headless-start/parameter-efficient-dfer

Parameter-Efficient Adaptation of Facial Expression Recognition Models for Driver Monitoring

Trained weights for the MSc thesis by Ayush Tiwari, BTU Cottbus-Senftenberg, 2026.

Code and training instructions · Weight files · Author profile

The code repository is currently a private preview; its link becomes accessible to everyone when the code is published. The weights here are public.

A public ImageNet-pretrained ViT-B/16 is fine-tuned on six-class FER+ (Stage 1), then its encoder is adapted to KMU-FED with a fresh classifier (Stage 2). Five strategies are provided: full fine-tuning, linear probe, AdaptFormer, LoRA and SSF.

Contents

There are 51 checkpoints: one selected FER+ model and ten held-out-fold models for each selected KMU-FED strategy. Per-fold results, predictions and search-selection summaries accompany the weights. No image data is included.

Path	Contents
stage1/best.ckpt	Complete FER+-adapted model
stage2/<strategy>/fold_00/ … fold_09/	Fold weights, results and predictions
stage2/<strategy>/selection.json	Search scores and selected configuration
load_weights.py	Loader for the accompanying code release

Full fine-tuning stores complete networks. The other Stage-2 strategies store trained parameters and their classifier; the loader combines them with the Stage-1 encoder. There is no separately trained all-data deployment model.

Results

FER+ six-class test: 93.21% accuracy, 81.40% UAR, 93.13% weighted F1. The learning rate (1e-5) and epoch (28) were selected on the official validation split; the selected model was tested once.

KMU-FED values below are fold means; accuracy ± is sample standard deviation across ten folds. The thesis and code README instead report pooled weighted F1, which differs from averaging fold F1 scores.

Strategy	Selected setting	Trainable parameters	Accuracy (%)	UAR (%)	Fold-mean weighted F1 (%)
Full fine-tuning	lr 3e-5	85,803,270	97.55 ± 4.11	96.87	97.25
Linear probe	lr 3e-3	4,614	93.27 ± 6.83	92.76	92.42
AdaptFormer	lr 1e-3, width 64 (reduction 12)	1,194,246	98.45 ± 2.43	98.21	98.40
LoRA	lr 1e-3, rank 4	152,070	99.00 ± 2.59	98.56	98.83
SSF	lr 3e-3	210,438	97.91 ± 2.88	97.48	97.75

Both the retained epoch and configuration were selected on the reported folds, so these are optimistic, best-observed estimates. Folds share drivers and sequences; they measure familiar-driver performance. Fold spread is descriptive, with one base seed (42) and deterministic fold-specific initialisation. The study does not establish equivalence, condition-specific robustness, the benefit of the intermediate FER+ stage, or inference speed.

Use the weights

Install the accompanying code release's requirements and run from its root directory. The loader needs its methods.py, data.py and experiment.py; the checkpoint files alone are not a Transformers pipeline.

Download the helper at the verified weight revision:

python - <<'PY'
from huggingface_hub import hf_hub_download
hf_hub_download(
    "headless-start/parameter-efficient-dfer", "load_weights.py",
    revision="92a70a21dc4810e346d7daf7b4e3c04b7dc721bb",
    local_dir="runs/released",
)
PY

import sys
import torch
from PIL import Image
from data import build_transform
from experiment import CLASS_NAMES

sys.path.insert(0, "runs/released")
from load_weights import load_stage2

model = load_stage2("lora", fold=0)
with Image.open("face_crop.jpg") as image:
    x = build_transform(False, kmufed=True)(image.convert("RGB"))[None]
with torch.inference_mode():
    predicted = model(x).argmax(1).item()
print(CLASS_NAMES[predicted])

The loader downloads the required weights and returns an evaluation-mode model. Use load_stage1() for FER+ or load_stage2(method, fold) for KMU-FED; methods are full_ft, linear_probe, adaptformer, lora, ssf, and folds are zero-based (0–9). Class order is angry, disgust, fear, happy, sad, surprise. Inputs must be face crops; preprocessing makes three-channel grayscale, resizes to 224 × 224 with bicubic interpolation and normalises with mean/std 0.5.

These original-format weights are for inference through the helper, not --resume inputs for the standalone training scripts. Hosted result JSONs also use the original format; the code release's plotter reads its own training outputs. The supplied loader retains adapter modules rather than merging them.

Licence and citation

Weights: CC BY-NC 4.0. The code release is MIT-licensed. Dataset and pretrained-weight terms remain with their providers. These are research models evaluated on a small driver corpus, without validation for operational use.

@mastersthesis{tiwari2026peft_dfer,
  author = {Ayush Tiwari},
  title = {Parameter-Efficient Adaptation of Facial Expression Recognition Models for Driver Monitoring},
  school = {Brandenburgische Technische Universit{\"a}t Cottbus-Senftenberg},
  year = {2026}
}