app.py

import streamlit as st
import pandas as pd
from transformers import BartForConditionalGeneration, TapexTokenizer, T5ForConditionalGeneration, T5Tokenizer
from prophet import Prophet

# Abrindo e lendo o arquivo CSS
with open("style.css", "r") as css:
    css_style = css.read()

# Markdown combinado com a importação da fonte e o HTML
html_content = f"""
<style>
{css_style}
@import url('https://fonts.googleapis.com/css2?family=Kanit:wght@700&display=swap');
</style>
<div style='display: flex; flex-direction: column; align-items: flex-start;'>
    <div style='display: flex; align-items: center;'>
        <div style='width: 20px; height: 5px; background-color: green; margin-right: 0px;'></div>
        <div style='width: 20px; height: 5px; background-color: red; margin-right: 0px;'></div>
        <div style='width: 20px; height: 5px; background-color: yellow; margin-right: 18px;'></div>
        <span style='font-size: 38px; font-weight: normal; font-family: "Kanit", sans-serif;'>NOSTRADAMUS</span>
    </div>
</div>
"""

# Aplicar o markdown combinado no Streamlit
st.markdown(html_content, unsafe_allow_html=True)

# Cache models to prevent re-loading on every run
#@st.cache_resource
def load_translation_model(model_name):
    return T5ForConditionalGeneration.from_pretrained(model_name)

#@st.cache_resource
def load_tapex_model():
    return BartForConditionalGeneration.from_pretrained("microsoft/tapex-large-finetuned-wtq")

#@st.cache_resource
def load_tapex_tokenizer():
    return TapexTokenizer.from_pretrained("microsoft/tapex-large-finetuned-wtq")

pt_en_translator = load_translation_model("unicamp-dl/translation-pt-en-t5")
en_pt_translator = load_translation_model("unicamp-dl/translation-en-pt-t5")
tapex_model = load_tapex_model()
tapex_tokenizer = load_tapex_tokenizer()
tokenizer = T5Tokenizer.from_pretrained("unicamp-dl/translation-pt-en-t5")

def translate(text, model, tokenizer, source_lang="pt", target_lang="en"):
    input_ids = tokenizer.encode(text, return_tensors="pt", add_special_tokens=True)
    outputs = model.generate(input_ids)
    translated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return translated_text

# Function to translate and interact with TAPEX model
def response(user_question, all_anomalies):
    question_en = translate(user_question, pt_en_translator, tokenizer, source_lang="pt", target_lang="en")
    encoding = tapex_tokenizer(table=all_anomalies, query=[question_en], padding=True, return_tensors="pt", truncation=True)
    outputs = tapex_model.generate(**encoding)
    response_en = tapex_tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
    response_pt = translate(response_en, en_pt_translator, tokenizer, source_lang="en", target_lang="pt")
    return response_pt

# Load and preprocess the data
def load_data(uploaded_file):
    if uploaded_file.name.endswith('.csv'):
        df = pd.read_csv(uploaded_file, quotechar='"', encoding='utf-8')
    elif uploaded_file.name.endswith('.xlsx'):
        df = pd.read_excel(uploaded_file)
    return df

def preprocess_data(df):
    new_df = df.iloc[2:,9:-1].fillna(0)
    new_df.columns = df.iloc[1,9:-1]
    new_df.columns = new_df.columns.str.replace(r" \(\d+\)", "", regex=True)
    month_dict = {
        'Jan': '01', 'Fev': '02', 'Mar': '03', 'Abr': '04',
        'Mai': '05', 'Jun': '06', 'Jul': '07', 'Ago': '08',
        'Set': '09', 'Out': '10', 'Nov': '11', 'Dez': '12'
    }
    
    def convert_column_name(column_name):
        # Check if the column name is 'Rótulos de Linha'
        if column_name == 'Rótulos de Linha':
            return column_name
    
        # Otherwise, proceed to convert
        parts = column_name.split('/')
        month = parts[0].strip()
        year = parts[1].strip()
    
        # Clean year in case there are extra characters
        year = ''.join(filter(str.isdigit, year))
    
        # Get month number from the dictionary
        month_number = month_dict.get(month, '00')  # Default '00' if month is not found
    
        # Return formatted date string
        return f"{month_number}/{year}"

    new_df.columns = [convert_column_name(col) for col in new_df.columns]
    new_df.columns = pd.to_datetime(new_df.columns, errors='coerce')
    new_df.rename(columns={new_df.columns[0]: 'Rotulo'}, inplace=True)
    df_clean = new_df.copy()
    return df_clean

# Cache the Prophet computation to avoid recomputing
@st.cache_data
def apply_prophet(df_clean):
    if df_clean.empty:
        st.error("DataFrame está vazio após o pré-processamento.")
        return pd.DataFrame()

    # Debugging: Check structure of df_clean
    #st.write("Estrutura do DataFrame df_clean:")
    #st.write(df_clean)

    # Criar um DataFrame vazio para armazenar todas as anomalias
    all_anomalies = pd.DataFrame()
    
    # Processar cada linha no DataFrame
    for index, row in df_clean.iterrows():
        # Extract timestamp and value columns
        date_columns = [col for col in df_clean.columns if isinstance(col, pd.Timestamp)]
        data = pd.DataFrame({
            'ds': date_columns,
            'y': row[date_columns].values
        })

        # Debugging: Check the data passed into Prophet
        #st.write(f"Dados para Prophet - Grupo {row['Rotulo']}:")
        #st.write(data)

        # Remove rows where 'y' is zero or missing
        data = data[data['y'] > 0].dropna().reset_index(drop=True)

        # Ensure there's enough data for Prophet to run
        if data.empty or len(data) < 2:
            #st.write(f"Pular grupo {row['Rotulo']} por não ter observações suficientes.")
            continue

        try:
            # Create and fit the Prophet model
            model = Prophet(interval_width=0.95)
            model.fit(data)
        except ValueError as e:
            #st.write(f"Pular grupo {row['Rotulo']} devido ao erro: {e}")
            continue

        # Make future predictions
        future = model.make_future_dataframe(periods=12, freq='M')
        forecast = model.predict(future)

        # Add real values and calculate anomalies
        real_values = list(data['y']) + [None] * (len(forecast) - len(data))
        forecast['real'] = real_values
        anomalies = forecast[(forecast['real'] < forecast['yhat_lower']) | (forecast['real'] > forecast['yhat_upper'])]

        # Debugging: Check the anomalies detected
        #st.write(f"Anomalias detectadas para o grupo {row['Rotulo']}:")
        #st.write(anomalies)

        # Add group label and append anomalies to all_anomalies DataFrame
        anomalies['group'] = row['Rotulo']
        all_anomalies = pd.concat([all_anomalies, anomalies[['ds', 'real', 'group']]], ignore_index=True)

    # Return the dataframe of all anomalies
    st.write(f"Concluída a aplicação do modelo de série tempotal")
    st.write(all_anomalies.head())
    all_anomalies.sort_values(by=['real'], ascending=False, inplace=True)
    all_anomalies['real'] = pd.to_numeric(all_anomalies['real'], errors='coerce')
    all_anomalies = all_anomalies[all_anomalies['real'] >= 10000000.00]
    all_anomalies = all_anomalies.astype(str)
    return all_anomalies

# Initialize session states
if 'all_anomalies' not in st.session_state:
    st.session_state['all_anomalies'] = pd.DataFrame()

if 'history' not in st.session_state:
    st.session_state['history'] = []

tab1, tab2 = st.tabs(["Meta Prophet", "Microsoft TAPEX"])

# Interface para carregar arquivo
uploaded_file = st.file_uploader("Carregue um arquivo CSV ou XLSX", type=['csv', 'xlsx'])

with tab1:
    if uploaded_file:
        df = load_data(uploaded_file)
        df_clean = preprocess_data(df)
        
        if df_clean.empty:
            st.warning("Não há dados válidos para processar.")
        else:
            # Cache the Prophet results
            #if st.session_state['all_anomalies'].empty:
                #with st.spinner('Aplicando modelo de série temporal...'):
                    #all_anomalies = apply_prophet(df_clean)
                    #st.session_state['all_anomalies'] = all_anomalies
            # Cache the Prophet results
            if st.session_state['all_anomalies'].empty:
                all_anomalies = apply_prophet(df_clean)
                st.session_state['all_anomalies'] = all_anomalies

with tab2:
    if 'all_anomalies' in st.session_state and not st.session_state['all_anomalies'].empty:
        user_question = st.text_input("Escreva sua questão aqui:", "")
        if user_question:
            bot_response = response(user_question, st.session_state['all_anomalies'])
            st.session_state['history'].append(('👤', user_question))
            st.session_state['history'].append(('🤖', bot_response))
        
        for sender, message in st.session_state['history']:
            st.markdown(f"**{sender} {message}**")
        
        if st.button("Limpar histórico"):
            st.session_state['history'] = []
    else:
        st.warning("Por favor, processe os dados no Meta Prophet primeiro.")