recsys_image_embedding.py

img2vec = Img2Vec(cuda=False)
movie_df_with_plots_posters_filepath = f'{BASE_FOLDER}/MovieLens-1M/movie_df_with_plots_posters.csv'

if os.path.exists(movie_df_with_plots_posters_filepath):
  movies_df = pd.read_csv(movie_df_with_plots_posters_filepath)
else:
  def get_movie_embedding(url):
    try:
      response = requests.get(url)
      img = Image.open(BytesIO(response.content)).convert('RGB')
      embedding = img2vec.get_vec(img, tensor=True)
      return embedding.squeeze().numpy().astype(float).tolist()
    except (UnidentifiedImageError, OSError):
      print(f'failed URL {url}')
      return np.zeros(512).tolist()
  movies_df['image_embedding'] = movies_df['URL'].progress_apply(lambda url: get_movie_embedding(url))
  movies_df.to_csv(movie_df_with_plots_posters_filepath)
  
all_embeddings = np.vstack(movies_df['image_embedding'].values)
item_matrix_images_trainset_loocv = get_item_matrix_with_inner_ids(all_embeddings, movies_df, train_loocv)
cosine_sim_images_trainset_loocv = cosine_similarity(item_matrix_images_trainset_loocv,
                                                        item_matrix_images_trainset_loocv)
item_matrix_images_trainset = get_item_matrix_with_inner_ids(all_embeddings, movies_df, trainset)
cosine_sim_images_trainset = cosine_similarity(item_matrix_images_trainset, item_matrix_images_trainset)

get_algorithm_report(CustomSimKNNAlgorithm, trainset, testset, train_loocv, test_loocv, movies_df,
                     target_movie_id='movie_1', target_user_id='user_1', top_k=10,
                     algo_kwargs_trainset=dict(similarities=cosine_sim_images_trainset, sim_options={'user_based': False}),
                     algo_kwargs_trainset_loocv=dict(similarities=cosine_sim_images_trainset_loocv, sim_options={'user_based': False}))