app.py

import gradio as gr
import torch
from PIL import Image
from torchvision import transforms
from model_loader import load_model
from index_to_attr import index_to_attr

# Modell laden
model = load_model("model/AttrPredModel_StateDict.pth")

# taskName pro Index extrahieren
def get_task_map(index_to_attr):
    task_map = {}
    for idx, desc in index_to_attr.items():
        if "(" in desc and ")" in desc:
            task = desc.split("(")[-1].split(")")[0]
            task_map[idx] = task
    return task_map

task_map = get_task_map(index_to_attr)

# Bildverarbeitungspipeline
preprocess = transforms.Compose([
    transforms.Resize((512, 512)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.6765, 0.6347, 0.6207],
                         std=[0.3284, 0.3371, 0.3379])
])

# Inferenz-Funktion mit Markierung für unsichere Kategorien
def predict(image):
    image = image.convert("RGB")
    input_tensor = preprocess(image).unsqueeze(0)
    with torch.no_grad():
        output = model(input_tensor)
        probs = torch.sigmoid(output).squeeze().numpy()

    result = {}
    threshold = 0.5
    top_per_task = {}

    for idx, score in enumerate(probs):
        task = task_map.get(idx, "unknown")
        if task not in top_per_task or score > top_per_task[task][1]:
            top_per_task[task] = (idx, score)

    for task, (idx, score) in top_per_task.items():
        label = index_to_attr.get(idx, f"Unknown ({idx})").split(" (")[0]
        result[task] = {
            "label": label,
            "score": round(float(score), 4),
            "confidence": "low" if score < threshold else "high"
        }

    return result

# Gradio Interface – stabil und einfach
demo = gr.Interface(
    fn=predict,
    inputs=gr.Image(type="pil", label="Upload image"),
    outputs="json",
    title="Fashion Attribute Predictor (mit Confidence)",
    description="Zeigt pro Attributgruppe die wahrscheinlichste Vorhersage + Confidence ('high' / 'low')."
)

if __name__ == "__main__":
    demo.launch()