import streamlit as st
import plotly.express as px
import torch

from torch import nn
from transformers import AutoTokenizer, AutoModelForSequenceClassification

defaultTxt = "I hate you cancerous insects so much"
txt = st.text_area('Text to analyze', defaultTxt)

# load tokenizer and model weights
tokenizer = AutoTokenizer.from_pretrained("s-nlp/roberta_toxicity_classifier")
model = AutoModelForSequenceClassification.from_pretrained("s-nlp/roberta_toxicity_classifier")
batch = tokenizer.encode(txt, return_tensors='pt')

# run encoding through model to get classification output
# e.g. "logits": tensor([[ 4.8982, -5.1952]], grad_fn=<AddmmBackward0>)
result = model(batch)

# transform logit to get probabilities
# e.g. tensor([[9.9996e-01, 4.2627e-05]], grad_fn=<SoftmaxBackward0>)
# first indice is neutral, second is toxic
prediction = nn.functional.softmax(result.logits, dim=-1)
neutralProb = prediction.data[0][0]
toxicProb = prediction.data[0][1]

# default text input ought to return:
# Neutral: 0.0052
# Toxic: 0.9948
st.write("Classification Probabilities")
st.write(f"{neutralProb:4.4} - NEUTRAL")
st.write(f"{toxicProb:4.4} - TOXIC")