app.py

import io
import pickle

import numpy as np
import torch
from fastapi import FastAPI, UploadFile, File, Form
from fastapi.middleware.cors import CORSMiddleware
from PIL import Image
from transformers import AutoTokenizer, AutoModel
import open_clip

device = "cuda" if torch.cuda.is_available() else "cpu"

TEXT_MODEL_NAME = "indobenchmark/indobert-large-p1"
tokenizer = AutoTokenizer.from_pretrained(TEXT_MODEL_NAME)
text_model = AutoModel.from_pretrained(TEXT_MODEL_NAME).to(device)
text_model.eval()

clip_model, _, clip_preprocess = open_clip.create_model_and_transforms(
    "EVA01-g-14-plus",
    pretrained="merged2b_s11b_b114k"
)
clip_model.to(device)
clip_model.eval()

with open("xgb_full.pkl", "rb") as f:
    xgb_model = pickle.load(f)

def preprocess_text(text: str) -> str:
    # nanti ditambahin preprocessingnya
    return text.strip()

app = FastAPI(
    title="Multimodal Water Pollution Risk API",
    description=(
        "Input: text + image + geospatial + time\n"
        "Model: IndoBERT + EVA-CLIP (HF Hub) + XGBoost (xgb.pkl)\n"
    ),
    version="1.0.0",
)

app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_methods=["*"],
    allow_headers=["*"],
)

@app.get("/")
def root():
    return {
        "status": "OK",
        "message": "Multimodal Water Pollution Risk API is running.",
        "info": "Use POST /predict with text, image, and features.",
    }

@app.post("/predict")
async def predict(
    text: str = Form(...),
    longitude: float = Form(...),
    latitude: float = Form(...),
    location_cluster: int = Form(...),
    hour: int = Form(...),
    dayofweek: int = Form(...),
    month: int = Form(...),
    image: UploadFile = File(...),
):
    # 1. Preprocess text
    cleaned_text = preprocess_text(text)

    # 2. Encode text -> IndoBERT CLS embedding (shape: [1, 1024])
    text_inputs = tokenizer(
        cleaned_text,
        return_tensors="pt",
        padding="max_length",
        truncation=True,
        max_length=128,
    )
    text_inputs = {k: v.to(device) for k, v in text_inputs.items()}
    with torch.no_grad():
        text_emb = text_model(**text_inputs).last_hidden_state[:, 0, :]
    text_emb = text_emb.cpu().numpy()

    # 3. Encode image -> EVA-CLIP image embedding (shape: [1, 1024] / sesuai model)
    img_bytes = await image.read()
    pil_img = Image.open(io.BytesIO(img_bytes)).convert("RGB")
    img_tensor = clip_preprocess(pil_img).unsqueeze(0).to(device)

    with torch.no_grad():
        img_emb = clip_model.encode_image(img_tensor)
    img_emb = img_emb.cpu().numpy()

    # 4. Additional numeric features (same order as training)
    add_feats = np.array(
        [[longitude, latitude, location_cluster, hour, dayofweek, month]],
        dtype=np.float32,
    )

    # 5. Concatenate: [image_emb, text_emb, add_feats]
    # pastikan bentuk-nya [1, dim_image + dim_text + 6]
    fused = np.concatenate([img_emb, text_emb, add_feats], axis=1)

    # 6. XGBoost prediction
    proba = xgb_model.predict_proba(fused)[0]  # shape: [2]
    pred_idx = int(np.argmax(proba))
    label = "KRITIS" if pred_idx == 1 else "WASPADA"

    return {
        "prediction": label,
        "probabilities": {
            "WASPADA": float(proba[0]),
            "KRITIS": float(proba[1]),
        },
    }

if __name__ == "__main__":
    import uvicorn

    uvicorn.run(app, host="0.0.0.0", port=7860)