app.py

import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from transformers import Trainer, TrainingArguments, EarlyStoppingCallback
barthez_tokenizer = AutoTokenizer.from_pretrained("moussaKam/barthez")
model = AutoModelForSequenceClassification.from_pretrained("Anvil-ML/detecteur-ia")

def interpret_pred(pred):
    low_bond = -6.748472
    high_bound = 6.7176056
    result = "IA" if pred.argmax(dim=-1).item() == 1 else "Humain"
    pred_value = pred[0][1].item()
    interpreted_pred = (pred_value - low_bond) / (high_bound - low_bond)
    is_ai_percent = round(100 * interpreted_pred)
    return result, is_ai_percent

def interpret_pred_with_sensibility(pred):
    low_bond = -6.748472
    high_bound = 6.7176056
    pred_value = pred[0][1].item()
    interpreted_pred = (pred_value - low_bond) / (high_bound - low_bond)
    if interpreted_pred < 0.5:
        proba = "très faible"
    elif interpreted_pred < 0.6:
        proba = "faible"
    elif interpreted_pred < 0.8:
        proba = "modérée"
    elif interpreted_pred < 0.95:
        proba = "élevée"
    else:
        proba = "très élevée"

    return proba


def main(text_sentence):

    input_ids = torch.tensor(
        [barthez_tokenizer.encode(text_sentence, truncation=True, padding=True, add_special_tokens=True)]
    )

    predict = model.forward(input_ids)[0]

    #result = (
    #    "Résultat : {}.\nCe texte a {}% de chances d'avoir été généré par de l'IA"
    #    .format(interpret_pred(predict)[0], interpret_pred(predict)[1])
    #)

    proba = interpret_pred_with_sensibility(predict)
    result = (
        "La probabilité que ce texte a été généré par de l'IA est {}"
        .format(proba)
    )
    return result


iface = gr.Interface(fn=main, inputs="text", outputs="text")
iface.launch()