{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "fd1758c9-040d-4727-96d8-951d385ba277",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "%cd .."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "eb53a9fc-90eb-4658-b24c-f6f33c731235",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "from pathlib import Path\n",
    "\n",
    "import matplotlib.pyplot as plt\n",
    "import pandas as pd\n",
    "import PIL.Image\n",
    "import torch\n",
    "from tqdm import tqdm\n",
    "from torch.utils.data import DataLoader\n",
    "\n",
    "from realfake.data import DictDataset, get_augs\n",
    "from realfake.models import RealFakeClassifier, RealFakeParams\n",
    "from realfake.utils import find_latest_checkpoint, get_user_name, read_jsonl"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "41a0c3c5-d01c-46ef-b50b-25d9c297a432",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "def load_from_checkpoint(checkpoint_dir, name=None, map_location=\"cpu\"):\n",
    "    checkpoint_dir = Path(checkpoint_dir)\n",
    "    path = find_latest_checkpoint(checkpoint_dir) if name is None else checkpoint_dir/name\n",
    "    checkpoint = torch.load(path, map_location)\n",
    "    params = RealFakeParams.parse_file(path.parent/\"params.json\")\n",
    "    params.pretrained = False\n",
    "    classifier = RealFakeClassifier(params)\n",
    "    classifier.load_state_dict(checkpoint[\"state_dict\"])\n",
    "    classifier.eval()\n",
    "    return classifier"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "14ca2c34-99f0-4088-99d8-9b0cd008097a",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "model = load_from_checkpoint(\"checkpoints/convnext_large_2m_e5\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "d3445075-07ba-424c-9849-26c9de6ce1a8",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "real = [{\"path\": str(p), \"label\": \"real\"} for p in Path(\"imagenet_val\").iterdir()]\n",
    "fake = [{\"path\": str(p), \"label\": \"fake\"} for p in Path(\"fakes\").glob(\"**/*.png\")]\n",
    "data = real + fake\n",
    "len(data)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "7b76d082-9e35-455d-8c86-0300ffa224d0",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "batch_size = 128\n",
    "scores = []\n",
    "\n",
    "with torch.inference_mode():\n",
    "    ds = DictDataset(data, get_augs(train=False))\n",
    "    dl = DataLoader(ds, batch_size=batch_size, num_workers=8, shuffle=False)\n",
    "\n",
    "    for batch in tqdm(dl):\n",
    "        _, logits, y_true_onehot = model(batch)\n",
    "        probs = logits.softmax(dim=1)\n",
    "        y_true = y_true_onehot.argmax(dim=1)\n",
    "        y_pred = probs.argmax(dim=1)\n",
    "        matched = y_true == y_pred\n",
    "        \n",
    "        scores += [\n",
    "            {\"fake_prob\": fake_prob.item(), \"match\": match.item()}\n",
    "            for fake_prob, match in zip(probs[:, 1], matched)\n",
    "        ]\n",
    "        \n",
    "scores = pd.DataFrame(scores)\n",
    "scores[\"label\"] = [r[\"label\"] for r in data]\n",
    "scores[\"path\"] = [r[\"path\"] for r in data]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "9c358c74-845d-42f1-9fcb-673e2a90ef69",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "def view_results(df: pd.DataFrame, \n",
    "                 query: str, \n",
    "                 img_size: int = 256, \n",
    "                 plot_size: int = 4,\n",
    "                 n_rows: int = 5,\n",
    "                 n_cols: int = 5):\n",
    "    \n",
    "    f, axes = plt.subplots(n_rows, n_cols, \n",
    "                           figsize=(n_cols*plot_size, n_rows*plot_size), \n",
    "                           gridspec_kw={\"hspace\": 0.1, \"wspace\": 0})\n",
    "    \n",
    "    f.subplots_adjust(hspace=0, wspace=0)\n",
    "     \n",
    "    sz = img_size\n",
    "    \n",
    "    items = (df.sort_values(by=\"fake_prob\")\n",
    "             .reset_index(drop=True)\n",
    "             .query(query)\n",
    "             .apply(lambda rec: (\n",
    "                 PIL.Image.open(rec.path).resize((sz,sz)), \n",
    "                 rec.fake_prob), axis=1)\n",
    "             .path.tolist())\n",
    "\n",
    "    for ax, (im, score) in zip(axes.flat, items):\n",
    "        ax.imshow(im)\n",
    "        ax.set_title(f\"P(fake)={score:2.2%}\")\n",
    "        ax.set_axis_off()\n",
    "        ax.set_aspect(\"equal\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "aeb284f7-46a3-408c-afa4-158ba9640571",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "view_results(scores, \"label == 'fake' and fake_prob >= 0.8\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "f518ea74-3c97-4bb9-a461-73c108dac75f",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "view_results(scores, \"label == 'fake' and fake_prob < 0.5\")"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "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": "3.8.10"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}