Blame for old_code/decision_boundary.py

from collections import deque
from pathlib import Path
import pickle

import gradio as gr

import matplotlib.pyplot as plt
import matplotlib.colors as mcolors

import black
import cv2
import inspect
import numpy as np
import pandas as pd
import io
from jinja2 import Template
from PIL import Image
import sklearn
from sklearn.linear_model import LogisticRegression
from sklearn.svm import LinearSVC
from sklearn.base import ClassifierMixin
from sklearn.datasets import load_iris
from sklearn.decomposition import PCA
from sklearn.metrics import classification_report
import traceback
import yaml

from util import *

import logging
logging.basicConfig(
    level=logging.INFO,  # set minimum level to capture (DEBUG, INFO, WARNING, ERROR, CRITICAL)
    format="%(asctime)s [%(levelname)s] %(message)s",  # log format
)
logger = logging.getLogger("ELVIS")

# TODO:
# - support for session: load a previous session and continue from there

def label2color(labels, cmap=None):
    '''
    Parameters
    ----------
    labels: a list of distinct strings
    '''
    if (cmap is not None) and (cmap != ''):
        # sample the number of needed colors from user-specified color map
        color_deque = deque(plt.get_cmap(cmap, len(labels)).colors) 
    elif len(labels) <= 10:
        #color_deque = deque(["red", "green", "blue", "yellow", "orange", "purple", "pink", "brown", "gray", "black"])
        #color_deque = deque(mcolors.TABLEAU_COLORS.keys())
        color_deque = deque([c.replace('tab:', '') for c in mcolors.TABLEAU_COLORS])
    elif len(labels) <= 148:
        color_deque = deque(mcolors.CSS4_COLORS)
    elif len(labels) <= 949: 
        color_deque = deque([c.replace('xkcd:', '') for c in mcolors.XKCD_COLORS])
    else: # very unlikely 
        color_deque = deque(plt.get_cmap('vridis', len(labels))) 
    
    colors = []

    for label in labels:
        print(label, color_deque)
        if label.lower() in color_deque:
            colors.append(label.lower())
            color_deque.remove(label.lower())
        else:
            colors.append(color_deque.popleft())

    return colors

def toydata():
    points = [['Red', 0.12375, 0.8516666666666667], 
              ['Red', 0.19, 0.8916666666666666], 
              ['Red', 0.27375, 0.9233333333333333], 
              ['Blue', 0.50625, 0.785], 
              ['Blue', 0.38375, 0.6733333333333333], 
              ['Blue', 0.28875, 0.595]]
    df = pd.DataFrame(points, columns=['label', 'F1', 'F2'])
    return df

class CoordinateProjection2d:
    """
    Project data on the two coordinates.
    """

    def __init__(self, dim0=0, dim1=1, **kwargs):
        self.dims = [dim0, dim1]

    def fit(self, X):
        self.mean = X.mean(axis=0)
        return self

    def transform(self, X):
        self._check_dims(X)
        return X[:, self.dims]

    def fit_transform(self, X):
        self.fit(X)
        return self.transform(X)

    def inverse_transform(self, Z):
        X = np.ones((len(Z), 1)) * self.mean
        self._check_dims(X)
        X[:, self.dims] = Z
        return X

    def _check_dims(self, X):
        n_features = X.shape[1]
        if self.dims[0] >= n_features:
            raise ValueError(f"dim0={self.dims[0]} exceeds the number of features {n_features}")
        if self.dims[1] >= n_features:
            raise ValueError(f"dim1={self.dims[1]} exceeds the number of features {n_features}")


class InteractiveDecisionBoundary:
    DATASET_FILE = "dataset.csv"
    MODEL_FILE = "model.pkl"
    FIGURE_BASENAME = "figure"

    CODE_FILE = "generated_code.py"
    EXPORT_CODE_TEMPLATE = "export_code_template.py.j2"

    def __init__(self, width, height):
        # initialized in draw_plot
        #self.canvas_width = -1
        #self.canvas_height = -1

        self.canvas_width = width
        self.canvas_height = height

        supported_classifier_names = yaml.safe_load(Path("model_imports.yaml").read_text())
        self.classifiers = {k: v for k, v in get_sklearn_classifiers().items() if k in supported_classifier_names}

        self.model_class = LinearSVC
        self.model_args = ""
        self.model = self.model_class()

        self.dataloaders = get_sklearn_dataloaders()

        supported_embedder_names = yaml.safe_load(Path("embedder_imports.yaml").read_text())
        self.embedders = {
            'CoordinateProjection2d': CoordinateProjection2d,
            'GaussianRandomProjection': sklearn.random_projection.GaussianRandomProjection,
            'SparseRandomProjection': sklearn.random_projection.SparseRandomProjection,
        }
        module = getattr(sklearn, 'decomposition')
        for cls_name, cls in inspect.getmembers(module, inspect.isclass):
            if cls_name in supported_embedder_names:
                self.embedders[cls_name] = cls

        # normalizers
        self.normalizers = {
            'None': None,
            'MinMaxScaler': sklearn.preprocessing.MinMaxScaler,
            'StandardScaler': sklearn.preprocessing.StandardScaler,
        }

        # data embedding and preprocessing values
        self.embedder_class = CoordinateProjection2d
        self.embedder_args = ""
        self.normalizer_class = None
        self.jitter_std = 0

        # todo: support arbitrary number of classes and user-defined class labels 
        #self.dataset = toydata()
        #iris = load_iris(as_frame=True)
        #self.dataset = pd.concat([iris.data, iris.target], axis=1)
        #self.dataset = self.dataset.rename(columns={'target': 'label'})
        self.dataset = pd.DataFrame(columns=['target', 'F1', 'F2'])
        self.dataset_type = 'Draw2D'
        self.custom_selected = True

        # options
        self.num_dots = 200
        self.dpi = 100
        self.cmap = None
        self.precision = 2 # number of decimal places to show in datatable
        self.marker_size = 100

        self.data_image = None
        self.boundary_image = None
        self._axis_topleft = (0, 0) 

        self.figure_extension = ".svg"

        self.css ="""
#my-button {
    height: 30px;
    font-size: 16px;
}

#rowheight {
    height: 90px;
}

.file-chooser {
    height: 150px;
}

.hidden-button {
    display: none;
}

.report-table {
    border: 0 !important;
}
.report-table tr, .report-table th, .report-table td, .report-table tbody, .report-table thead {
    border: 0 !important;
    padding: 6px 12px;
    text-align: center;
}"""

    def _get_features(self):
        """Get the feature values from the current dataset, applying normalization if set."""
        X = self.dataset.loc[:, self.dataset.columns != 'target'].values
        if len(X) == 0:
            raise ValueError("The dataset is empty or not properly formatted.")
        return X

    def _process_features(self, features):
        if self.normalizer_class is not None:
            normalizer = self.normalizer_class()
            features = normalizer.fit_transform(features)
        if self.jitter_std > 0:
            noise = np.random.normal(0, self.jitter_std, features.shape)
            features += noise
        return features

    def _embed_features(self, features, return_embedder=False):
        embedder = self.embedder_class(n_components=2, **parse_param_string(self.embedder_args))
        features = embedder.fit_transform(features)

        if return_embedder:
            return features, embedder
        return features

    def _reset_data_processing_and_embedding(self):
        # Reset the values
        self.normalizer_class = None
        self.jitter_std = 0
        self.embedder_class = CoordinateProjection2d
        self.embedder_args = ""

    def plot(self, decision_boundary=False, save_figure=False):
        '''
        Plot data and decision boundary with matplotlib and return as PIL image.
        '''

        logger.info("Initializing figure")
        fig = plt.figure(figsize=(self.canvas_width/100., self.canvas_height/100.0), dpi=self.dpi)
        # set entire figure to be the canvas to allow simple conversion of mouse
        # position to coordinates in the figure
        ax = fig.add_axes([0., 0., 1., 1.]) # 
        ax.margins(x=0, y=0) # no padding in both directions

        if self.dataset_type == 'Draw2D':
            # draw canvas boundary
            #ax.scatter([0, 0, 1, 1], [0, 1, 0, 1], color='brown')
            # DO NOT CHANGE THE COLOR OF THE BOUNDARY
            # IT WILL BREAK THE ORIGIN COORDINATE DETECTION
            ax.plot([0, 0, 1, 1, 0], [0, 1, 1, 0, 0], color='black')
        
        for spine in ax.spines.values():
            spine.set_color((0.1, 0.1, 0.1))

        # TODO: allow showing x and y axes with ticks and labels
        if (self.dataset is not None and len(self.dataset) > 0):
            try: 
                X = self._get_features()
                y = self.dataset.target.values
                logger.info("Data:\n" + str(X))
                logger.info("Target:\n" + str(y)) 

                # preprocess features
                X = self._process_features(X)

                # embed features to 2D for visualization
                Z, embedder = self._embed_features(X, return_embedder=True)

                #ax.set_title("Click to add points")
                labels = np.unique(y)
                colors = label2color(labels, cmap=self.cmap)
                logger.info("Classes:\n" + str(labels)) 
                logger.info("Colors:\n" + str(colors)) 
                l2c = dict(zip(labels, colors))

                # scatter plots for data
                for l, label in enumerate(labels):
                    #print('class', label)
                    #ax.scatter(*zip(*self.dataset[self.dataset.label == label].features), color=label, label=label)
                    subset = Z[y == label]
                    ax.scatter(subset[:, 0], subset[:, 1], color=colors[l], label=label, s=self.marker_size)
                    ax.legend()

                # plot the decision boundary
                if decision_boundary:
                    model = self.model_class(**parse_param_string(self.model_args))
                    model.fit(X, y)
                    self.model = model

                    # plot decision boundary in the projected space
                    # xx, yy = np.meshgrid(np.linspace(Z[:, 0].min(), 1, 100), np.linspace(0, 1, 100))
                    # Note: Should not apply normalization/jittering to meshgrid points
                    if self.dataset_type == 'Draw2D':
                        xx, yy = np.meshgrid(np.linspace(0, 1, self.num_dots),
                                             np.linspace(0, 1, self.num_dots))
                    else:
                        xx, yy = np.meshgrid(np.linspace(Z[:, 0].min(), Z[:, 0].max(), self.num_dots),
                                             np.linspace(Z[:, 1].min(), Z[:, 1].max(), self.num_dots))

                    grid = np.c_[xx.ravel(), yy.ravel()]
                    #scores = clf.decision_function(grid)[:, 1].reshape(xx.shape)
                    #scores = clf.decision_function(grid).reshape(xx.shape)
                    #ax.contour(xx, yy, scores)#, levels=[0], colors="black", linestyles="--")
                    print('grid', grid)
                    print('inverse', embedder.inverse_transform(grid))
                    preds = model.predict(embedder.inverse_transform(grid)).reshape(xx.shape)
                    #print(preds.shape, xx.shape, yy.shape)
                    ax.scatter(xx.ravel(), yy.ravel(), c=[l2c[l] for l in preds.ravel()], s=1, alpha=0.5)

            except Exception as e:
                print(traceback.format_exc())
                #raise gr.Error(f"⚠️   {e}")
                gr.Info(f"⚠️   {e}")

        buf = io.BytesIO()
        ax.figure.savefig(buf, format="png", bbox_inches="tight", pad_inches=0)
        plt.close(fig)
        buf.seek(0)
        img = Image.open(buf)

        if save_figure:
            ax.figure.savefig(f"{self.FIGURE_BASENAME}{self.figure_extension}")

        # detect axis pixel positions
        if self.dataset_type == 'Draw2D':
            array = np.array(img.convert("RGB"))
            bgr = cv2.cvtColor(array, cv2.COLOR_RGB2BGR)
            gray = cv2.cvtColor(bgr, cv2.COLOR_BGR2GRAY)
            black_mask = gray < 0.05 * 255
            contours, _ = cv2.findContours(black_mask.astype(np.uint8), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

            # find the contour with the largest area
            max_area = 0
            most_likely_topleft = 0, 0
            for contour in contours:
                x, y, w, h = cv2.boundingRect(contour)
                area = w * h
                if w * h > max_area:
                    max_area = area
                    most_likely_topleft = x, y

            self._axis_topleft = most_likely_topleft
        else:
            self._axis_topleft = 0, 0
        
        # TODO: add a save function for saving screenshot
        #img.save('image.png')

        return img

    def update_resolution(self, num_dots):
        self.num_dots = num_dots
        return self.plot(decision_boundary=True)

    def update_dpi(self, dpi):
        self.dpi = dpi
        return self.plot(decision_boundary=True)

    def update_cmap(self, cmap):
        self.cmap = cmap
        return self.plot(decision_boundary=True)

    def update_precision(self, precision):
        self.precision = precision
        data_table = gr.Dataframe(
            value=self.dataset.round(self.precision), 
            visible=True,
            headers=list(self.dataset.columns),
        )
        return data_table

    def update_marker_size(self, size):
        self.marker_size = size
        return self.plot(decision_boundary=True)

    def add_point(self, evt: gr.SelectData, label):
        '''
        Mouse click to add a point.
        '''
        if self.custom_selected:
            if self.dataset_type != 'Draw2D':
                self.dataset = pd.DataFrame(columns=['target', 'F1', 'F2'])
                self.dataset_type = 'Draw2D'

            shift_x, shift_y = self._axis_topleft
            
            # normalize clicked position to [0, 1]
            x = (evt.index[0] - shift_x)  / self.canvas_width
            y = 1 - (evt.index[1] - shift_y) / self.canvas_height  # flip y-axis to match matplotlib

            if 0 <= x <= 1 and 0 <= y <= 1:
                self.dataset.loc[len(self.dataset)] = [label, x, y]

            logger.info(f'clicked ({evt.index[0]}, {evt.index[1]}), mapped to ({x}, {y})')
        
        vis = self.plot()
        data_table = gr.Dataframe(
            value=self.dataset.round(self.precision), 
            visible=True,
            headers=list(self.dataset.columns),
        )

        return vis, data_table

    # train a model and show decision boundary
    def train(self):
        image = self.plot(decision_boundary=True)
        try:
            X = self.dataset.loc[:, self.dataset.columns != 'target'].values
            y = self.dataset.target.values
            pred = self.model.predict(X)
            df = pd.DataFrame(classification_report(y, pred, output_dict=True)).T
            summary = df.to_html(classes="report-table", float_format="%.2f")

            return image, gr.HTML(visible=True), "<b>Classification report</b><br>" + summary
        except Exception as e:
            print(traceback.format_exc())
            gr.Info(f"⚠️   {e}")
            
            return image, gr.HTML(visible=False), ""

    # clear data points and data preprocessing
    def clear(self):
        self.dataset = pd.DataFrame(columns=['target', 'F1', 'F2'])
        return self.plot(), gr.Dataframe(visible=False)

    def save(self):
        self.save_data()

    # save dataset 
    def save_data(self):
        # TODO: allow user-specified filename
        self.dataset.to_csv(self.DATASET_FILE, index=False)
        logger.info(f"{self.DATASET_FILE} updated")
        return self.DATASET_FILE

    def update_model(self, classifier_name):
        self.model_class = self.classifiers[classifier_name]
        self.args_textbox.value = ""
        logger.info(f'Updated model to {self.model_class}')

        return ""

    def save_model(self):
        with open(self.MODEL_FILE, "wb") as f:
            pickle.dump(self.model, f)
        logger.info(f"{self.MODEL_FILE} updated")
        return self.MODEL_FILE

    def save_code(self):
        model_class = str(self.model_class.__name__)
        model_imports = yaml.safe_load(Path("model_imports.yaml").read_text())
        if model_class not in model_imports:
            raise ValueError(f"Model {model_class} not found in model_imports.yaml")
        model_import_stmt = f"{model_imports[model_class]}"

        embedder_class = str(self.embedder_class.__name__)
        if embedder_class == "CoordinateProjection2d":
            embedder_import_stmt = f"\n\n\n{inspect.getsource(CoordinateProjection2d)}".rstrip()
        else:
            embedder_imports = yaml.safe_load(Path("embedder_imports.yaml").read_text())
            if embedder_class not in embedder_imports:
                raise ValueError(f"Embedder {embedder_class} not found in embedder_imports.yaml")
            embedder_import_stmt = f"\n{embedder_imports[embedder_class]}"

        if self.normalizer_class is not None:
            normalizer_class = str(self.normalizer_class.__name__)
            normalizer_imports = yaml.safe_load(Path("normalizer_imports.yaml").read_text())
            if normalizer_class not in normalizer_imports:
                raise ValueError(f"Normalizer {normalizer_class} not found in normalizer_imports.yaml")
            normalizer_import_stmt = f"\n{normalizer_imports[normalizer_class]}"
        else:
            normalizer_import_stmt = ""

        if self.dataset_type == 'Draw2D':
            x_min = 0
            x_max = 1
            y_min = 0
            y_max = 1
        else:
            x_min = "X_embedded[:, 0].min()"
            x_max = "X_embedded[:, 0].max()"
            y_min = "X_embedded[:, 1].min()"
            y_max = "X_embedded[:, 1].max()"

        model_params = parse_param_string(self.model_args)
        if len(model_params) == 0:
            model_params_text = ""
        else:
            model_params_text = "".join([f"\n\t\t{k}={repr(v)}," for k, v in model_params.items()]) + "\n\t"

        if embedder_class == "CoordinateProjection2d":
            embedder_args = {**parse_param_string(self.embedder_args)}
        else:
            embedder_args = {"n_components": 2, **parse_param_string(self.embedder_args)}
        embedder_args_text = "".join([f"\n\t\t{k}={repr(v)}," for k, v in embedder_args.items()]) + "\n\t"

        template = Template(Path(self.EXPORT_CODE_TEMPLATE).read_text())
        variables = {
            'model_import_statement': model_import_stmt,
            'embedder_import_statement': embedder_import_stmt,
            "normalizer_import_statement": normalizer_import_stmt,
            'dataset_file': self.DATASET_FILE,
            'embedder_class': embedder_class,
            'embedder_args': embedder_args_text,
            'model_class': model_class,
            'model_params': model_params_text,
            'fig_width': self.canvas_width / 100,
            'fig_height': self.canvas_height / 100,
            'dpi': 100,
            'num_dots': self.num_dots,
            'x_min': x_min,
            'x_max': x_max,
            'y_min': y_min,
            'y_max': y_max,
            "normalize": self.normalizer_class is not None,
            "normalizer_class": self.normalizer_class.__name__ if self.normalizer_class is not None else "",
            "jitter": self.jitter_std > 0,
            "jitter_scale": self.jitter_std,
        }

        rendered_code = template.render(variables)
        rendered_code = black.format_str(rendered_code, mode=black.FileMode())
        Path(self.CODE_FILE).write_text(rendered_code)
        logger.info(f"{self.CODE_FILE} updated")
        return self.CODE_FILE

    def update_figure_extension(self, ext):
        self.figure_extension = ext
        print('updated figure extension:', self.figure_extension)

    def save_figure(self):
        self.plot(decision_boundary=True, save_figure=True)
        return f"{self.FIGURE_BASENAME}{self.figure_extension}"

    def update_args(self, model_args):
        self.model_args = model_args
        print('updated model_args:', self.model_args)

    def update_embedder(self, embedder):
        self.embedder_class = self.embedders[embedder]
        print('updated Embedder:', self.embedder_class)
        return self.plot()

    def update_embedder_args(self, embedder_args):
        self.embedder_args = embedder_args
        print('updated Embedder args:', self.embedder_args)
        return self.plot()

    def update_normalizer(self, normalizer):
        self.normalizer_class = self.normalizers[normalizer]
        print('updated Normalizer:', self.normalizer_class)
        
        data_table = gr.Dataframe(
            value=self.dataset[:100].round(self.precision), 
            visible=True,
            headers=list(self.dataset.columns),
        )

        return self.plot(), data_table

    def update_jittering(self, jitter_std):
        try:
            self.jitter_std = float(jitter_std)
        except ValueError:
            self.jitter_std = 0
        print('updated Jittering std:', self.jitter_std)
        return self.plot()

    def handle_dataset_radio(self, type):
        if type == 'Draw2D':
            self.custom_selected = True
            self.dataset_type = "Draw2D"
            new_fields = gr.File(visible=False), gr.Dropdown(visible=False, value="None"), gr.Dropdown(visible=False, value="None"), gr.Dropdown(visible=False, value=0), gr.Dropdown(visible=False, value="CoordinateProjection2d"), gr.Textbox(visible=False), gr.Textbox(visible=True), gr.Button(visible=True), gr.Button(visible=False)
        elif type == 'Upload':
            self.dataset_type = "Upload"
            self.custom_selected = False
            new_fields = gr.File(visible=True), gr.Dropdown(visible=False, value="None"), gr.Dropdown(visible=True, value="None"), gr.Dropdown(visible=True, value=0), gr.Dropdown(visible=True, value="CoordinateProjection2d"), gr.Textbox(visible=True), gr.Textbox(visible=False), gr.Button(visible=False), gr.Button(visible=False)
        elif type == 'sklearn':
            self.dataset_type = "sklearn"
            self.custom_selected = False
            new_fields = gr.File(visible=False), gr.Dropdown(visible=True, value="None"), gr.Dropdown(visible=True, value="None"), gr.Dropdown(visible=True, value=0), gr.Dropdown(visible=True, value="CoordinateProjection2d"), gr.Textbox(visible=True), gr.Textbox(visible=False), gr.Button(visible=False), gr.Button(visible=False)
        else:
            # TODO: better error handling
            print('Error - unknown dataset type:', type)

        self._reset_data_processing_and_embedding()
        plot, data_table = self.clear()

        return plot, data_table, *new_fields

    def load_local_data_and_plot(self, filename):
        if filename is not None:
            self.dataset = read(filename)
            self.dataset.target = self.dataset.target.astype(str)
            self.dataset_type = 'Upload'
            logger.info(f'Loaded dataset from {filename}') 

        vis = self.plot()
        #data_html = self.dataset.to_html(classes="report-table", float_format="%.2f")
        # TODO: need to make it explicit that this only shows first 100 points 
        data_table = gr.Dataframe(
            value=self.dataset[:100].round(self.precision), 
            visible=True,
            headers=list(self.dataset.columns)
        )

        return vis, data_table

    def load_sklearn_data_and_plot(self, datasetname):
        if datasetname is not None and datasetname != "None":
            dataset = self.dataloaders[datasetname]()
            X = dataset.data
            y = dataset.target
            if hasattr(dataset, 'feature_names'):
                feature_names = dataset.feature_names
            else:
                feature_names = ['F{%d}' % i for i in range(len(X[0]))]

            if hasattr(dataset, 'target_names'):
                labels = dataset.target_names
            else:
                labels = ['C{%d}' % i for i in range(len(np.unique(y)))]

            y = np.array([labels[i] for i in y])

            self.dataset = pd.DataFrame(X, columns=feature_names)
            self.dataset['target'] = y.astype(str)
            self.dataset_type = 'sklearn'
            logger.info(f'Loaded dataset {datasetname}') 

        vis = self.plot()
        #data_html = self.dataset.to_html(classes="report-table", float_format="%.2f")
        # TODO: need to make it explicit that this only shows first 100 points 
        data_table = gr.Dataframe(
            value=self.dataset[:100].round(self.precision), visible=True,
            headers=list(self.dataset.columns)
        )
        return vis, data_table

    def launch(self):
        # build the Gradio interface
        with gr.Blocks(css=self.css) as demo:
            # app title
            gr.HTML("<div style='text-align:left; font-size:40px; font-weight: bold;'>Interactive Decision Boundary Visualizer</div>")

            # GUI elements and layout 
            with gr.Row():
                with gr.Column(scale=2):
                    self.data_image = gr.Image(
                        value=self.plot(), 
                        container=True, 
                        show_share_button=False, 
                        show_fullscreen_button=False,
                        show_download_button=False,
                        show_label=False,
                    )

                with gr.Column(scale=1): 
                    with gr.Tab("Dataset"):
                        dataset_radio = gr.Radio(
                            ["Draw2D", "Upload", "sklearn"], 
                            value="Draw2D", 
                            label="Dataset type", 
                            elem_id="rowheight",
                        )

                        # upload data
                        file_chooser = gr.File(label="Choose a file", visible=False, elem_classes="file-chooser")
                        self.file_chooser = file_chooser

                        # sklearn data dropdown menu 
                        sklearn_data_selector = gr.Dropdown(
                            choices=self.dataloaders, 
                            label='Select dataset',
                            value='None',
                            visible=False,
                            allow_custom_value=True,
                        )
                        self.sklearn_data_selector = sklearn_data_selector

                        # normalization
                        normalizer_selector = gr.Dropdown(
                            choices=self.normalizers,
                            label='Select normalizer',
                            value='None',
                            visible=False,
                        )
                        self.normalizer_selector = normalizer_selector

                        # jittering
                        jittering_textbox = gr.Textbox(label="Set jittering std", value="0", visible=False)
                        self.jittering_textbox = jittering_textbox

                        # embedder
                        embedder_selector = gr.Dropdown(
                            choices=self.embedders, 
                            label='Select embedder (only for plotting)', 
                            value='CoordinateProjection2d',
                            visible=False,
                            allow_custom_value=True,
                        )
                        self.embedder_selector = embedder_selector

                        embedder_args_textbox = gr.Textbox(label="Embedder arguments", visible=False)
                        self.embedder_args_textbox = embedder_args_textbox

                        # custom data
                        label = gr.Radio(["Gray", "Orange", "Blue"], value="Gray", label="Choose point label", visible=True, elem_id="rowheight")
                        self.label = label
      
                        with gr.Row():
                            btn_clear = gr.Button("Clear", visible=True, elem_id="my-button")
                            self.btn_clear = btn_clear

                            btn_save = gr.Button("Save", visible=False, elem_id="my-button")
                            self.btn_save = btn_save
                        
                        #data_html = gr.HTML(visible=True)
                        data_table = gr.Dataframe(visible=False)

                    # classifier selector
                    with gr.Tab("Classifier"):
                        # specify model
                        model_selector = gr.Dropdown(choices=self.classifiers,
                                               #label='', 
                                               #value='Select classifier',
                                               label='Select Classifier', 
                                               value='LinearSVC',
                                               allow_custom_value=True)
                        self.model_selector = model_selector
                   
                        # specify arguments
                        args_textbox = gr.Textbox(label="Classifier arguments")
                        self.args_textbox = args_textbox

                        model_selector.change(fn=self.update_model, inputs=model_selector, outputs=args_textbox)
                        
                        btn_train = gr.Button("Train Model")

                        classification_summary = gr.HTML(visible=False)
                       
                    with gr.Tab("Export"):
                        # use hidden download button to generate files on the fly
                        # https://github.com/gradio-app/gradio/issues/9230#issuecomment-2323771634

                        btn_export_data = gr.Button("Data")
                        btn_export_data_hidden = gr.DownloadButton(label="You should not see this", elem_id="btn_export_data_hidden", elem_classes="hidden-button")
                        
                        btn_export_model = gr.Button('Model')
                        btn_export_model_hidden = gr.DownloadButton(label="You should not see this", elem_id="btn_export_model_hidden", elem_classes="hidden-button")

                        btn_export_code = gr.Button('Code')
                        btn_export_code_hidden = gr.DownloadButton(label="You should not see this", elem_id="btn_export_code_hidden", elem_classes="hidden-button")

                        with gr.Row():
                            btn_export_figure = gr.Button('Figure')
                            btn_export_figure_hidden = gr.DownloadButton(label="You should not see this", elem_id="btn_export_figure_hidden", elem_classes="hidden-button")
                            figure_extension_selector = gr.Dropdown(choices=['.svg', '.pdf', '.png', '.jpeg'], label="File extension", value=".svg")

                    with gr.Tab("Options"):
                        grid_resolution_slider = gr.Slider(minimum=100, maximum=1000, value=200, step=10, label="Decision boundary grid resolution")
                        # image_dpi_slider = gr.Slider(minimum=100, maximum=1000, value=100, step=10, label="Image DPI")
                        cmap_textbox = gr.Textbox(label="Colormap")
                        precision_slider = gr.Slider(minimum=0, maximum=20, value=2, step=1, label="# decimal place in datatable")
                        marker_size_slider = gr.Slider(minimum=0, maximum=200, value=100, step=5, label="Marker size")

                    with gr.Tab("Usage"):
                        gr.Markdown(''.join(open('usage.md', 'r').readlines()))


            # event handlers for GUI elements
            self.data_image.select(self.add_point, inputs=label,
                                   outputs=(self.data_image, data_table))

            dataset_radio.change(
                fn=self.handle_dataset_radio,
                inputs=dataset_radio, 
                outputs=(
                    self.data_image, data_table, file_chooser, sklearn_data_selector, normalizer_selector, jittering_textbox, embedder_selector, embedder_args_textbox, label, btn_clear, btn_save
                ),
            )

            # events for custom dataset
            btn_clear.click(fn=self.clear, outputs=(self.data_image, data_table))
            btn_save.click(fn=self.save)

            # events for local dataset
            file_chooser.change(fn=self.load_local_data_and_plot,
                                inputs=file_chooser, 
                                outputs=(self.data_image, data_table))

            # events for sklearn dataset
            sklearn_data_selector.change(fn=self.load_sklearn_data_and_plot,
                                         inputs=sklearn_data_selector,
                                         outputs=(self.data_image, data_table))

            embedder_selector.change(fn=self.update_embedder,
                                     inputs=embedder_selector,
                                     outputs=self.data_image)

            embedder_args_textbox.change(
                fn=self.update_embedder_args,
                inputs=embedder_args_textbox,
                outputs=self.data_image,
            )

            normalizer_selector.change(
                fn=self.update_normalizer,
                inputs=normalizer_selector,
                outputs=(self.data_image, data_table),
            )

            jittering_textbox.change(
                fn=self.update_jittering,
                inputs=jittering_textbox,
                outputs=self.data_image,
            )

            btn_train.click(fn=self.update_args, inputs=args_textbox)
            btn_train.click(fn=self.train, outputs=(self.data_image, classification_summary, classification_summary))

            # events for export
            # create files on the fly using hidden download buttons
            # https://github.com/gradio-app/gradio/issues/9230#issuecomment-2323771634
            btn_export_data.click(
                fn=self.save_data, 
                inputs=None, 
                outputs=[btn_export_data_hidden]
            ).then(
                fn=None, inputs=None, outputs=None, js="() => document.querySelector('#btn_export_data_hidden').click()"
            )
            btn_export_model.click(
                fn=self.save_model, 
                inputs=None, 
                outputs=[btn_export_model_hidden]
            ).then(
                fn=None, inputs=None, outputs=None, js="() => document.querySelector('#btn_export_model_hidden').click()"
            )
            btn_export_code.click(
                fn=self.save_code, 
                inputs=None, 
                outputs=[btn_export_code_hidden]
            ).then(
                fn=None, inputs=None, outputs=None, js="() => document.querySelector('#btn_export_code_hidden').click()"
            )
            btn_export_figure.click(
                fn=self.save_figure,
                inputs=None,
                outputs=[btn_export_figure_hidden]
            ).then(
                fn=None, inputs=None, outputs=None, js="() => document.querySelector('#btn_export_figure_hidden').click()"
            )
            figure_extension_selector.change(self.update_figure_extension, inputs=figure_extension_selector)

            # events for options
            grid_resolution_slider.change(self.update_resolution, inputs=grid_resolution_slider, outputs=self.data_image)
            # image_dpi_slider.change(self.update_dpi, inputs=image_dpi_slider, outputs=self.data_image)
            cmap_textbox.submit(self.update_cmap, inputs=cmap_textbox, outputs=self.data_image)
            precision_slider.change(self.update_precision, inputs=precision_slider, outputs=data_table)
            marker_size_slider.change(self.update_marker_size, inputs=marker_size_slider, outputs=self.data_image)

        demo.launch()

visualizer = InteractiveDecisionBoundary(width=1200, height=900)
visualizer.launch()