{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# JMangaTranslator-Fast v1\n",
    "\n",
    "Japanese → Simplified Chinese translation of manga speech bubbles, one bubble at a time.\n",
    "日译简中漫画对话框翻译，一次翻译一个框。\n",
    "\n",
    "- **Colab**: Runtime → Change runtime type → **T4 GPU**.\n",
    "- **Kaggle**: Settings → Accelerator → **GPU T4**, and turn **Internet** on.\n",
    "- Without a GPU the notebook runs on the CPU with ONNX Runtime (slower). 没有 GPU 时改用 CPU 上的 ONNX Runtime（较慢）。\n",
    "\n",
    "Model card: https://huggingface.co/muscgab/JMangaTranslator-Fast"
   ]
  },
  {
   "cell_type": "code",
   "metadata": {},
   "execution_count": null,
   "outputs": [],
   "source": [
    "# 1. Install dependencies (PyTorch is preinstalled on Colab and Kaggle)\n",
    "import subprocess, sys\n",
    "import torch\n",
    "\n",
    "GPU = torch.cuda.is_available()\n",
    "pkgs = [\"transformers>=5.14\", \"safetensors\", \"sentencepiece\", \"tokenizers\", \"huggingface_hub\"]\n",
    "if not GPU:\n",
    "    pkgs.append(\"onnxruntime\")\n",
    "subprocess.run([sys.executable, \"-m\", \"pip\", \"install\", \"-q\", *pkgs], check=True)\n",
    "print(\"GPU:\", torch.cuda.get_device_name(0) if GPU else \"none, using the CPU (ONNX Runtime)\")"
   ]
  },
  {
   "cell_type": "code",
   "metadata": {},
   "execution_count": null,
   "outputs": [],
   "source": [
    "# 2. Download only the files this runtime needs (about 1.5 GB)\n",
    "from huggingface_hub import snapshot_download\n",
    "\n",
    "files = [\"config.json\", \"vocab.json\", \"joint.model\", \"tokenizer/*\", \"jmt_fast/*\", \"translate.py\"]\n",
    "files += [\"model.safetensors\", \"norm_scales.json\"] if GPU else [\"onnx/*\"]\n",
    "path = snapshot_download(\"muscgab/JMangaTranslator-Fast\", allow_patterns=files)\n",
    "print(path)"
   ]
  },
  {
   "cell_type": "code",
   "metadata": {},
   "execution_count": null,
   "outputs": [],
   "source": [
    "# 3. Load the model once (CUDA Graphs fp16 on a GPU, ONNX Runtime on the CPU)\n",
    "import time\n",
    "sys.path.insert(0, path)\n",
    "from jmt_fast import load, pick\n",
    "\n",
    "t = time.perf_counter()\n",
    "tr = load(path)\n",
    "tr.translate(\"こんにちは\")   # warm-up: the first call is slower\n",
    "print(f\"backend: {pick()}, loaded in {time.perf_counter() - t:.1f} s\")"
   ]
  },
  {
   "cell_type": "code",
   "metadata": {},
   "execution_count": null,
   "outputs": [],
   "source": [
    "# 4. Translate. Each call is one speech bubble; remove line breaks inside a bubble first.\n",
    "bubbles = [\n",
    "    \"堪忍袋の緒が切れた！\",\n",
    "    \"お前、まだそんなこと言ってるのか？\",\n",
    "    \"今日こそ決着をつけてやる！\",\n",
    "    \"べ、別にあんたのためじゃないんだからね！\",\n",
    "    \"腹減ったなあ……\",\n",
    "]\n",
    "for b in bubbles:\n",
    "    t = time.perf_counter()\n",
    "    out = tr.translate(b)\n",
    "    print(f\"{1000 * (time.perf_counter() - t):6.1f} ms  {b}  →  {out}\")"
   ]
  },
  {
   "cell_type": "code",
   "metadata": {},
   "execution_count": null,
   "outputs": [],
   "source": [
    "# 5. Your own text\n",
    "print(tr.translate(\"ここに日本語のセリフを入力してください\"))"
   ]
  },
  {
   "cell_type": "code",
   "metadata": {},
   "execution_count": null,
   "outputs": [],
   "source": [
    "# 6. A file with one bubble per line: input.txt → output.txt\n",
    "from pathlib import Path\n",
    "\n",
    "src = Path(\"input.txt\")\n",
    "if src.exists():\n",
    "    lines = src.read_text(encoding=\"utf-8\").splitlines()\n",
    "    Path(\"output.txt\").write_text(\"\\n\".join(tr.translate(x) if x.strip() else \"\" for x in lines) + \"\\n\", encoding=\"utf-8\")\n",
    "    print(f\"translated {len(lines)} lines → output.txt\")\n",
    "else:\n",
    "    print(\"Upload input.txt (one bubble per line) and run this cell again.\")"
   ]
  }
 ],
 "metadata": {
  "accelerator": "GPU",
  "colab": {
   "gpuType": "T4",
   "provenance": []
  },
  "kernelspec": {
   "display_name": "Python 3",
   "name": "python3"
  },
  "language_info": {
   "name": "python"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 0
}