app.py

"""
Gradio demo for Hugging Face Spaces.
Upload a short Minecraft gameplay clip → top-3 action predictions.

Deploy: create Space with sdk=gradio, app_file=app.py
Requires voxel_scripted.pt in the model repo (download on startup).
"""

import os
from pathlib import Path

import gradio as gr
import torch

from hf_inference import CLASS_NAMES, load_model, predict_video

MODEL_ID = os.environ.get("HF_MODEL_ID", "fotographer/VoxelMind-2M-Minecraft-Classifier")
MODEL_FILE = os.environ.get("VOXELMIND_WEIGHTS", "models/voxel_scripted.pt")
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

_model = None


def get_model():
    global _model
    if _model is None:
        path = Path(MODEL_FILE)
        if not path.exists():
            from huggingface_hub import hf_hub_download

            path = Path(
                hf_hub_download(
                    repo_id=MODEL_ID,
                    filename=MODEL_FILE,
                    repo_type="model",
                )
            )
        _model = load_model(path, DEVICE)
    return _model


def run(video_path: str):
    if not video_path:
        return "Upload a video first."
    preds = predict_video(get_model(), video_path, device=DEVICE, topk=3)
    lines = [f"**{name}** — {prob * 100:.1f}%" for name, prob in preds]
    return "\n\n".join(lines)


demo = gr.Interface(
    fn=run,
    inputs=gr.Video(label="Minecraft clip (≥22 frames recommended)"),
    outputs=gr.Markdown(label="Top-3 predictions"),
    title="VoxelMind action classifier",
    description=(
        "Classifies player actions from 22×64×64 grayscale frames. "
        f"Classes: {', '.join(CLASS_NAMES)}. "
        "Inference uses a TorchScript export — architecture source is not published."
    ),
    examples=[],
)

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