app.py

"""
Dota 2 Edge — Multi-Model Ensemble Prediction API v3
=====================================================
HuggingFace Space serving V1, V2, and V3 (temporal CV) models.
V3 = production-grade, leak-free temporal cross-validation.
"""

import gradio as gr
import pickle
import numpy as np
import json
import os
from pathlib import Path

# ── Model directories ──
MODEL_DIR_V1 = Path("models")
MODEL_DIR_V2 = Path("models_v2")
MODEL_DIR_V3 = Path("models_v3")

# ── V3 Moments (production) ──
V3_MOMENTS = {
    'm5':  '@5 minutes',
    'm8':  '@8 minutes',
    'm10': '@10 minutes',
    'm15': '@15 minutes',
    'm20': '@20 minutes',
    'm25': '@25 minutes',
    'm30': '@30 minutes',
}

# ── V3 Feature columns ──
V3_FEATURES = [
    'gold_delta', 'xp_delta', 'lh_delta', 'dn_delta',
    'rad_gpm', 'dire_gpm', 'gpm_delta',
    'rad_xpm', 'dire_xpm', 'xpm_delta',
    'carry_gap', 'support_gap',
    'cs_eff_delta', 'deny_delta',
    'gold_log_ratio', 'xp_log_ratio', 'gold_lead_pct',
    'gold_momentum', 'xp_momentum',
]

# ── V1 Moments (legacy) ──
V1_MOMENTS = {
    'm0':  'Draft (pre-game)',
    'm1':  '@10 minutes',
    'm15': '@15 minutes',
    'm2':  '@20 minutes',
    'm3':  '@30 minutes',
}

# ── Load V3 bundles (production) ──
v3_bundles = {}
for mk, label in V3_MOMENTS.items():
    path = MODEL_DIR_V3 / f"temporal_v3_{mk}.pkl"
    if path.exists():
        with open(path, 'rb') as f:
            v3_bundles[mk] = pickle.load(f)

# ── Load V1 bundles (legacy) ──
v1_bundles = {}
for mk in V1_MOMENTS:
    path = MODEL_DIR_V1 / f"ensemble_{mk}.pkl"
    if path.exists():
        with open(path, 'rb') as f:
            v1_bundles[mk] = pickle.load(f)

# ── Load V2 bundles ──
v2_bundles = {}
V2_MOMENTS = {'m0_5': '@5min', 'm1_8': '@8min', 'm1': '@10min', 'm15': '@15min',
              'm2': '@20min', 'm25': '@25min', 'm3': '@30min'}
for mk in V2_MOMENTS:
    path = MODEL_DIR_V2 / f"ensemble_{mk}.pkl"
    if path.exists():
        with open(path, 'rb') as f:
            v2_bundles[mk] = pickle.load(f)

# Load summaries
summaries = {}
for name, path in [('v1', MODEL_DIR_V1 / 'ensemble_summary.json'),
                    ('v2', MODEL_DIR_V2 / 'ensemble_v2_summary.json'),
                    ('v3', MODEL_DIR_V3 / 'ensemble_v3_summary.json')]:
    if path.exists():
        with open(path) as f:
            summaries[name] = json.load(f)


def predict_v3(minute_key: str, features_json: str) -> dict:
    """
    V3 production prediction using temporal CV models.
    """
    if minute_key not in v3_bundles:
        return {"error": f"Moment '{minute_key}' not loaded. Available: {list(v3_bundles.keys())}"}

    bundle = v3_bundles[minute_key]
    feature_cols = bundle['feature_cols']

    try:
        features = json.loads(features_json) if isinstance(features_json, str) else features_json
    except json.JSONDecodeError as e:
        return {"error": f"Invalid JSON: {e}"}

    X = np.array([[features.get(col, 0.0) for col in feature_cols]])
    X_scaled = bundle['scaler'].transform(X)

    probs = {}
    for name in bundle['models']:
        model = bundle['models'][name]
        if name in ('logreg', 'mlp'):
            p = model.predict_proba(X_scaled)[:, 1][0]
        else:
            p = model.predict_proba(X)[:, 1][0]
        probs[name] = round(float(p), 4)

    # Stacking
    stack_input = np.array([[probs[n] for n in bundle['models']]])
    stacking = round(float(bundle['meta_learner'].predict_proba(stack_input)[:, 1][0]), 4)

    # Best individual = LogReg (proven by temporal CV)
    logreg_prob = probs.get('logreg', stacking)

    return {
        "version": "v3_temporal_cv",
        "moment": minute_key,
        "moment_name": V3_MOMENTS[minute_key],
        "n_features": len(feature_cols),
        "individual_models": probs,
        "stacking_ensemble": stacking,
        "logreg_prob": logreg_prob,
        "recommended_prob": logreg_prob,  # LogReg is best with temporal CV
        "prediction": "Radiant" if logreg_prob > 0.5 else "Dire",
        "confidence": round(abs(logreg_prob - 0.5) * 2, 4),
        "holdout_auc": bundle['results'].get('logreg', {}).get('auc', 'N/A'),
    }


def predict_v1(moment: str, features_json: str) -> dict:
    """V1 legacy prediction."""
    if moment not in v1_bundles:
        return {"error": f"V1 moment '{moment}' not loaded. Available: {list(v1_bundles.keys())}"}

    bundle = v1_bundles[moment]
    feature_cols = bundle['feature_cols']

    try:
        features = json.loads(features_json) if isinstance(features_json, str) else features_json
    except json.JSONDecodeError as e:
        return {"error": f"Invalid JSON: {e}"}

    X = np.array([[features.get(col, 0.0) for col in feature_cols]])
    X_scaled = bundle['scaler'].transform(X)

    probs = {}
    for name in bundle['base_model_names']:
        model = bundle['models'][name]
        if name in ('logreg', 'mlp'):
            p = model.predict_proba(X_scaled)[:, 1][0]
        else:
            p = model.predict_proba(X)[:, 1][0]
        probs[name] = round(float(p), 4)

    stack_input = np.array([[probs[n] for n in bundle['base_model_names']]])
    stacking = round(float(bundle['meta_learner'].predict_proba(stack_input)[:, 1][0]), 4)

    return {
        "version": "v1",
        "moment": moment,
        "moment_name": V1_MOMENTS.get(moment, moment),
        "individual_models": probs,
        "stacking_ensemble": stacking,
        "prediction": "Radiant" if stacking > 0.5 else "Dire",
        "confidence": round(abs(stacking - 0.5) * 2, 4),
    }


def get_model_info() -> dict:
    """Get info about all loaded models."""
    info = {"v3_models": {}, "v2_models": {}, "v1_models": {}, "summaries": summaries}

    for mk, bundle in v3_bundles.items():
        info["v3_models"][mk] = {
            "name": V3_MOMENTS[mk],
            "n_features": len(bundle['feature_cols']),
            "n_train": bundle.get('n_train', 'N/A'),
            "n_test": bundle.get('n_test', 'N/A'),
            "feature_cols": bundle['feature_cols'],
            "results": bundle['results'],
            "cutoff_date": bundle.get('cutoff_date', 'N/A'),
            "top5_features": bundle.get('top5_features', []),
        }

    for mk, bundle in v1_bundles.items():
        info["v1_models"][mk] = {
            "name": V1_MOMENTS.get(mk, mk),
            "n_features": bundle.get('n_features', len(bundle['feature_cols'])),
            "results": bundle['results'],
        }

    return info


# ── AUC Reference Table ──
AUC_TABLE = """
| Minute | V3 Holdout AUC | V2 K-Fold AUC | Confidence Level |
|--------|---------------|---------------|-----------------|
| @5min  | 0.652 | 0.661 | Low |
| @8min  | 0.685 | 0.697 | Low-Medium |
| @10min | 0.720 | 0.718 | Medium |
| @15min | 0.775 | 0.770 | Medium-High |
| @20min | 0.830 | 0.836 | High |
| @25min | 0.872 | 0.893 | Very High |
| @30min | 0.906 | 0.905 | Very High |
"""


# ================================================================
# GRADIO INTERFACE
# ================================================================

def gradio_predict_v3(moment, features_text):
    try:
        result = predict_v3(moment, features_text)
        return json.dumps(result, indent=2)
    except Exception as e:
        return json.dumps({"error": str(e)}, indent=2)


def gradio_predict_v1(moment, features_text):
    try:
        result = predict_v1(moment, features_text)
        return json.dumps(result, indent=2)
    except Exception as e:
        return json.dumps({"error": str(e)}, indent=2)


def gradio_info():
    return json.dumps(get_model_info(), indent=2, default=str)


# Example V3 features
example_v3 = {
    "gold_delta": 3500, "xp_delta": 2200, "lh_delta": 35, "dn_delta": 5,
    "rad_gpm": 2200, "dire_gpm": 1850, "gpm_delta": 350,
    "rad_xpm": 2800, "dire_xpm": 2500, "xpm_delta": 300,
    "carry_gap": 800, "support_gap": -200,
    "cs_eff_delta": 0.05, "deny_delta": 3,
    "gold_log_ratio": 0.15, "xp_log_ratio": 0.08, "gold_lead_pct": 0.12,
    "gold_momentum": 100, "xp_momentum": 50,
}

with gr.Blocks(
    title="Dota 2 Edge — Ensemble Predictor v3",
    theme=gr.themes.Base(primary_hue="red", secondary_hue="green"),
) as demo:
    gr.Markdown("""
    # Dota 2 Edge — Ensemble Predictor v3

    Predict professional Dota 2 match outcomes using temporally-validated models.

    **V3 (Production):** Trained with expanding-window temporal CV on 4,068 pro matches.
    No data leakage. LogReg = best individual model.

    **Holdout AUC:** @10min=0.720 | @15min=0.775 | @20min=0.830 | @25min=0.872 | @30min=0.906
    """)

    with gr.Tab("V3 Predict (Production)"):
        with gr.Row():
            v3_moment = gr.Dropdown(
                choices=list(V3_MOMENTS.keys()),
                value="m10",
                label="Minute",
                info="Select game minute"
            )
        v3_features = gr.Textbox(
            label="Features (JSON — 19 temporal features)",
            value=json.dumps(example_v3, indent=2),
            lines=10,
        )
        v3_btn = gr.Button("Predict (V3)", variant="primary")
        v3_output = gr.Textbox(label="Result", lines=15)
        v3_btn.click(gradio_predict_v3, inputs=[v3_moment, v3_features], outputs=v3_output)

        gr.Markdown(AUC_TABLE)

    with gr.Tab("V1 Predict (Legacy)"):
        with gr.Row():
            v1_moment = gr.Dropdown(
                choices=list(V1_MOMENTS.keys()),
                value="m1",
                label="Prediction Moment",
            )
        v1_features = gr.Textbox(
            label="Features (JSON)",
            value=json.dumps({"gold_delta_10": 2500, "xp_delta_10": 1800}, indent=2),
            lines=10,
        )
        v1_btn = gr.Button("Predict (V1)", variant="secondary")
        v1_output = gr.Textbox(label="Result", lines=15)
        v1_btn.click(gradio_predict_v1, inputs=[v1_moment, v1_features], outputs=v1_output)

    with gr.Tab("Model Info"):
        info_btn = gr.Button("Load Model Info")
        info_output = gr.Textbox(label="Model Details", lines=30)
        info_btn.click(gradio_info, outputs=info_output)

    with gr.Tab("API"):
        gr.Markdown("""
        ## API Usage

        ```python
        from gradio_client import Client

        client = Client("Finish-him/dota2-edge-ensemble")

        # V3 Production prediction
        result = client.predict(
            moment="m10",
            features_text='{"gold_delta": 3500, "xp_delta": 2200, "gpm_delta": 350, ...}',
            api_name="/predict_v3"
        )
        print(result)

        # V1 Legacy prediction
        result = client.predict(
            moment="m1",
            features_text='{"gold_delta_10": 2500}',
            api_name="/predict_v1"
        )
        ```

        ### V3 Feature List (19 features)
        `gold_delta`, `xp_delta`, `lh_delta`, `dn_delta`, `rad_gpm`, `dire_gpm`,
        `gpm_delta`, `rad_xpm`, `dire_xpm`, `xpm_delta`, `carry_gap`, `support_gap`,
        `cs_eff_delta`, `deny_delta`, `gold_log_ratio`, `xp_log_ratio`, `gold_lead_pct`,
        `gold_momentum`, `xp_momentum`
        """)


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