import torch
import gradio as gr
import os
from detoxify import Detoxify
import pandas as pd
import json
import spaces
import logging
import datetime

# Load model for first time cache
model = Detoxify("unbiased-small")


@spaces.GPU
def classify(query):
    torch_device = "cuda" if torch.cuda.is_available() else "cpu"
    model = Detoxify("unbiased-small", device="cuda")

    all_result = []
    request_type = type(query)
    try:
        data = json.loads(query)
        if type(data) != list:
            data = [query]
        else:
            request_type = type(data)
    except Exception as e:
        print(e)
        data = [query]
        pass

    start_time = datetime.datetime.now()
    for i in range(len(data)):
        result = {}

        df = pd.DataFrame(model.predict(str(data[i])), index=[0])
        columns = df.columns

        for i, label in enumerate(columns):
            result[label] = df[label][0].round(3).astype("float")

        all_result.append(result)

    end_time = datetime.datetime.now()
    elapsed_time = end_time - start_time

    logging.debug("elapsed predict time: %s", str(elapsed_time))
    print("elapsed predict time:", str(elapsed_time))

    output = {}
    output["time"] = str(elapsed_time)
    output["device"] = torch_device
    output["result"] = all_result

    return json.dumps(output)


demo = gr.Interface(fn=classify, inputs=["text"], outputs="text")
demo.launch()