Prediction confidence analysis

nbs/inference.ipynb

@@ -2,38 +2,34 @@
  "cells": [
   {
    "cell_type": "code",
-   "execution_count": 1,
+   "execution_count": null,
    "id": "fd1758c9-040d-4727-96d8-951d385ba277",
    "metadata": {
     "tags": []
    },
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "/admin/home-devforfu/realfake\n"
-     ]
-    }
-   ],
+   "outputs": [],
    "source": [
     "%cd .."
    ]
   },
   {
    "cell_type": "code",
-   "execution_count": 98,
+   "execution_count": null,
    "id": "eb53a9fc-90eb-4658-b24c-f6f33c731235",
    "metadata": {
     "tags": []
    },
    "outputs": [],
    "source": [
-    "import random\n",
     "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"
@@ -41,7 +37,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 68,
+   "execution_count": null,
    "id": "41a0c3c5-d01c-46ef-b50b-25d9c297a432",
    "metadata": {
     "tags": []
@@ -62,7 +58,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 69,
+   "execution_count": null,
    "id": "14ca2c34-99f0-4088-99d8-9b0cd008097a",
    "metadata": {
     "tags": []
@@ -74,128 +70,113 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 109,
-   "id": "fb594a6e-0f75-4d06-a65e-68dda1d39ef6",
-   "metadata": {
-    "tags": []
-   },
-   "outputs": [
-    {
-     "data": {
-      "text/plain": [
-       "2504"
-      ]
-     },
-     "execution_count": 109,
-     "metadata": {},
-     "output_type": "execute_result"
-    }
-   ],
-   "source": [
-    "fake = [{\"path\": str(fn), \"label\": \"fake\"} for fn in Path(\"fakes\").glob(\"**/*.png\")]\n",
-    "len(fake)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 110,
-   "id": "da8ae077-53d7-494a-b330-9d674c257734",
+   "execution_count": null,
+   "id": "d3445075-07ba-424c-9849-26c9de6ce1a8",
    "metadata": {
     "tags": []
    },
    "outputs": [],
    "source": [
-    "imagenet_validation = list(Path(f\"/fsx/{get_user_name()}/data/imagenet-1k/validation\").glob(\"*.JPEG\"))"
+    "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": 111,
-   "id": "59c21450-c7bb-479e-8d07-5861dca3dac7",
+   "execution_count": null,
+   "id": "7b76d082-9e35-455d-8c86-0300ffa224d0",
    "metadata": {
     "tags": []
    },
    "outputs": [],
    "source": [
-    "real = [{\"path\": str(fn), \"label\": \"real\"} for fn in random.choices(imagenet_validation, k=len(fakes))]"
+    "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": 113,
-   "id": "1d445977-e1f1-40c2-8975-1eb1f7c04493",
+   "execution_count": null,
+   "id": "9c358c74-845d-42f1-9fcb-673e2a90ef69",
    "metadata": {
     "tags": []
    },
    "outputs": [],
    "source": [
-    "records = fake + real"
+    "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": 117,
-   "id": "3a25ded0-c284-4148-b29a-25897cce9c5b",
+   "execution_count": null,
+   "id": "aeb284f7-46a3-408c-afa4-158ba9640571",
    "metadata": {
     "tags": []
    },
    "outputs": [],
    "source": [
-    "random.shuffle(records)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 120,
-   "id": "7b76d082-9e35-455d-8c86-0300ffa224d0",
-   "metadata": {
-    "tags": []
-   },
-   "outputs": [
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 157/157 [03:45<00:00,  1.44s/it]\n"
-     ]
-    }
-   ],
-   "source": [
-    "batch_size = 128\n",
-    "    \n",
-    "with torch.inference_mode():\n",
-    "    ds = DictDataset(records, get_augs(train=False))\n",
-    "    dl = DataLoader(ds, batch_size=32, num_workers=8, shuffle=False)\n",
-    "\n",
-    "    matched, total = 0, len(ds)\n",
-    "\n",
-    "    for batch in tqdm(dl):\n",
-    "        _, logits, y_true_onehot = model(batch)\n",
-    "        y_true = y_true_onehot.argmax(dim=1)\n",
-    "        y_pred = logits.softmax(dim=1).argmax(dim=1)\n",
-    "        equals = y_true == y_pred\n",
-    "        # print(equals.float().mean())\n",
-    "        matched += equals.sum().item()"
+    "view_results(scores, \"label == 'fake' and fake_prob >= 0.8\")"
    ]
   },
   {
    "cell_type": "code",
-   "execution_count": 121,
-   "id": "6077a3f7-4ad5-4fc0-bc53-7e939b64a658",
+   "execution_count": null,
+   "id": "f518ea74-3c97-4bb9-a461-73c108dac75f",
    "metadata": {
     "tags": []
    },
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Accuracy: 99.58%\n"
-     ]
-    }
-   ],
+   "outputs": [],
    "source": [
-    "print(f\"Accuracy: {matched/total:2.2%}\")"
+    "view_results(scores, \"label == 'fake' and fake_prob < 0.5\")"
    ]
   }
  ],