import numpy as np
import pandas as pd
import time
import streamlit as st
import matplotlib.pyplot as plt
import seaborn as sns

import jax
import jax.numpy as jnp

from transformers import AlbertTokenizer

from custom_modeling_albert_flax import CustomFlaxAlbertForMaskedLM

@st.cache(show_spinner=True,allow_output_mutation=True)
def load_model():
    tokenizer = AlbertTokenizer.from_pretrained('albert-xxlarge-v2')
    model = CustomFlaxAlbertForMaskedLM.from_pretrained('albert-xxlarge-v2',from_pt=True)
    return tokenizer,model

def clear_data():
    for key in st.session_state:
        del st.session_state[key]

if __name__=='__main__':

    # Config
    max_width = 1500
    padding_top = 0
    padding_right = 2
    padding_bottom = 0
    padding_left = 2

    define_margins = f"""
    <style>
        .appview-container .main .block-container{{
            max-width: {max_width}px;
            padding-top: {padding_top}rem;
            padding-right: {padding_right}rem;
            padding-left: {padding_left}rem;
            padding-bottom: {padding_bottom}rem;
        }}
    </style>
    """
    hide_table_row_index = """
                <style>
                tbody th {display:none}
                .blank {display:none}
                </style>
                """
    st.markdown(define_margins, unsafe_allow_html=True)
    st.markdown(hide_table_row_index, unsafe_allow_html=True)

    tokenizer,model = load_model()
    mask_id = tokenizer('[MASK]').input_ids[1:-1][0]

    sent_1 = st.sidebar.text_input('Sentence 1',value='It is better to play a prank on Samuel than Craig because he gets angry less often.',on_change=clear_data)
    sent_2 = st.sidebar.text_input('Sentence 2',value='It is better to play a prank on Samuel than Craig because he gets angry more often.',on_change=clear_data)
    input_ids_1 = tokenizer(sent_1).input_ids
    input_ids_2 = tokenizer(sent_2).input_ids
    input_ids = np.array([input_ids_1,input_ids_2])

    outputs = model(input_ids, interv_type='swap', interv_dict = {0:{'lay':[(8,1,[0,1])]}})
    logprobs = jax.nn.log_softmax(outputs.logits, axis = -1)
    preds = [np.random.choice(np.arange(len(probs)),p=np.exp(probs)/np.sum(np.exp(probs))) for probs in logprobs[0]]
    st.write([tokenizer.decode([token]) for token in preds])