{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Generating images with MX-Font model from a reference style\n",
    "In this example we'll generate images with trained MX-Font model from a reference style.\n",
    "If you want to generate multiple styles, please check using `eval.py` instead of using this example file (because it is much simpler to load the referece styles)."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 1. Loading packages\n",
    "* First, load the packages used in this code.\n",
    "* All of the packages are avilable in `pip`."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import json\n",
    "from pathlib import Path\n",
    "from PIL import Image\n",
    "\n",
    "import torch\n",
    "from sconf import Config\n",
    "from torchvision import transforms"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "* These modules are defined in this repository."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "import models\n",
    "from datasets import read_font, render\n",
    "from utils import save_tensor_to_image"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 2. Build model\n",
    "* Build and load the trained model.\n",
    "* `weight_path` : \n",
    "    * The location of the trained model weight."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<All keys matched successfully>"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "########################################################\n",
    "weight_path = \"generator.pth\"  # path to weight to infer\n",
    "########################################################\n",
    "\n",
    "cfg = Config(\"cfgs/eval.yaml\", default=\"cfgs/defaults.yaml\")\n",
    "transform = transforms.Compose(\n",
    "    [transforms.Resize((128, 128)), transforms.ToTensor(), transforms.Normalize([0.5], [0.5])]\n",
    ")\n",
    "decomposition = json.load(open(\"data/chn_decomposition.json\"))\n",
    "\n",
    "g_kwargs = cfg.get('g_args', {})\n",
    "gen = models.Generator(1, cfg.C, 1, **g_kwargs).cuda().eval()\n",
    "weight = torch.load(weight_path)\n",
    "if \"generator_ema\" in weight:\n",
    "    weight = weight[\"generator_ema\"]\n",
    "gen.load_state_dict(weight)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 3. Load reference images.\n",
    "* `ref_path`: \n",
    "    * The path of reference font or images.\n",
    "    * If you are using a ttf file, set this to the location of the ttf file.\n",
    "    * If you want to use rendered images, set this to the path to the directory which contains the reference images.\n",
    "* `extension`:\n",
    "    * If you are using image files, set this to their extension(png, jpg, etc..). \n",
    "    * This will be ignored if `use_ttf` is True.\n",
    "* `batch_size`:\n",
    "    * The number of images inferred at once."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "########################################################\n",
    "ref_path = \"data/images/test/ZCOOLKuaiLe-Regular\"  # Path to the reference images\n",
    "extension = \"png\"  # Extension of the reference images\n",
    "batch_size = 3  # The batch size\n",
    "########################################################\n",
    "\n",
    "ref_paths = Path(ref_path).glob(f\"*.{extension}\")\n",
    "ref_imgs = torch.stack([transform(Image.open(str(p))) for p in ref_paths]).cuda()\n",
    "ref_batches = torch.split(ref_imgs, batch_size)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 4. Extract style factors from reference images."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "style_facts = {}\n",
    "\n",
    "for batch in ref_batches:\n",
    "    style_fact = gen.factorize(gen.encode(batch), 0)\n",
    "    for k in style_fact:\n",
    "        style_facts.setdefault(k, []).append(style_fact[k])\n",
    "        \n",
    "style_facts = {k: torch.cat(v).mean(0, keepdim=True) for k, v in style_facts.items()}"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 5. Generate the images.\n",
    "* `gen_chars`: The characters to generate.\n",
    "* `save_dir`: Path to save the generated images.\n",
    "* `source_path`: Path to the font file used as the source font."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "########################################################\n",
    "gen_chars = \"意味岸\"  # Characters to generate\n",
    "save_dir = Path(\"./results\")  # Directory where you want to save generated images\n",
    "source_path = \"data/chn_source.ttf\"  # Path to the font file to render the source images\n",
    "########################################################\n",
    "\n",
    "save_dir.mkdir(parents=True, exist_ok=True)\n",
    "\n",
    "source_font = read_font(source_path)\n",
    "for char in gen_chars:\n",
    "    source_img = transform(render(source_font, char)).unsqueeze(0).cuda()\n",
    "    char_facts = gen.factorize(gen.encode(source_img), 1)\n",
    "    \n",
    "    gen_feats = gen.defactorize([style_facts, char_facts])\n",
    "    out = gen.decode(gen_feats).detach().cpu()[0]\n",
    "\n",
    "    path = save_dir / f\"{char}.png\"\n",
    "    save_tensor_to_image(out, path)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.8.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}