d6d94ea01463b0160dd98c", | |
| "train_time_sec": 18755.88986349106, | |
| "optimizer_updates": 66212 | |
| }, | |
| { | |
| "config_id": "full_ft__lr1.0e-04", | |
| "lr": 0.0001, | |
| "epoch": 94, | |
| "val_war": 0.9411177644710579, | |
| "val_uar": 0.8119364341735839, | |
| "val_loss": 0.2515738859267054, | |
| "epochs_completed": 120, | |
| "n_retained_max": 2, | |
| "checkpoint_sha256": "09d0e3ab357c1f141afb2d15b379ae2c66681fe0d74b95939b04e0081ec6b350", | |
| "train_time_sec": 18814.422446727753, | |
| "optimizer_updates": 66213 | |
| } | |
| ], | |
| "decision": { | |
| "rule_id": "set_rule_war_then_uar_v1", | |
| "rule": "retain every candidate whose validation WAR is within 1e-4 of the global maximum; within that set retain every candidate whose validation UAR is within 1e-4 of the retained maximum; within that, every candidate whose validation loss is within 1e-4 of the retained minimum; break remaining ties with the declared ordered keys. Order-independent by construction: no pairwise tolerance comparison is made.", | |
| "tolerance": 0.0001, | |
| "tie_breakers": [ | |
| "lr", | |
| "config_id" | |
| ], | |
| "n_candidates": 3, | |
| "max_val_war": 0.9436127744510978, | |
| "n_retained_by_war": 1, | |
| "max_val_uar_in_retained": 0.8422065711699508, | |
| "n_retained_by_uar": 1, | |
| "min_val_loss_in_retained": 0.16450884051903517, | |
| "n_retained_by_loss": 1, | |
| "selected": { | |
| "config_id": "full_ft__lr1.0e-05", | |
| "lr": 1e-05, | |
| "epoch": 28, | |
| "val_war": 0.9436127744510978, | |
| "val_uar": 0.8422065711699508, | |
| "val_loss": 0.16450884051903517, | |
| "epochs_completed": 120, | |
| "n_retained_max": 2, | |
| "checkpoint_sha256": "3094f103c17bd13558f960c22b91ed3316679cadeab88ba269d862019c4dd58a", | |
| "train_time_sec": 18801.249953985214, | |
| "optimizer_updates": 66212 | |
| } | |
| }, | |
| "selected_run": "outputs/stage1_lr_study_2026-08-09_00-43-00/candidates/full_ft__lr1.0e-05", | |
| "accounting": { | |
| "candidates": [ | |
| 1e-05, | |
| 3e-05, | |
| 0.0001 | |
| ], | |
| "n_candidates": 3, | |
| "epochs_each": 120, | |
| "training_fits": 3, | |
| "reportable_test_evaluations": 1, | |
| "label": "three-candidate learning-rate selection study" | |
| } | |
| } |