Unified_IAA

Unified IAA (PARA pt=70 + LAPIS pt=137)

This release includes:

• PARA PIAA-MIR (pt=70)
  • best_model_vit_small_patch16_224_piaamir_super-yogurt-742.pth
  • best_model_swin_tiny_patch4_window7_224_piaamir_fanciful-blaze-742.pth
• PARA PIAA-ICI (pt=70)
  • best_model_swin_tiny_patch4_window7_224_piaaici_ethereal-cherry-741.pth
  • best_model_vit_small_patch16_224_piaaici_laced-bird-742.pth
• LAPIS PIAA-MIR (pt=137)
  • lapis_best_model_resnet50_piaamir_azure-gorge-1153.pth
• LAPIS PIAA-ICI (pt=137)
  • lapis_best_model_resnet50_piaaici_dutiful-serenity-1076.pth
• PARA GIAA prior vector (prior_mean_vector.pt, shape [70])

Scope and compatibility

• ✅ PARA supported: num_pt=70, disable_onehot=false
• ✅ LAPIS supported: num_pt=137 (use direct trait vector input)
• ✅ Supports both:
  • PARA GIAA prior inference (using uploaded prior_mean_vector.pt)
  • PARA PIAA personalized inference (user provides demographics + Big5)
  • LAPIS inference (user provides a trait vector)

See configs/compatibility.json for exact artifact mapping and hashes.

Files

• models/best_model_vit_small_patch16_224_piaamir_super-yogurt-742.pth
• models/best_model_swin_tiny_patch4_window7_224_piaamir_fanciful-blaze-742.pth
• models/best_model_swin_tiny_patch4_window7_224_piaaici_ethereal-cherry-741.pth
• models/best_model_vit_small_patch16_224_piaaici_laced-bird-742.pth
• models/lapis_best_model_resnet50_piaamir_azure-gorge-1153.pth
• models/lapis_best_model_resnet50_piaaici_dutiful-serenity-1076.pth
• inference/prior_mean_vector.pt
• inference/demographics_encoder.py
• inference/predict_piaa.py
• inference/prior_giaa.py
• infer_unified_iaa.sh
• run_LAPIS_PIAA.sh

Python SDK (pip style)

from unified_iaa import UnifiedIAA

m = UnifiedIAA.from_pretrained("stupidog04/Unified_IAA")

score_piaa = m.predict_piaa(
    image="/path/to/test.jpg",
    demographics={
        "age": "20-29",
        "gender": "female",
        "EducationalLevel": "Bachelor",
        "artExperience": "medium",
        "photographyExperience": "low",
    },
    big5={
        "personality-E": 6,
        "personality-A": 7,
        "personality-N": 4,
        "personality-O": 8,
        "personality-C": 6,
    },
    task="mir",
    backbone="vit_small_patch16_224",
)

score_giaa = m.predict_giaa_prior(
    image="/path/to/test.jpg",
    task="ici",
    backbone="swin_tiny_patch4_window7_224",
)

# LAPIS (pt=137) with direct trait vector
lapis_traits = [0.0] * 137
score_lapis = m.predict_with_traits(
    image="/path/to/test.jpg",
    traits=lapis_traits,
    task="mir",
    backbone="resnet50",
    dataset="lapis",
)

Quick usage

1) Build demographics encoder

python inference/demographics_encoder.py \
  --userinfo_csv /mnt/d/datasets/PARA/annotation/PARA-UserInfo.csv \
  --out_json inference/demographics_encoder.json

2) Personalized PIAA inference (user demographics input)

python inference/predict_piaa.py \
  --task mir \
  --backbone vit_small_patch16_224 \
  --checkpoint models/best_model_vit_small_patch16_224_piaamir_super-yogurt-742.pth \
  --image /path/to/image.jpg \
  --encoder_json inference/demographics_encoder.json \
  --demographics_json /path/to/user_demo.json

3) GIAA prior inference (uploaded prior vector)

python inference/prior_giaa.py \
  --task ici \
  --backbone swin_tiny_patch4_window7_224 \
  --checkpoint models/best_model_swin_tiny_patch4_window7_224_piaaici_ethereal-cherry-741.pth \
  --image /path/to/image.jpg \
  --prior_vector_path inference/prior_mean_vector.pt

4) Run LAPIS PIAA eval pipeline

bash run_LAPIS_PIAA.sh

user_demo.json format

{
  "age": "20-29",
  "gender": "female",
  "EducationalLevel": "Bachelor",
  "artExperience": "medium",
  "photographyExperience": "low",
  "personality-E": 6,
  "personality-A": 7,
  "personality-N": 4,
  "personality-O": 8,
  "personality-C": 6
}

Category strings must match encoder categories produced from your PARA PARA-UserInfo.csv.