Upload New File

Retrieve_combined_features.ipynb

+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "f8b76e74",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import pandas as pd\n",
+    "import numpy as np\n",
+    "import matplotlib.pyplot as plt\n",
+    "from PIL import Image\n",
+    "from annoy import AnnoyIndex\n",
+    "import random\n",
+    "\n",
+    "base_dir = 'path_to_the_dataset'\n",
+    "top_count = 100\n",
+    "top_K = 11"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "43c9c917",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#the 1st column is index, the second is path to the paintings, rest columns are pre-calculated embeddings\n",
+    "df_artist = pd.read_csv('dino_features_wikiart_artist.csv')\n",
+    "df_genre = pd.read_csv('dino_features_wikiart_genre.csv')\n",
+    "df_style = pd.read_csv('dino_features_wikiart_style.csv')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "be5120fe",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def build_index(df, filename):\n",
+    "    NUMBER_OF_TREES = 25\n",
+    "    feature_dim = len(df.columns) - 2\n",
+    "\n",
+    "    t = AnnoyIndex(feature_dim, metric='euclidean')\n",
+    "\n",
+    "    for i in range(len(df)):\n",
+    "        vector = df.loc[i][2:]\n",
+    "        t.add_item(i, vector)\n",
+    "    \n",
+    "    _  = t.build(NUMBER_OF_TREES)\n",
+    "\n",
+    "    #save indexes\n",
+    "    t.save(base_dir + filename)\n",
+    "    return t"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "afa13d41",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "t_genre = build_index(df_genre, 'dino_genre.ann')\n",
+    "t_style = build_index(df_style, 'dino_style.ann')\n",
+    "t_artist = build_index(df_artist, 'dino_artist.ann')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "feff4c2f",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#download saved indexes\n",
+    "feature_dim = len(df_genre.columns) - 2\n",
+    "\n",
+    "t_genre = AnnoyIndex(feature_dim, metric='euclidean')\n",
+    "t_genre.load('dino_genre.ann')\n",
+    "\n",
+    "t_style = AnnoyIndex(feature_dim, metric='euclidean')\n",
+    "t_style.load('dino_style.ann')\n",
+    "\n",
+    "t_artist = AnnoyIndex(feature_dim, metric='euclidean')\n",
+    "t_artist.load('dino_artist.ann')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "ee4c29ab",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#receive similar images using ANNOY\n",
+    "def get_similar_images_annoy(vector, t):\n",
+    "    indices, dists = t.get_nns_by_vector(vector, top_count+1, include_distances=True)\n",
+    "    return indices, dists"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "1b10c4d0",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def show_similar(img_str, img_list, k, distances):\n",
+    "    img = Image.open(base_dir + img_str)\n",
+    "    plt.title('Request painting')\n",
+    "    plt.axis(\"off\")\n",
+    "    plt.imshow(img)\n",
+    "\n",
+    "    plt.figure(figsize=(30, 30))\n",
+    "    ax = plt.subplot(1, k, 1)\n",
+    "    \n",
+    "    i = 0\n",
+    "    for key in img_list:\n",
+    "        ax = plt.subplot(1, k, i+1)\n",
+    "        img = Image.open(base_dir + key)\n",
+    "        plt.title(\"{}\".format(round(distances[i], 4)))\n",
+    "        plt.imshow(img)\n",
+    "        i+=1\n",
+    "\n",
+    "        plt.axis(\"off\")\n",
+    "    plt.show()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "29ccae77",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#request image\n",
+    "img_str = 'Impressionism/pierre-auguste-renoir_children-on-the-seashore-1883.jpg'\n",
+    "\n",
+    "#pre-calculated embeddings of the request painting\n",
+    "img_req_g = np.array(df_genre[df_genre['file_path']==img_str])[0][2:]\n",
+    "img_req_s = np.array(df_style[df_style['file_path']==img_str])[0][2:]\n",
+    "img_req_a = np.array(df_artist[df_artist['file_path']==img_str])[0][2:]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "9c2c37c1",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "similar_images_g = []\n",
+    "similar_images_s = []\n",
+    "similar_images_a = []\n",
+    "\n",
+    "similar_img_ids_g, distances_g = get_similar_images_annoy(img_req_g, t_genre)\n",
+    "similar_images_g = list(df_genre.iloc[similar_img_ids_g]['file_path'])\n",
+    "\n",
+    "similar_img_ids_s, distances_s = get_similar_images_annoy(img_req_s, t_style)\n",
+    "similar_images_s = list(df_style.iloc[similar_img_ids_s]['file_path'])\n",
+    "\n",
+    "similar_img_ids_a, distances_a = get_similar_images_annoy(img_req_a, t_artist)\n",
+    "similar_images_a = list(df_artist.iloc[similar_img_ids_a]['file_path'])"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "eac7636a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#pre-calculated values of mean and std of eucledian distances on WikiArt dataset\n",
+    "mean_genre = 19.6129883775738\n",
+    "std_genre = 2.9717604654152274\n",
+    "mean_style = 18.328733753985198\n",
+    "std_style = 2.5707980486965547\n",
+    "mean_artist = 20.973407871326454\n",
+    "std_artist = 2.9063593849149183"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "08e0941d",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Calculating cosine distance\n",
+    "#request painting belongs to the dataset, so we ignore the first similar painting\n",
+    "df_genre_sim = df_genre[df_genre['file_path'].isin(similar_images_g)]\n",
+    "cosines_genre = distances_g[1:]\n",
+    "\n",
+    "df_style_sim = df_style[df_style['file_path'].isin(similar_images_s)]\n",
+    "cosines_style = distances_s[1:]\n",
+    "\n",
+    "df_artist_sim = df_artist[df_artist['file_path'].isin(similar_images_a)]\n",
+    "cosines_artist = distances_a[1:]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "b48542d3",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "d = {'file_path': list(df_genre_sim['file_path'].iloc[1:]), 'd_g': cosines_genre}\n",
+    "df_g = pd.DataFrame(data=d)\n",
+    "\n",
+    "d = {'file_path': list(df_style_sim['file_path'].iloc[1:]), 'd_s': cosines_style}\n",
+    "df_s = pd.DataFrame(data=d)\n",
+    "\n",
+    "d = {'file_path': list(df_artist_sim['file_path'].iloc[1:]), 'd_a': cosines_artist}\n",
+    "df_a = pd.DataFrame(data=d)\n",
+    "\n",
+    "df_u = pd.merge(df_g, df_s, on='file_path', how='outer')\n",
+    "df_u = pd.merge(df_u, df_a, on='file_path', how='outer')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "92bb15c8",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def func(x):\n",
+    "    missing_indices = x.index[x.isnull()]\n",
+    "    for index in missing_indices:\n",
+    "        if x.name == 'd_g':\n",
+    "            img_sim = df_u.loc[index, 'file_path']\n",
+    "            eucl_distance = np.linalg.norm(img_req_g - df_genre[df_genre['file_path'] == img_sim].iloc[0, 2:])\n",
+    "            x.loc[index] = eucl_distance\n",
+    "        elif x.name == 'd_s':   \n",
+    "            img_sim = df_u.loc[index, 'file_path']\n",
+    "            eucl_distance = np.linalg.norm(img_req_s - df_style[df_style['file_path'] == img_sim].iloc[0, 2:])\n",
+    "            x.loc[index] = eucl_distance\n",
+    "        elif x.name == 'd_a':   \n",
+    "            img_sim = df_u.loc[index, 'file_path']\n",
+    "            eucl_distance = np.linalg.norm(img_req_a - df_artist[df_artist['file_path'] == img_sim].iloc[0, 2:])\n",
+    "            x.loc[index] = eucl_distance\n",
+    "\n",
+    "    return x\n",
+    "\n",
+    "df_u = df_u.apply(func, axis=0)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "7e464ba0",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "df_u['d_g'] = (df_u['d_g'] - mean_genre)/std_genre\n",
+    "df_u['d_s'] = (df_u['d_s'] - mean_style)/std_style\n",
+    "df_u['d_a'] = (df_u['d_a'] - mean_style)/std_style"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "5ea5a057",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#combined recommendations with the same weights\n",
+    "df_u['combined'] = df_u['d_g'] + df_u['d_s'] + df_u['d_a']"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "b9e0d83f",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "paintings = list(df_u.sort_values(by=['combined'])['file_path'])[:top_K]\n",
+    "dist = list(df_u.sort_values(by=['combined'])['combined'])[:top_K]\n",
+    "show_similar(img_str, paintings, top_K, dist)"
+   ]
+  }
+ ],
+ "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": "3.10.14"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}