Merge pull request #2 from adiIspas/lightfm-version

Lightfm version

Author: adiIspas

dataset/utils/collect_posters_in_clusters.py

@@ -0,0 +1,57 @@
+import os
+import pandas as pd
+import csv
+import shutil
+
+dataset = '../../../king-rec-dataset/ml-latest-small/images/'
+number_of_clusters = 7
+model = 'vgg16'
+clusters_dir = '../../../king-rec-dataset/ml-latest-small/results/clusters/sanity-check/' + model + '/' + str(number_of_clusters) + '/'
+
+
+def collect_posters():
+    data = pd.read_csv('sanity_check_movies_1_poster_clusters_' + model + '.csv')
+
+    # create directories
+    for idx in range(1, number_of_clusters + 1):
+        os.makedirs(clusters_dir + str(idx), exist_ok=True)
+
+    # move posters into associated cluster
+    for index, row in data.iterrows():
+        src = dataset + str(int(row['0'])) + '/posters/' + str(int(row['1'])) + '.jpg'
+        dest = clusters_dir + str(int(row['cluster_' + str(number_of_clusters)]) + 1) + '/' + str(int(row['0'])) + '_' + str(int(row['1'])) + '.jpg'
+
+        if os.path.isfile(src):
+            shutil.copy(src, dest)
+
+    print('Done')
+
+
+collect_posters()
+
+dataset2 = '../../../king-rec-dataset/ml-latest-small/'
+
+
+def get_items_ids():
+    item_ids = set()
+
+    with open(dataset2 + 'movies.csv', 'r') as movies_file:
+        reader = csv.reader(movies_file, delimiter=',')
+        next(reader)  # skip header
+
+        for row in reader:
+            item_ids.add(int(row[0]))
+
+    return item_ids
+
+
+def count_movies():
+    movies = get_items_ids()
+
+    idx = 1
+    for item in movies:
+        print(idx, item)
+        idx = idx + 1
+
+
+# count_movies()

dataset/utils/create_clusters.py

@@ -0,0 +1,88 @@
+import numpy as np
+import matplotlib.pyplot as plt
+import csv
+import pandas as pd
+
+from sklearn import metrics
+from sklearn.cluster import KMeans
+
+
+dataset = '../../../king-rec-dataset/ml-latest-small/'
+
+
+def get_items_ids():
+    item_ids = set()
+
+    with open(dataset + 'movies.csv', 'r') as movies_file:
+        reader = csv.reader(movies_file, delimiter=',')
+        next(reader)  # skip header
+
+        for row in reader:
+            item_ids.add(int(row[0]))
+
+    return item_ids
+
+
+def explore_clusters():
+    clusters = range(2, 22, 2)
+    models_results = dict()
+    colors = ['r', 'y', 'b', 'g', 'c']
+
+    models = ['vgg16', 'vgg19', 'inception_v3', 'resnet50', 'NASNet']
+
+    for model in models:
+        print('Reading data ...')
+        feature_list = np.loadtxt('./posters_features/1000-movies/' + model + '1000-movies_1-posters.csv', delimiter=',')
+        print('Complete read data.')
+
+        movie_poster_clusters = pd.DataFrame(feature_list[:, :2])
+
+        feature_list = feature_list[:, 2:]
+        feature_list_np = np.array(feature_list)
+        for n_clusters in clusters:
+            k_means = KMeans(n_clusters=n_clusters).fit(feature_list_np)
+
+            name = model
+            result = metrics.silhouette_score(feature_list_np, k_means.labels_)
+
+            if name not in models_results:
+                results = []
+            else:
+                results = models_results.pop(name)
+
+            cluster_name = 'cluster_' + str(n_clusters)
+            movie_poster_clusters[cluster_name] = pd.Series(k_means.labels_)
+
+            results.append(result)
+            models_results.update({name: results})
+            print('silhouette score on', name, 'with', n_clusters, 'clusters:', result)
+
+        movie_poster_clusters.to_csv('movies_1_poster_clusters_' + name + '.csv')
+
+    n_groups = len(list(clusters))
+    index = np.arange(n_groups)
+    bar_width = 0.15
+    current_index = 0
+
+    for key, values in models_results.items():
+        plt.bar(index + bar_width * current_index, values, bar_width,
+                color=colors[current_index],
+                label=key)
+        current_index += 1
+
+    plt.xlabel('Number of clusters')
+    plt.ylabel('Silhouette score')
+    plt.title('Silhouette score by model')
+    plt.xticks(index + bar_width, list(clusters))
+    plt.legend()
+    plt.tight_layout()
+    plt.savefig('silhouette-score.jpg')
+    plt.show()
+
+
+def main():
+    explore_clusters()
+
+
+if __name__ == "__main__":
+    main()

dataset/utils/create_clusters_minibatch.py

@@ -0,0 +1,80 @@
+import csv
+import numpy as np
+import pandas as pd
+
+from sklearn.cluster import MiniBatchKMeans
+
+
+dataset = '../../../king-rec-dataset/ml-latest-small/'
+
+
+def get_items_ids():
+    item_ids = set()
+
+    with open(dataset + 'movies.csv', 'r') as movies_file:
+        reader = csv.reader(movies_file, delimiter=',')
+        next(reader)  # skip header
+
+        for row in reader:
+            item_ids.add(int(row[0]))
+
+    return item_ids
+
+
+def explore_clusters():
+    batch_size = 40
+
+    # models = ['vgg16', 'vgg19', 'inception_v3', 'resnet50', 'NASNet']
+    models = ['resnet50']
+
+    for model in models:
+        # csv_path = './' + model + '-1-posters.csv'
+        csv_path = './posters_features/sanity-check/' + model + '-sanity-check.csv'
+
+        movie_poster_clusters = pd.DataFrame()
+        for n_clusters in [7]:
+            final_clusters = pd.Series()
+            print('Process cluster', n_clusters)
+
+            k_means = MiniBatchKMeans(n_clusters=n_clusters, batch_size=batch_size, compute_labels=True)
+
+            reader_chunks = pd.read_csv(csv_path, delimiter=',', header=None, chunksize=batch_size)
+            for chunk in reader_chunks:
+                print('Processing chunk ...')
+
+                feature_list = pd.DataFrame(data=chunk)
+
+                movie_poster_clusters = movie_poster_clusters.append(feature_list.iloc[:, :2])
+
+                feature_list = feature_list.iloc[:, 2:]
+                feature_list_np = np.array(feature_list)
+
+                k_means.partial_fit(feature_list_np)
+
+            reader_chunks = pd.read_csv(csv_path, delimiter=',', header=None, chunksize=batch_size)
+            for chunk in reader_chunks:
+                print('Predicting chunk ...')
+
+                feature_list = pd.DataFrame(data=chunk)
+
+                feature_list = feature_list.iloc[:, 2:]
+                feature_list_np = np.array(feature_list)
+
+                labels = k_means.predict(feature_list_np)
+
+                final_clusters = final_clusters.append(pd.Series(labels))
+
+            name = model
+
+            cluster_name = 'cluster_' + str(n_clusters)
+            movie_poster_clusters[cluster_name] = pd.Series(final_clusters.values, index=movie_poster_clusters.index)
+
+        movie_poster_clusters.to_csv('test-chunk-movies_1_poster_clusters_' + name + '.csv')
+
+
+def main():
+    explore_clusters()
+
+
+if __name__ == "__main__":
+    main()

dataset/utils/downloader.py

@@ -0,0 +1,85 @@
+import os
+import csv
+import sys
+import requests
+import urllib.request
+
+api_key = sys.argv[1]
+
+dataset = '../../king-rec-dataset/ml-latest-small/'
+tmdb_api = 'https://api.themoviedb.org/3/movie/$MOVIE_ID/images?include_image_language=en,null&api_key=$API_KEY'
+tmdb_images_url = 'https://image.tmdb.org/t/p/original/'
+
+
+def get_tmdb_posters(tmdb_api_key, max_movie_index=10):
+    tmdb_movies_id = get_tmdb_ids()
+    download_images(tmdb_api_key, tmdb_movies_id, max_movie_index)
+
+    return tmdb_movies_id
+
+
+def download_images(tmdb_api_key, tmdb_movies_id, max_movie_index=10):
+    images = dataset + 'images/'
+
+    movie_index = 1
+    total_movies = len(tmdb_movies_id.items())
+
+    for key, value in tmdb_movies_id.items():
+        posters = images + str(key) + '/posters/'
+        backdrops = images + str(key) + '/backdrops/'
+
+        if not os.path.exists(posters):
+            os.makedirs(posters)
+
+        if not os.path.exists(backdrops):
+            os.makedirs(backdrops)
+
+        if len(os.listdir(posters)) == 0 or len(os.listdir(backdrops)) == 0:
+            current_url = tmdb_api.replace('$MOVIE_ID', str(value)).replace('$API_KEY', tmdb_api_key)
+            response = requests.get(current_url)
+
+            if response.status_code == 200:
+                json = response.json()
+
+                if len(os.listdir(posters)) == 0:
+                    image_idx = 1
+                    for poster in json['posters']:
+                        if poster['iso_639_1'] == 'en':
+                            print(movie_index, '/', total_movies, '- Process movie', value, 'and poster', image_idx)
+                            poster_url = poster['file_path']
+                            urllib.request.urlretrieve(tmdb_images_url + poster_url, posters + str(image_idx) + '.jpg')
+                            image_idx += 1
+
+                if len(os.listdir(backdrops)) == 0:
+                    image_idx = 1
+                    for backdrop in json['backdrops']:
+                        if backdrop['iso_639_1'] == 'xx' or backdrop['iso_639_1'] is None:
+                            print(movie_index, '/', total_movies, '- Process movie', value, 'and backdrop', image_idx)
+                            backdrop_url = backdrop['file_path']
+                            urllib.request.urlretrieve(tmdb_images_url + backdrop_url,
+                                                       backdrops + str(image_idx) + '.jpg')
+                            image_idx += 1
+
+            else:
+                print('Status code:', response.status_code, 'on movie', key, '-', value)
+
+        if movie_index == max_movie_index:
+            break
+
+        movie_index += 1
+
+
+def get_tmdb_ids(tmdb_index=2):
+    links = dataset + 'links.csv'
+    with open(links, 'r') as links_file:
+        reader = csv.reader(links_file, delimiter=',', )
+        next(reader)  # skip header
+
+        tmdb_movies_id = dict()
+        for row in reader:
+            tmdb_movies_id.update({row[0]: row[tmdb_index]})
+
+    return tmdb_movies_id
+
+
+get_tmdb_posters(api_key, max_movie_index=20)

dataset/utils/extract_features.py

@@ -0,0 +1,81 @@
+import os
+import numpy as np
+import csv
+import pandas as pd
+
+from keras.layers import Input
+from keras.preprocessing import image
+from keras.applications.vgg16 import VGG16
+from keras.applications.vgg19 import VGG19
+from keras.applications.inception_v3 import InceptionV3
+from keras.applications.resnet50 import ResNet50
+from keras.applications.nasnet import NASNetLarge
+from keras.applications.imagenet_utils import preprocess_input
+
+dataset = '../../../king-rec-dataset/ml-latest-small/'
+base_path = 'images/'
+# base_path = 'clusters_sanity_check/'
+max_posters_per_movie = 1
+
+
+def get_int(filename):
+    return int(filename.split('.')[0])
+
+
+def get_items_ids():
+    item_ids = set()
+
+    with open(dataset + 'movies.csv', 'r') as movies_file:
+        reader = csv.reader(movies_file, delimiter=',')
+        next(reader)  # skip header
+
+        for row in reader:
+            item_ids.add(int(row[0]))
+
+    return item_ids
+
+
+def extract_images_features():
+    movies = list(get_items_ids())
+    # movies = [1, 3, 4, 5, 7, 19, 22, 23]
+    subdir = [dataset + base_path + str(movie) + '/posters/' for movie in movies]
+    models = [
+              VGG16(weights='imagenet', include_top=False),
+              VGG19(weights='imagenet', include_top=False),
+              InceptionV3(weights='imagenet', include_top=False),
+              ResNet50(weights='imagenet', include_top=False),
+              NASNetLarge(weights='imagenet', include_top=False, input_tensor=Input(shape=(224, 224, 3)))
+             ]
+    total_movies = len(subdir)
+    for current_movie, dirname in enumerate(subdir):
+        movie_idx = int([s for s in dirname.split('/') if s.isdigit()][0])
+        filenames = sorted(os.listdir(dirname), key=get_int)[0:max_posters_per_movie]
+
+        for _, file_name in enumerate(filenames):
+            poster_idx = int(file_name.split('.')[0])
+
+            img = image.load_img(dirname + '/' + file_name, target_size=(224, 224))
+            img_data = image.img_to_array(img)
+            img_data = np.expand_dims(img_data, axis=0)
+            img_data = preprocess_input(img_data)
+
+            for model in models:
+                feature = model.predict(img_data)
+                feature_np = np.array(feature)
+                feature = feature_np.flatten()
+
+                data_to_save = np.append([movie_idx, poster_idx], feature)
+                data = pd.DataFrame([data_to_save])
+                data.to_csv(model.name + '-' + str(max_posters_per_movie) + '-posters' + '.csv',
+                            mode='a', sep=',', index=False, header=False)
+
+                print(str(current_movie + 1) + '/' + str(total_movies) + ':', 'movie id:', movie_idx, '  poster id:', poster_idx,
+                      '  model name:', model.name, '  total features:', len(feature))
+
+
+def main():
+    extract_images_features()
+
+
+if __name__ == "__main__":
+    main()