from transformers import RobertaTokenizer, AutoModelForSequenceClassification
from scipy.special import expit
import numpy as np
import os
import gradio as gr
import requests
from datetime import datetime

# set up model
authtoken = os.environ.get("TOKEN")
tokenizer = RobertaTokenizer.from_pretrained("guidecare/feelings_and_issues_large_v2", token=authtoken, use_safetensors=True)
tokenizer.do_lower_case = True
model = AutoModelForSequenceClassification.from_pretrained("guidecare/feelings_and_issues_large_v2", token=authtoken, use_safetensors=True)
all_label_names = list(model.config.id2label.values())


def predict(text):
    probs = expit(model(**tokenizer([text], return_tensors="pt", padding=True)).logits.detach().numpy())
    # can't use numpy for whatever reason
    probs = [float(np.round(i, 2)) for i in probs[0]]
    # break out issue, harm, sentiment, feeling
    zipped_list = list(zip(all_label_names, probs))
    print(text, zipped_list)
    issues = [(i, j) for i, j in zipped_list if i.startswith('issue')]
    feelings = [(i, j) for i, j in zipped_list if i.startswith('feeling')]
    harm = [(i, j) for i, j in zipped_list if i.startswith('harm')]
    sentiment = [(i, j) for i, j in zipped_list if i.startswith('sentiment')]
    # keep tops for each one
    issues = sorted(issues, key=lambda x: x[1])[::-1]
    feelings = sorted(feelings, key=lambda x: x[1])[::-1]
    harm = sorted(harm, key=lambda x: x[1])[::-1]
    sentiment = sorted(sentiment, key=lambda x: x[1])[::-1]
    # top is the combo of these
    top = issues + feelings + harm + sentiment

    d = {i: j for i, j in top}
    return d


iface = gr.Interface(
  fn=predict, 
  inputs="text",
  outputs="label",
  #examples=["This test tomorrow is really freaking me out."]
)

iface.launch()