classification_model_monitor.py

import streamlit as st
import pandas as pd
import numpy as np
import seaborn as sns
import plotly.express as px
import matplotlib.pyplot as plt
# from read_from_prediction_db import PredictionDBRead
# from read_daily_metrics_from_db import MetricsDBRead
from sklearn.metrics import balanced_accuracy_score, accuracy_score
import logging
from config import (CLASSIFIER_ADJUSTMENT_THRESHOLD,
                    PERFORMANCE_THRESHOLD,
                    CLASSIFIER_THRESHOLD)

logging.basicConfig(format='%(asctime)s %(levelname)s: %(message)s', level=logging.INFO)


def filter_prediction_data(data: pd.DataFrame):
    try:
        logging.info("Entering filter_prediction_data()")
        if data is None:
            raise Exception("Input Prediction Data frame in None")   

        # filtered_prediction_data = data.loc[(data['y_true'].isin(['WEATHER', 'EDUCATION', 'ASTROLOGY', 'OTHERS']) == False) & 
        #                                     (data['y_pred'].isin(['WEATHER', 'EDUCATION', 'ASTROLOGY', 'OTHERS']) == False) & 
        #                                     (data['y_true_proba'] > CLASSIFIER_THRESHOLD)].copy()

        # filtered_prediction_data = data.loc[data['y_true_proba'] == 1].copy()

        filtered_prediction_data = data.loc[(data['y_true_proba'] == 1) & (data['used_for_training'].astype("str").str.contains("_train")==False) & 
                                                                          (data['used_for_training'].astype("str").str.contains("_excluded")==False) & 
                                                                          (data['used_for_training'].astype("str").str.contains("_validation")==False) 
                                           ].copy()
            
        logging.info("Exiting filter_prediction_data()")
        return filtered_prediction_data
    except Exception as e:
        logging.critical(f"Error in filter_prediction_data(): {e}")
        return None


def get_adjusted_predictions(df):
    try:
        logging.info("Entering get_adjusted_predictions()")
        if df is None:
            raise Exception('Input Filtered Prediction Data Frame is None')
        df = df.copy()
        df.reset_index(drop=True, inplace=True)
        df.loc[df['y_pred_proba']<CLASSIFIER_ADJUSTMENT_THRESHOLD, 'y_pred'] = 'NATION'
        # df.loc[(df['text'].str.contains('Pakistan')) & (df['y_pred'] == 'NATION'), 'y_pred'] = 'WORLD'
        # df.loc[(df['text'].str.contains('Zodiac Sign', case=False)) | (df['text'].str.contains('Horoscope', case=False)), 'y_pred'] = 'SCIENCE'
        logging.info("Exiting get_adjusted_predictions()")
        return df
    except Exception as e:
        logging.info(f"Error in get_adjusted_predictions(): {e}")
        return None


def display_kpis(data: pd.DataFrame, adj_data: pd.DataFrame):
    try:
        logging.info("Entering display_kpis()")
        if data is None:
            raise Exception("Input Prediction Data frame in None")
        if adj_data is None:
            raise Exception('Input Adjusted Data frame is None')    
        
        n_samples = len(data)
        balanced_accuracy = np.round(balanced_accuracy_score(data['y_true'], data['y_pred']), 4)
        accuracy = np.round(accuracy_score(data['y_true'], data['y_pred']), 4)
            
        adj_balanced_accuracy = np.round(balanced_accuracy_score(adj_data['y_true'], adj_data['y_pred']), 4)
        adj_accuracy = np.round(accuracy_score(adj_data['y_true'], adj_data['y_pred']), 4)

        st.write('''<style>
                        [data-testid="column"] {
                        width: calc(33.3333% - 1rem) !important;
                        flex: 1 1 calc(33.3333% - 1rem) !important;
                        min-width: calc(33% - 1rem) !important;
                        }
                    </style>''', 
                 unsafe_allow_html=True)

        col1, col2= st.columns(2)
        with col1:
            metric1 = st.metric(label="Balanced Accuracy", value=balanced_accuracy)
        with col2:
            metric2 = st.metric(label="Adj Balanced Accuracy", value=adj_balanced_accuracy)
            
        col3, col4= st.columns(2)
        with col3:
            metric3 = st.metric(label="Accuracy", value=accuracy)
        with col4:
            metric4 = st.metric(label="Adj Accuracy", value=adj_accuracy)
    
        col5, col6= st.columns(2)
        with col5:
            metric5 = st.metric(label="Bal Accuracy Threshold", value=PERFORMANCE_THRESHOLD)
        with col6:
            metric6 = st.metric(label="N Samples", value=n_samples)
        logging.info("Exiting display_kpis()")
    except Exception as e:
        logging.critical(f'Error in display_kpis(): {e}')
        st.error("Couldn't display KPIs")


def plot_daily_metrics(metrics_df: pd.DataFrame):
    try:
        logging.info("Entering plot_daily_metrics()")
        st.write(" ")
        if metrics_df is None:
            raise Exception('Input Metrics Data Frame is None')
            
        metrics_df['evaluation_date'] = pd.to_datetime(metrics_df['evaluation_date'])
        metrics_df['mean_score_minus_std'] = np.round(metrics_df['mean_balanced_accuracy_score'] - metrics_df['std_balanced_accuracy_score'], 4)
        metrics_df['mean_score_plus_std'] = np.round(metrics_df['mean_balanced_accuracy_score'] + metrics_df['std_balanced_accuracy_score'], 4)
        
        hover_data={'mean_balanced_accuracy_score': True,
                    'std_balanced_accuracy_score': False, 
                    'mean_score_minus_std': True,
                    'mean_score_plus_std': True,
                    'evaluation_window_days': True,
                    'n_splits': True,
                    'sample_start_date': True,
                    'sample_end_date': True, 
                    'sample_size_of_each_split': True}
    
        hover_labels = {'mean_balanced_accuracy_score': "Mean Score",
                        'mean_score_minus_std': "Mean Score - Stdev",
                        'mean_score_plus_std': "Mean Score + Stdev",
                        'evaluation_window_days': "Observation Window (Days)",
                        'sample_start_date': "Observation Window Start Date",
                        'sample_end_date': "Observation Window End Date",
                        'n_splits': "N Splits For Evaluation",
                        'sample_size_of_each_split': "Sample Size of Each Split"}
        
        fig = px.line(data_frame=metrics_df, x='evaluation_date', 
                      y='mean_balanced_accuracy_score', 
                      error_y='std_balanced_accuracy_score', 
                      title="Daily Balanced Accuracy",
                      color_discrete_sequence=['black'],
                      hover_data=hover_data, labels=hover_labels, markers=True)
    
        fig.add_hline(y=PERFORMANCE_THRESHOLD, line_dash="dash", line_color="green", 
                      annotation_text=f"<b>THRESHOLD</b>",
                      annotation_position="left top")
        
        fig.update_layout(dragmode='pan')
        fig.update_layout(margin=dict(l=0, r=0, t=110, b=10))
        st.plotly_chart(fig, use_container_width=True)
        logging.info("Exiting plot_daily_metrics()")
    except Exception as e:
        logging.critical(f'Error in plot_daily_metrics(): {e}')
        st.error("Couldn't Plot Daily Model Metrics")


def get_misclassified_classes(data):
    try:
        logging.info("Entering get_misclassified_classes()")
        if data is None:
            raise Exception("Input Prediction Data Frame is None")
        
        data = data.copy()
        data['match'] = (data['y_true'] == data['y_pred']).astype('int')
        y_pred_counts = data['y_pred'].value_counts()
        
        misclassified_examples = data.loc[data['match'] == 0, ['text', 'y_true', 'y_pred', 'y_pred_proba', 'url']].copy()
        misclassified_examples.sort_values(by=['y_pred', 'y_pred_proba'], ascending=[True, False], inplace=True)
        
        misclassifications = data.loc[data['match'] == 0, 'y_pred'].value_counts()

        missing_classes = [i for i in y_pred_counts.index if i not in misclassifications.index]
        for i in missing_classes:
            misclassifications[i] = 0
        
        misclassifications = misclassifications[y_pred_counts.index].copy()
        misclassifications /= y_pred_counts
        misclassifications.sort_values(ascending=False, inplace=True)
        logging.info("Exiting get_misclassified_classes()")
        return np.round(misclassifications, 2), misclassified_examples
    except Exception as e:
        logging.critical(f'Error in get_misclassified_classes(): {e}')
        return None, None

    
def display_misclassified_examples(misclassified_classes, misclassified_examples):
    try:
        logging.info("Entering display_misclassified_examples()")
        st.write(" ")
        if misclassified_classes is None:
            raise Exception('Misclassified Classes Distribution Data Frame is None')
        if misclassified_examples is None:
            raise Exception('Misclassified Examples Data Frame is None')
        
        fig, ax = plt.subplots(figsize=(10, 4.5))
        misclassified_classes.plot(kind='bar', ax=ax, color='black', title="Misclassification percentage")
        plt.yticks([])
        plt.xlabel("")
        ax.bar_label(ax.containers[0]);
        st.pyplot(fig)
    
        st.markdown("<b>Misclassified examples</b>", unsafe_allow_html=True)    
        st.dataframe(misclassified_examples, hide_index=True)
        st.markdown(
                    """
                    <style>
                    [data-testid="stElementToolbar"] {
                        display: none;
                    }
                    </style>
                    """,
                    unsafe_allow_html=True
                )
        logging.info("Exiting display_misclassified_examples()")
    except Exception as e:
        logging.critical(f'Error in display_misclassified_examples(): {e}')
        st.error("Couldn't display Misclassification Data")


def classification_model_monitor(prediction_data, metrics_df):
    try:
        st.write("<h4>Classification Model Monitor</h4>", unsafe_allow_html=True)

        # prediction_db = PredictionDBRead()
        # metrics_db = MetricsDBRead()
        
        # Read Prediction Data From DB
        # prediction_data = prediction_db.read_predictions_from_db()
        # Filter Prediction Data
        filtered_prediction_data = filter_prediction_data(prediction_data)
        # Get Adjusted Prediction Data
        adjusted_filtered_prediction_data = get_adjusted_predictions(filtered_prediction_data)
        # Display KPIs
        display_kpis(filtered_prediction_data, adjusted_filtered_prediction_data)
        
        # Read Daily Metrics From DB
        # metrics_df = metrics_db.read_metrics_from_db()
        # Display daily Metrics Line Plot
        plot_daily_metrics(metrics_df)
        
        # Get misclassified class distribution and misclassified examples from Prediction Data
        misclassified_classes, misclassified_examples = get_misclassified_classes(filtered_prediction_data)
        # Display Misclassification Data
        display_misclassified_examples(misclassified_classes, misclassified_examples)
        
    except Exception as e:
        logging.critical(f"Error in classification_model_monitor(): {e}")
        st.error("Unexpected Error. Couldn't display Model Monitor")