src/api.py

from fastapi import FastAPI, Query
from pydantic import BaseModel
from typing import List
from transformers import BertTokenizer, BertForSequenceClassification
import torch
import pickle
import random
from collections import defaultdict

app = FastAPI()

# Model və tokenizer yükləmə
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
model = BertForSequenceClassification.from_pretrained("best_model")
model.eval()

with open("best_model/label_encoder.pkl", "rb") as f:
    label_encoder = pickle.load(f)

class PredictionResponse(BaseModel):
    disease: str
    probability: float

@app.get("/predict", response_model=List[PredictionResponse])
def predict(symptoms: str = Query(..., description="Comma-separated symptoms")):
    symptoms_list = [s.strip() for s in symptoms.split(",") if s.strip()]
    agg_probs = defaultdict(float)
    n_shuffles = 10

    for _ in range(n_shuffles):
        random.shuffle(symptoms_list)
        shuffled_text = ", ".join(symptoms_list)
        inputs = tokenizer(shuffled_text, return_tensors="pt", truncation=True, padding=True, max_length=128)
        with torch.no_grad():
            outputs = model(**inputs)
            probs = torch.nn.functional.softmax(outputs.logits, dim=-1).squeeze()
        for i, p in enumerate(probs):
            agg_probs[i] += p.item()
    for k in agg_probs:
        agg_probs[k] /= n_shuffles
    top_3 = sorted(agg_probs.items(), key=lambda x: x[1], reverse=True)[:3]

    results = []
    for idx, prob in top_3:
        label = label_encoder.classes_[idx]
        results.append({"disease": label, "probability": float(prob)})

    return results