app.py

import gradio as gr
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import seaborn as sns
import logging
import requests
from typing import List, Tuple, Dict, Optional
from datasets import load_dataset
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import accuracy_score, confusion_matrix, classification_report
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import LabelEncoder

# Logging configuration
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

# Dataset repository configuration
REPO_CONFIG = {
    "Core (Clean)": {
        "repo": "QSBench/QSBench-Core-v1.0.0-demo",
        "meta_url": "https://huggingface.co/datasets/QSBench/QSBench-Core-v1.0.0-demo/raw/metadata/meta/meta.json",
        "report_url": "https://huggingface.co/datasets/QSBench/QSBench-Core-v1.0.0-demo/raw/metadata/meta/report.json"
    },
    "Depolarizing Noise": {
        "repo": "QSBench/QSBench-Depolarizing-Demo-v1.0.0",
        "meta_url": "https://huggingface.co/datasets/QSBench/QSBench-Depolarizing-Demo-v1.0.0/raw/meta/meta/meta.json",
        "report_url": "https://huggingface.co/datasets/QSBench/QSBench-Depolarizing-Demo-v1.0.0/raw/meta/meta/report.json"
    },
    "Amplitude Damping": {
        "repo": "QSBench/QSBench-Amplitude-v1.0.0-demo",
        "meta_url": "https://huggingface.co/datasets/QSBench/QSBench-Amplitude-v1.0.0-demo/raw/meta/meta/meta.json",
        "report_url": "https://huggingface.co/datasets/QSBench/QSBench-Amplitude-v1.0.0-demo/raw/meta/meta/report.json"
    },
    "Transpilation (10q)": {
        "repo": "QSBench/QSBench-Transpilation-v1.0.0-demo",
        "meta_url": "https://huggingface.co/datasets/QSBench/QSBench-Transpilation-v1.0.0-demo/raw/meta/meta/meta.json",
        "report_url": "https://huggingface.co/datasets/QSBench/QSBench-Transpilation-v1.0.0-demo/raw/meta/meta/report.json"
    }
}

# Define non-feature columns to exclude from training
NON_FEATURE_COLS = {
    "sample_id", "sample_seed", "circuit_hash", "split", "circuit_qasm", 
    "qasm_raw", "qasm_transpiled", "circuit_type_resolved", "circuit_type_requested",
    "noise_type", "noise_prob", "observable_bases", "observable_mode", "backend_device", 
    "precision_mode", "circuit_signature", "entanglement", "shots", "gpu_requested", "gpu_available"
}

_ASSET_CACHE = {}

def load_all_assets(key: str) -> Dict:
    """
    Fetch and cache dataset and metadata from Hugging Face.
    """
    if key not in _ASSET_CACHE:
        logger.info(f"Fetching {key} assets...")
        ds = load_dataset(REPO_CONFIG[key]["repo"])
        meta = requests.get(REPO_CONFIG[key]["meta_url"]).json()
        report = requests.get(REPO_CONFIG[key]["report_url"]).json()
        _ASSET_CACHE[key] = {"df": pd.DataFrame(ds["train"]), "meta": meta, "report": report}
    return _ASSET_CACHE[key]

def load_guide_content() -> str:
    """
    Load Markdown content for the Methodology/Guide tab.
    """
    try:
        with open("GUIDE.md", "r", encoding="utf-8") as f:
            return f.read()
    except FileNotFoundError:
        return "### ⚠️ GUIDE.md not found."

def sync_ml_metrics(ds_name: str) -> gr.update:
    """
    Filter and return available numerical features for the selected dataset.
    """
    assets = load_all_assets(ds_name)
    df = assets["df"]
    numeric_cols = df.select_dtypes(include=[np.number]).columns.tolist()
    valid_features = [c for c in numeric_cols if c not in NON_FEATURE_COLS and not any(p in c for p in ["ideal_", "noisy_", "error_"])]
    defaults = [f for f in ["gate_entropy", "meyer_wallach", "adjacency", "depth", "cx_count"] if f in valid_features]
    return gr.update(choices=valid_features, value=defaults)

def train_classifier(ds_name: str, features: List[str]) -> Tuple[Optional[plt.Figure], str]:
    """
    Perform multi-class classification on circuit families and return metrics/plots.
    """
    if not features: 
        return None, "### ❌ Error: No features selected."
    
    assets = load_all_assets(ds_name)
    df = assets["df"]
    
    # Target column selection fallback logic
    target_col = 'circuit_type_resolved' if 'circuit_type_resolved' in df.columns else 'circuit_type_requested'
    
    # Data preprocessing and cleaning
    train_df = df.dropna(subset=features + [target_col])
    if 'mixed' in train_df[target_col].unique() and len(train_df[target_col].unique()) > 1:
        train_df = train_df[train_df[target_col] != 'mixed']

    X = train_df[features]
    y = train_df[target_col]

    if len(y.unique()) < 2:
        return None, f"### ❌ Error: Dataset contains insufficient classes for training ({y.unique()})."

    # Label encoding and dataset splitting
    le = LabelEncoder()
    y_encoded = le.fit_transform(y)
    
    try:
        X_train, X_test, y_train, y_test = train_test_split(X, y_encoded, test_size=0.2, random_state=42, stratify=y_encoded)
    except (ValueError, TypeError):
        X_train, X_test, y_train, y_test = train_test_split(X, y_encoded, test_size=0.2, random_state=42)

    # Model initialization and training
    clf = RandomForestClassifier(n_estimators=100, max_depth=12, n_jobs=-1, random_state=42)
    clf.fit(X_train, y_train)
    preds = clf.predict(X_test)

    # Visualization generation
    sns.set_theme(style="whitegrid")
    fig, axes = plt.subplots(1, 2, figsize=(20, 8))

    # Confusion Matrix Plot
    cm = confusion_matrix(y_test, preds)
    sns.heatmap(cm, annot=True, fmt='d', cmap='viridis', xticklabels=le.classes_, yticklabels=le.classes_, ax=axes[0], cbar=False)
    axes[0].set_title(f"Confusion Matrix (Accuracy: {accuracy_score(y_test, preds):.2%})")

    # Feature Importance Plot
    importances = clf.feature_importances_
    idx = np.argsort(importances)[-10:]
    axes[1].barh([features[i] for i in idx], importances[idx], color='#2ecc71')
    axes[1].set_title("Top-10 Predictive Features")

    plt.tight_layout()
    
    # Performance metrics string generation
    cls_report = classification_report(y_test, preds, target_names=le.classes_, output_dict=False)
    results_md = f"### 🏆 Classification Results\n**Target:** `{target_col}`\n**Accuracy:** {accuracy_score(y_test, preds):.2%}\n\n**Metrics:**\n```text\n{cls_report}\n```"
    
    return fig, results_md

def update_explorer(ds_name: str, split_name: str) -> Tuple[gr.update, pd.DataFrame, str, str, str]:
    """
    Refresh the Explorer view based on dataset and split selection.
    """
    assets = load_all_assets(ds_name)
    df = assets["df"]
    splits = df["split"].unique().tolist() if "split" in df.columns else ["train"]
    
    if split_name not in splits:
        split_name = splits[0]
        
    filtered = df[df["split"] == split_name] if "split" in df.columns else df
    display_df = filtered.head(10)
    
    raw_qasm = display_df["qasm_raw"].iloc[0] if "qasm_raw" in display_df.columns and not display_df.empty else "// N/A"
    transpiled_qasm = display_df["qasm_transpiled"].iloc[0] if "qasm_transpiled" in display_df.columns and not display_df.empty else "// N/A"
    
    return (
        gr.update(choices=splits, value=split_name), 
        display_df, 
        raw_qasm, 
        transpiled_qasm, 
        f"### 📋 {ds_name} Explorer"
    )

# Gradio interface definition
with gr.Blocks(theme=gr.themes.Soft(), title="QSBench Classifier") as demo:
    gr.Markdown("# 🌌 QSBench: Circuit Family Classifier")
    
    with gr.Tabs():
        with gr.TabItem("🔎 Explorer"):
            meta_label = gr.Markdown("### Initializing...")
            with gr.Row():
                ds_dropdown = gr.Dropdown(list(REPO_CONFIG.keys()), value="Core (Clean)", label="Dataset Type")
                split_dropdown = gr.Dropdown(["train"], value="train", label="Split")
            explorer_df = gr.Dataframe(interactive=False)
            with gr.Row():
                raw_qasm_code = gr.Code(label="Logical QASM", language="python")
                tr_qasm_code = gr.Code(label="Transpiled QASM", language="python")

        with gr.TabItem("🧠 Classification"):
            with gr.Row():
                with gr.Column(scale=1):
                    ml_ds_dropdown = gr.Dropdown(list(REPO_CONFIG.keys()), value="Core (Clean)", label="Noise Environment")
                    ml_feature_checks = gr.CheckboxGroup(label="Input Metrics", choices=[])
                    run_btn = gr.Button("Train & Evaluate", variant="primary")
                with gr.Column(scale=2):
                    plot_output = gr.Plot()
                    results_output = gr.Markdown()

        with gr.TabItem("📖 Guide"):
            gr.Markdown(load_guide_content())

    gr.Markdown("--- \n ### 🔗 [Website](https://qsbench.github.io) | [Hugging Face](https://huggingface.co/QSBench) | [GitHub](https://github.com/QSBench)")

    # UI Event bindings
    ds_dropdown.change(update_explorer, [ds_dropdown, split_dropdown], [split_dropdown, explorer_df, raw_qasm_code, tr_qasm_code, meta_label])
    split_dropdown.change(update_explorer, [ds_dropdown, split_dropdown], [split_dropdown, explorer_df, raw_qasm_code, tr_qasm_code, meta_label])
    ml_ds_dropdown.change(sync_ml_metrics, [ml_ds_dropdown], [ml_feature_checks])
    run_btn.click(train_classifier, [ml_ds_dropdown, ml_feature_checks], [plot_output, results_output])
    
    # Application startup triggers
    demo.load(update_explorer, [ds_dropdown, split_dropdown], [split_dropdown, explorer_df, raw_qasm_code, tr_qasm_code, meta_label])
    demo.load(sync_ml_metrics, [ml_ds_dropdown], [ml_feature_checks])

if __name__ == "__main__":
    demo.launch()