alecccdd/ratebv-frwo-r1-model

Continuous image scorer: maps an image to a score in [0, 1000]. Backbone: vendored Sapiens2 ViT (sapiens2_0.8b) (frozen) with a LoRA adapter applied at load time, plus a small regression head. Exported from training run jvcead4r.

Usage

from transformers import AutoModel

model = AutoModel.from_pretrained("alecccdd/ratebv-frwo-r1-model", trust_remote_code=True).eval()
model.predict("photo.jpg")            # -> float in [0, 1000]
model.predict(["a.jpg", "b.jpg"])     # -> list[float]

The base backbone facebook/sapiens2-pretrain-0.8b is gated on the HuggingFace Hub. Before loading, accept its license and make sure HF_TOKEN is set — only the LoRA adapter and head are stored here; the base weights are downloaded at load time.

from huggingface_hub import login
login("hf_...")   # a token with the base model's license accepted

What's inside

File	Purpose
model.py / sapiens2.py	self-contained model code (loaded via trust_remote_code)
config.json	ScorerConfig (auto_map -> ScorerConfig / ScorerModel)
adapter/adapter_config.json, adapter/adapter_model.safetensors	the LoRA adapter (NOT merged)
head.pt	the regression head weights

How it works

1. Load the frozen backbone facebook/sapiens2-pretrain-0.8b and apply the LoRA adapter (peft).
2. Pool the CLS embedding and run the regression head -> a sigmoid in [0, 1].
3. Denormalize: score = sigmoid * (1000 - 0) + 0.

Preprocessing: preserve_aspect=True, image_size=[296, 222], patch_size=16, image_mean=[0.485, 0.456, 0.406], image_std=[0.229, 0.224, 0.225]. Head: hidden_size=1280, hidden_dims=[] (minimal head), dropout=0.45.

Provenance

Source run jvcead4r — val_mae=63.50330494869472, val_r2=0.9078422463529856.

Requirements

torch, torchvision, transformers>=5.9, peft>=0.19, huggingface_hub, safetensors, numpy, pillow.