app.py

import gradio as gr
import numpy as np
import pandas as pd
import matplotlib
import matplotlib.pyplot as plt

from datasets import load_dataset
from sklearn.ensemble import GradientBoostingClassifier
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score, confusion_matrix

matplotlib.use('Agg')

################################################################################
# SUGGESTED_DATASETS: These must actually exist on huggingface.co/datasets
#
# "scikit-learn/iris" -> A small, classic Iris dataset with a "train" split
# "uci/wine"          -> Another small dataset with a "train" split
# "SKIP/ENTER_CUSTOM" -> Placeholder to let the user enter a custom dataset ID
################################################################################
SUGGESTED_DATASETS = [
    "scikit-learn/iris",
    "uci/wine",
    "SKIP/ENTER_CUSTOM"
]

def update_columns(dataset_id, custom_dataset_id):
    """
    After the user chooses a dataset from the dropdown or enters their own,
    this function loads the dataset's "train" split, converts it to a DataFrame,
    and returns the columns. These columns are used to populate the Label and
    Feature selectors in the UI.
    """
    if dataset_id != "SKIP/ENTER_CUSTOM":
        final_id = dataset_id
    else:
        final_id = custom_dataset_id.strip()

    try:
        ds = load_dataset(final_id, split="train")
        df = pd.DataFrame(ds)
        cols = df.columns.tolist()

        message = (
            f"**Loaded dataset**: `{final_id}`\n\n"
            f"**Columns found**: {cols}"
        )
        return (
            gr.update(choices=cols, value=None),   # label_col dropdown
            gr.update(choices=cols, value=[]),     # feature_cols checkbox group
            message
        )
    except Exception as e:
        err_msg = f"**Error loading** `{final_id}`: {e}"
        return (
            gr.update(choices=[], value=None),
            gr.update(choices=[], value=[]),
            err_msg
        )

def train_model(dataset_id, custom_dataset_id, label_column, feature_columns,
                learning_rate, n_estimators, max_depth, test_size):
    """
    1. Decide which dataset ID to load (from dropdown or custom).
    2. Load that dataset's 'train' split, turn into DataFrame, extract X (features) and y (label).
    3. Train a GradientBoostingClassifier on X_train, y_train.
    4. Compute accuracy and confusion matrix on X_test, y_test.
    5. Plot and return feature importances + confusion matrix heatmap + textual summary.
    """
    # Resolve final dataset ID
    if dataset_id != "SKIP/ENTER_CUSTOM":
        final_id = dataset_id
    else:
        final_id = custom_dataset_id.strip()

    # Load dataset -> df
    ds = load_dataset(final_id, split="train")
    df = pd.DataFrame(ds)

    # Validate columns
    if label_column not in df.columns:
        raise ValueError(f"Label column '{label_column}' not found in dataset columns.")
    for fc in feature_columns:
        if fc not in df.columns:
            raise ValueError(f"Feature column '{fc}' not found in dataset columns.")

    # Convert to NumPy arrays
    X = df[feature_columns].values
    y = df[label_column].values

    # Train/test split
    X_train, X_test, y_train, y_test = train_test_split(
        X, y, test_size=test_size, random_state=42
    )

    # Instantiate and train GradientBoostingClassifier
    clf = GradientBoostingClassifier(
        learning_rate=learning_rate,
        n_estimators=int(n_estimators),
        max_depth=int(max_depth),
        random_state=42
    )
    clf.fit(X_train, y_train)

    # Evaluate
    y_pred = clf.predict(X_test)
    accuracy = accuracy_score(y_test, y_pred)
    cm = confusion_matrix(y_test, y_pred)

    # Create Matplotlib figure with feature importances + confusion matrix
    fig, axs = plt.subplots(1, 2, figsize=(10, 4))

    # Subplot 1: Feature Importances
    importances = clf.feature_importances_
    axs[0].barh(range(len(feature_columns)), importances, color='skyblue')
    axs[0].set_yticks(range(len(feature_columns)))
    axs[0].set_yticklabels(feature_columns)
    axs[0].set_xlabel("Importance")
    axs[0].set_title("Feature Importances")

    # Subplot 2: Confusion Matrix Heatmap
    im = axs[1].imshow(cm, interpolation='nearest', cmap=plt.cm.Blues)
    axs[1].set_title("Confusion Matrix")
    plt.colorbar(im, ax=axs[1])
    axs[1].set_xlabel("Predicted")
    axs[1].set_ylabel("True")

    # Optionally annotate each cell with numeric counts
    thresh = cm.max() / 2.0
    for i in range(cm.shape[0]):
        for j in range(cm.shape[1]):
            color = "white" if cm[i, j] > thresh else "black"
            axs[1].text(j, i, str(cm[i, j]), ha="center", va="center", color=color)

    plt.tight_layout()

    # Textual summary
    text_summary = (
        f"**Dataset used**: `{final_id}`\n\n"
        f"**Label column**: `{label_column}`\n\n"
        f"**Feature columns**: `{feature_columns}`\n\n"
        f"**Accuracy**: {accuracy:.3f}\n\n"
    )

    return text_summary, fig

###############################################################################
# Gradio UI
###############################################################################
with gr.Blocks() as demo:

    # High-level title and description
    gr.Markdown(
        """
        # Introduction to Gradient Boosting

        This Space demonstrates how to train a [GradientBoostingClassifier](https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.GradientBoostingClassifier.html#gradientboostingclassifier) from **scikit-learn** on **tabular datasets** hosted on the [Hugging Face Hub](https://huggingface.co/datasets).

        Gradient Boosting is an ensemble machine learning technique that combines many weak learners (usually small decision trees) in an iterative, stage-wise fashion to create a stronger overall model. 
        In each step, the algorithm fits a new weak learner to the current errors of the combined ensemble, effectively allowing the model to focus on the hardest-to-predict data points. 
        By repeatedly adding these specialized trees, Gradient Boosting can capture complex patterns and deliver high predictive accuracy, especially on tabular data.

        **Put simply, Gradient Boosting makes a big deal out of small anomolies!**

        **Purpose**:
        - Easily explore hyperparameters (_learning_rate, n_estimators, max_depth_) and quickly train an ML model on real data.
        - Visualise model performance via confusion matrix heatmap and a feature importance plot.

        **Notes**:
        - The dataset must have a **"train"** split with tabular columns (i.e., no nested structures).
        - Large datasets may take time to download/train.
        - The confusion matrix helps you see how predictions compare to ground-truth labels. The diagonal cells show correct predictions; off-diagonal cells indicate misclassifications.
        - The feature importance plot shows which features the model relies on the most for its predictions.

        ---

        **Usage**:
        1. Select one of the suggested datasets from the dropdown _or_ enter any valid dataset from the [Hugging Face Hub](https://huggingface.co/datasets).
        2. Click **Load Columns** to retrieve the column names from the dataset's **train** split.
        3. Choose exactly _one_ **Label column** (the target) and one or more **Feature columns** (the inputs).
        4. Adjust hyperparameters (learning_rate, n_estimators, max_depth, test_size).
        5. Click **Train & Evaluate** to train a Gradient Boosting model and see its accuracy, feature importances, and confusion matrix.

        You are now a machine learning engineer, congratulations 🤗 
        
        ---
        """
    )

    with gr.Row():
        dataset_dropdown = gr.Dropdown(
            label="Choose suggested dataset",
            choices=SUGGESTED_DATASETS,
            value=SUGGESTED_DATASETS[0]
        )
        custom_dataset_id = gr.Textbox(
            label="Or enter a custom dataset ID",
            placeholder="e.g. user/my_custom_dataset"
        )

    load_cols_btn = gr.Button("Load Columns")
    load_cols_info = gr.Markdown()

    with gr.Row():
        label_col = gr.Dropdown(choices=[], label="Label column (choose 1)")
        feature_cols = gr.CheckboxGroup(choices=[], label="Feature columns (choose 1 or more)")

    # Model Hyperparameters
    learning_rate_slider = gr.Slider(
        minimum=0.01, maximum=1.0, value=0.1, step=0.01, 
        label="learning_rate"
    )
    n_estimators_slider = gr.Slider(
        minimum=50, maximum=300, value=100, step=50, 
        label="n_estimators"
    )
    max_depth_slider = gr.Slider(
        minimum=1, maximum=10, value=3, step=1, 
        label="max_depth"
    )
    test_size_slider = gr.Slider(
        minimum=0.1, maximum=0.9, value=0.3, step=0.1, 
        label="test_size fraction (0.1-0.9)"
    )

    train_button = gr.Button("Train & Evaluate")

    output_text = gr.Markdown()
    output_plot = gr.Plot()

    # Link the "Load Columns" button -> update_columns function
    load_cols_btn.click(
        fn=update_columns,
        inputs=[dataset_dropdown, custom_dataset_id],
        outputs=[label_col, feature_cols, load_cols_info],
    )

    # Link "Train & Evaluate" -> train_model function
    train_button.click(
        fn=train_model,
        inputs=[
            dataset_dropdown,
            custom_dataset_id,
            label_col,
            feature_cols,
            learning_rate_slider,
            n_estimators_slider,
            max_depth_slider,
            test_size_slider
        ],
        outputs=[output_text, output_plot],
    )

demo.launch()