diff --git a/Phase 2/task_11.ipynb b/Phase 2/task_11.ipynb
new file mode 100644
index 0000000..9b2b34d
--- /dev/null
+++ b/Phase 2/task_11.ipynb
@@ -0,0 +1,322 @@
+{
+ "cells": [
+ {
+ "cell_type": "code",
+ "execution_count": 1,
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "from utils import *\n",
+ "warnings.filterwarnings('ignore')\n",
+ "%matplotlib inline\n"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 2,
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "fd_collection = getCollection(\"team_5_mwdb_phase_2\", \"fd_collection\")\n"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 12,
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "class ImageGraph:\n",
+ " \"\"\"\n",
+ " Construct image-similarity graph and apply personalized pagerank algorithm to get PPR scores and find relevant images\n",
+ " \"\"\"\n",
+ "\n",
+ " def __init__(self, fd_collection, verbose=False):\n",
+ " self.fd_collection = fd_collection\n",
+ " self.similarity_graph = None\n",
+ " self.verbose = verbose\n",
+ "\n",
+ " def create_similarity_graph(\n",
+ " self, n, feature_model, semantic_data=None, dim_reduction_method=None\n",
+ " ):\n",
+ " if semantic_data is None:\n",
+ " # Similarity graph from feature models\n",
+ " image_sim_matrix = find_image_image_similarity(fd_collection, feature_model)\n",
+ " if self.verbose:\n",
+ " print(\"Image-image similarity matrix constructed from\", feature_model)\n",
+ " else:\n",
+ " # Similarity graph from image-semantic latent space\n",
+ " # LS3, LS4\n",
+ " if \"sim-matrix\" in semantic_data:\n",
+ " # for now, don't work with LS3\n",
+ " # TODO: do similar to task 7 and 10\n",
+ " image_sim_matrix = np.array(semantic_data[\"sim-matrix\"])\n",
+ " if image_sim_matrix.shape[0] != NUM_IMAGES:\n",
+ " raise TypeError(\n",
+ " \"Functionality to construct similarity graph from LS3 not yet done\"\n",
+ " )\n",
+ " if self.verbose:\n",
+ " print(\"Using image-image similarity matrix from semantic data\")\n",
+ " # LS1, LS2\n",
+ " else:\n",
+ " image_semantic = semantic_data[\"image-semantic\"]\n",
+ " # SVD, CP\n",
+ " if \"semantics-core\" in semantic_data:\n",
+ " semantics_core = np.array(semantic_data[\"semantics-core\"])\n",
+ " if len(semantics_core.shape) == 1:\n",
+ " semantics_core = np.diag(semantics_core)\n",
+ " image_semantic = np.matmul(image_semantic, semantics_core)\n",
+ "\n",
+ " image_sim_matrix = np.zeros((NUM_IMAGES, NUM_IMAGES))\n",
+ " # Calculate half and fill the other\n",
+ " for i in range(NUM_IMAGES):\n",
+ " for j in range(i + 1, NUM_IMAGES):\n",
+ " # Note: lower the value, lower the distance => higher the similarity\n",
+ " distance_measure = (\n",
+ " kl_divergence_measure\n",
+ " if dim_reduction_method == \"lda\"\n",
+ " else euclidean_distance_measure\n",
+ " )\n",
+ " image_sim_matrix[j][i] = distance_measure(\n",
+ " np.array(image_semantic[i]),\n",
+ " np.array(image_semantic[j]),\n",
+ " )\n",
+ " image_sim_matrix[i][j] = image_sim_matrix[j][i]\n",
+ " if self.verbose:\n",
+ " print(\n",
+ " \"Image-image similarity matrix constructed from given image-semantic\"\n",
+ " )\n",
+ "\n",
+ " # Create an unweighted directed similarity graph, with no self-loops\n",
+ " self.similarity_graph = []\n",
+ " for i in range(len(image_sim_matrix)):\n",
+ " # distances should be small, so sort in ascending order\n",
+ " similar_image_ids = np.argsort(image_sim_matrix[i])[\n",
+ " 1 : n + 1\n",
+ " ] # exclude self\n",
+ " self.similarity_graph.extend(\n",
+ " [(i * 2, j * 2) for j in similar_image_ids]\n",
+ " ) # i*2 cuz even IDs\n",
+ " if self.verbose:\n",
+ " print(\"Similarity graph created\")\n",
+ "\n",
+ " def personalized_pagerank(\n",
+ " self, label, m, damping_factor=0.85, max_iter=1000, tol=1e-6\n",
+ " ):\n",
+ " import time\n",
+ " if self.similarity_graph is None:\n",
+ " raise ValueError(\n",
+ " \"Similarity graph not created. Call create_similarity_graph() first.\"\n",
+ " )\n",
+ "\n",
+ " label_indices = [\n",
+ " img[\"image_id\"] for img in self.fd_collection.find({\"true_label\": label})\n",
+ " ] # IDs of images with the given label\n",
+ "\n",
+ " pr_scores = np.ones(NUM_IMAGES) / NUM_IMAGES # Initialize PageRank scores\n",
+ " if self.verbose:\n",
+ " print(\"Initialized pagerank scores\")\n",
+ "\n",
+ " for _iter in range(max_iter):\n",
+ " prev_scores = np.copy(pr_scores)\n",
+ " # for every node,\n",
+ " for i in range(NUM_IMAGES):\n",
+ " tic = time.time()\n",
+ " # add sum of connected nodes' PR scores\n",
+ " pr_scores[i] = damping_factor * sum(\n",
+ " pr_scores[j]\n",
+ " for j in range(NUM_IMAGES)\n",
+ " if (i * 2, j * 2) in self.similarity_graph\n",
+ " # and add the prob for random teleport *if node in given label*\n",
+ " ) + (1 - damping_factor) * (1 if i * 2 in label_indices else 0) / len(\n",
+ " label_indices\n",
+ " )\n",
+ " toc = time.time()\n",
+ " print(toc-tic)\n",
+ "\n",
+ " pr_scores /= sum(pr_scores) # Normalize\n",
+ "\n",
+ " # check for convergence\n",
+ " conv_tol = np.sum(np.abs(prev_scores - pr_scores))\n",
+ " if self.verbose:\n",
+ " print(f\"Iter {_iter}, conv_tol={conv_tol}\")\n",
+ " if conv_tol < tol:\n",
+ " if self.verbose:\n",
+ " print(f\"Converged\")\n",
+ " break\n",
+ "\n",
+ " # Select top m images based on PageRank scores\n",
+ " top_m_indices = np.argsort(pr_scores)[::-1][\n",
+ " :m\n",
+ " ] # sort indices, reverse and take top m\n",
+ " return top_m_indices * 2 # again, even IDs\n"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 15,
+ "metadata": {},
+ "outputs": [
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "image_sim-cm_fd-svd-10-semantics.json loaded\n"
+ ]
+ }
+ ],
+ "source": [
+ "n = int(input(\"Enter value of n (no. of edges for each image in similarity graph): \"))\n",
+ "if n < 1:\n",
+ " raise ValueError(\"n should be a positive integer\")\n",
+ "\n",
+ "l = int(input(\"Enter target label l:\"))\n",
+ "if l < 0 or l > 100:\n",
+ " raise ValueError(\"l should be in range[0,100]\")\n",
+ "\n",
+ "m = int(input(\"Enter value of m (no. of significant images relative to given label): \"))\n",
+ "if m < 1:\n",
+ " raise ValueError(\"m should be a positive integer\")\n",
+ "\n",
+ "given_ls = int(\n",
+ " (input(\"Enter 0 to select a feature model, 1 to select a latent space: \"))\n",
+ ")\n",
+ "\n",
+ "selected_feature_model = valid_feature_models[\n",
+ " str(input(\"Enter feature model - one of \" + str(list(valid_feature_models.keys()))))\n",
+ "]\n",
+ "\n",
+ "if given_ls:\n",
+ " selected_latent_space = valid_latent_spaces[\n",
+ " str(\n",
+ " input(\n",
+ " \"Enter latent space - one of \" + str(list(valid_latent_spaces.keys()))\n",
+ " )\n",
+ " )\n",
+ " ]\n",
+ "\n",
+ " k = int(input(\"Enter value of k (no. of latent semantics): \"))\n",
+ " if k < 1:\n",
+ " raise ValueError(\"k should be a positive integer\")\n",
+ "\n",
+ " if selected_latent_space != \"cp\":\n",
+ " selected_dim_reduction_method = str(\n",
+ " input(\n",
+ " \"Enter dimensionality reduction method - one of \"\n",
+ " + str(list(valid_dim_reduction_methods.keys()))\n",
+ " )\n",
+ " )\n",
+ "\n",
+ " # Loading latent semantics\n",
+ " match selected_latent_space:\n",
+ " # LS1\n",
+ " case \"\":\n",
+ " file_prefix = (\n",
+ " f\"{selected_feature_model}-{selected_dim_reduction_method}-{k}\"\n",
+ " )\n",
+ " file_name = file_prefix + \"-semantics.json\"\n",
+ " model_name = file_prefix + \"-model.joblib\"\n",
+ " if os.path.exists(file_name):\n",
+ " data = json.load(open(file_name))\n",
+ " print(file_name + \" loaded\")\n",
+ " else:\n",
+ " raise Exception(file_name + \" does not exist\")\n",
+ " # LDA model\n",
+ " if selected_dim_reduction_method == \"lda\":\n",
+ " if os.path.exists(model_name):\n",
+ " data_model = load(model_name)\n",
+ " print(model_name + \" loaded\")\n",
+ " else:\n",
+ " raise Exception(model_name + \" does not exist\")\n",
+ " # LS2\n",
+ " case \"cp\":\n",
+ " file_name = f\"{selected_feature_model}-cp-{k}-semantics.json\"\n",
+ " if os.path.exists(file_name):\n",
+ " data = json.load(open(file_name))\n",
+ " print(file_name + \" loaded\")\n",
+ " else:\n",
+ " raise Exception(file_name + \" does not exist\")\n",
+ " # LS3, LS4\n",
+ " case _:\n",
+ " file_name = f\"{selected_latent_space}-{selected_feature_model}-{selected_dim_reduction_method}-{k}-semantics.json\"\n",
+ " if os.path.exists(file_name):\n",
+ " data = json.load(open(file_name))\n",
+ " print(file_name + \" loaded\")\n",
+ " else:\n",
+ " raise Exception(file_name + \" does not exist\")\n"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 16,
+ "metadata": {},
+ "outputs": [
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "Using image-image similarity matrix from semantic data\n",
+ "Similarity graph created\n",
+ "Initialized pagerank scores\n"
+ ]
+ },
+ {
+ "ename": "KeyboardInterrupt",
+ "evalue": "",
+ "output_type": "error",
+ "traceback": [
+ "\u001b[1;31m---------------------------------------------------------------------------\u001b[0m",
+ "\u001b[1;31mKeyboardInterrupt\u001b[0m Traceback (most recent call last)",
+ "\u001b[1;32mc:\\Kaushik\\ASU\\CSE 515 - Multimedia and Web Databases\\Project\\Phase 2\\task_11.ipynb Cell 5\u001b[0m line \u001b[0;36m1\n\u001b[0;32m 6\u001b[0m img_graph \u001b[39m=\u001b[39m ImageGraph(fd_collection, \u001b[39mTrue\u001b[39;00m)\n\u001b[0;32m 7\u001b[0m img_graph\u001b[39m.\u001b[39mcreate_similarity_graph(\n\u001b[0;32m 8\u001b[0m n,\n\u001b[0;32m 9\u001b[0m selected_feature_model,\n\u001b[0;32m 10\u001b[0m data,\n\u001b[0;32m 11\u001b[0m selected_dim_reduction_method\n\u001b[0;32m 12\u001b[0m )\n\u001b[1;32m---> 13\u001b[0m imgs \u001b[39m=\u001b[39m img_graph\u001b[39m.\u001b[39;49mpersonalized_pagerank(l, m)\n",
+ "\u001b[1;32mc:\\Kaushik\\ASU\\CSE 515 - Multimedia and Web Databases\\Project\\Phase 2\\task_11.ipynb Cell 5\u001b[0m line \u001b[0;36m9\n\u001b[0;32m 93\u001b[0m \u001b[39m# for every node,\u001b[39;00m\n\u001b[0;32m 94\u001b[0m \u001b[39mfor\u001b[39;00m i \u001b[39min\u001b[39;00m \u001b[39mrange\u001b[39m(NUM_IMAGES):\n\u001b[0;32m 95\u001b[0m \u001b[39m# add sum of connected nodes' PR scores\u001b[39;00m\n\u001b[1;32m---> 96\u001b[0m pr_scores[i] \u001b[39m=\u001b[39m damping_factor \u001b[39m*\u001b[39m \u001b[39msum\u001b[39;49m(\n\u001b[0;32m 97\u001b[0m pr_scores[j]\n\u001b[0;32m 98\u001b[0m \u001b[39mfor\u001b[39;49;00m j \u001b[39min\u001b[39;49;00m \u001b[39mrange\u001b[39;49m(NUM_IMAGES)\n\u001b[0;32m 99\u001b[0m \u001b[39mif\u001b[39;49;00m (i \u001b[39m*\u001b[39;49m \u001b[39m2\u001b[39;49m, j \u001b[39m*\u001b[39;49m \u001b[39m2\u001b[39;49m) \u001b[39min\u001b[39;49;00m \u001b[39mself\u001b[39;49m\u001b[39m.\u001b[39;49msimilarity_graph\n\u001b[0;32m 100\u001b[0m \u001b[39m# and add the prob for random teleport *if node in given label*\u001b[39;49;00m\n\u001b[0;32m 101\u001b[0m ) \u001b[39m+\u001b[39m (\u001b[39m1\u001b[39m \u001b[39m-\u001b[39m damping_factor) \u001b[39m*\u001b[39m (\u001b[39m1\u001b[39m \u001b[39mif\u001b[39;00m i \u001b[39m*\u001b[39m \u001b[39m2\u001b[39m \u001b[39min\u001b[39;00m label_indices \u001b[39melse\u001b[39;00m \u001b[39m0\u001b[39m) \u001b[39m/\u001b[39m \u001b[39mlen\u001b[39m(\n\u001b[0;32m 102\u001b[0m label_indices\n\u001b[0;32m 103\u001b[0m )\n\u001b[0;32m 104\u001b[0m pr_scores \u001b[39m/\u001b[39m\u001b[39m=\u001b[39m \u001b[39msum\u001b[39m(pr_scores) \u001b[39m# Normalize\u001b[39;00m\n\u001b[0;32m 106\u001b[0m \u001b[39m# check for convergence\u001b[39;00m\n",
+ "\u001b[1;32mc:\\Kaushik\\ASU\\CSE 515 - Multimedia and Web Databases\\Project\\Phase 2\\task_11.ipynb Cell 5\u001b[0m line \u001b[0;36m9\n\u001b[0;32m 93\u001b[0m \u001b[39m# for every node,\u001b[39;00m\n\u001b[0;32m 94\u001b[0m \u001b[39mfor\u001b[39;00m i \u001b[39min\u001b[39;00m \u001b[39mrange\u001b[39m(NUM_IMAGES):\n\u001b[0;32m 95\u001b[0m \u001b[39m# add sum of connected nodes' PR scores\u001b[39;00m\n\u001b[0;32m 96\u001b[0m pr_scores[i] \u001b[39m=\u001b[39m damping_factor \u001b[39m*\u001b[39m \u001b[39msum\u001b[39m(\n\u001b[0;32m 97\u001b[0m pr_scores[j]\n\u001b[0;32m 98\u001b[0m \u001b[39mfor\u001b[39;00m j \u001b[39min\u001b[39;00m \u001b[39mrange\u001b[39m(NUM_IMAGES)\n\u001b[1;32m---> 99\u001b[0m \u001b[39mif\u001b[39;00m (i \u001b[39m*\u001b[39;49m \u001b[39m2\u001b[39;49m, j \u001b[39m*\u001b[39;49m \u001b[39m2\u001b[39;49m) \u001b[39min\u001b[39;49;00m \u001b[39mself\u001b[39;49m\u001b[39m.\u001b[39;49msimilarity_graph\n\u001b[0;32m 100\u001b[0m \u001b[39m# and add the prob for random teleport *if node in given label*\u001b[39;00m\n\u001b[0;32m 101\u001b[0m ) \u001b[39m+\u001b[39m (\u001b[39m1\u001b[39m \u001b[39m-\u001b[39m damping_factor) \u001b[39m*\u001b[39m (\u001b[39m1\u001b[39m \u001b[39mif\u001b[39;00m i \u001b[39m*\u001b[39m \u001b[39m2\u001b[39m \u001b[39min\u001b[39;00m label_indices \u001b[39melse\u001b[39;00m \u001b[39m0\u001b[39m) \u001b[39m/\u001b[39m \u001b[39mlen\u001b[39m(\n\u001b[0;32m 102\u001b[0m label_indices\n\u001b[0;32m 103\u001b[0m )\n\u001b[0;32m 104\u001b[0m pr_scores \u001b[39m/\u001b[39m\u001b[39m=\u001b[39m \u001b[39msum\u001b[39m(pr_scores) \u001b[39m# Normalize\u001b[39;00m\n\u001b[0;32m 106\u001b[0m \u001b[39m# check for convergence\u001b[39;00m\n",
+ "\u001b[1;31mKeyboardInterrupt\u001b[0m: "
+ ]
+ }
+ ],
+ "source": [
+ "if not given_ls:\n",
+ " img_graph = ImageGraph(fd_collection, True)\n",
+ " img_graph.create_similarity_graph(n, selected_feature_model)\n",
+ " imgs = img_graph.personalized_pagerank(l, m)\n",
+ "else:\n",
+ " img_graph = ImageGraph(fd_collection, True)\n",
+ " img_graph.create_similarity_graph(\n",
+ " n,\n",
+ " selected_feature_model,\n",
+ " data,\n",
+ " selected_dim_reduction_method\n",
+ " )\n",
+ " imgs = img_graph.personalized_pagerank(l, m)\n"
+ ]
+ },
+ {
+ "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": "3.10.5"
+ }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}
diff --git a/Phase 2/task_7.ipynb b/Phase 2/task_7.ipynb
index 0eaebd4..f61bed9 100644
--- a/Phase 2/task_7.ipynb
+++ b/Phase 2/task_7.ipynb
@@ -310,7 +310,7 @@
"\n",
" distances = []\n",
" for i in range(NUM_LABELS):\n",
- " if i != label:\n",
+ " if i != img_label:\n",
" distances.append(\n",
" {\n",
" \"label\": i,\n",
@@ -326,7 +326,7 @@
" similar_images = []\n",
" for i in range(len(dataset)):\n",
" _, l = dataset[i]\n",
- " if l == label:\n",
+ " if l == img_label:\n",
" similar_images.append(i)\n",
"\n",
" similar_images = random.sample(similar_images, k_2)\n",
diff --git a/Phase 2/utils.py b/Phase 2/utils.py
index cf7b7df..b3e7ef4 100644
--- a/Phase 2/utils.py
+++ b/Phase 2/utils.py
@@ -5,6 +5,7 @@ import random
import cv2
import numpy as np
from scipy.stats import pearsonr
+
# from scipy.sparse.linalg import svds
# from sklearn.decomposition import NMF
from sklearn.decomposition import LatentDirichletAllocation
@@ -355,11 +356,14 @@ def pearson_distance_measure(img_1_fd, img_2_fd):
# such that lower distance implies more similarity
return 0.5 * (1 - pearsonr(img_1_fd_reshaped, img_2_fd_reshaped).statistic)
+
def kl_divergence_measure(p, q):
+ p_f = p.flatten()
+ q_f = q.flatten()
# Avoid division by zero
epsilon = 1e-10
- return np.sum(p * np.log((p + epsilon) / (q + epsilon)))
+ return np.sum(p_f * np.log((p_f + epsilon) / (q_f + epsilon)))
valid_feature_models = {