Table of Contents
Was this article helpful?
Try Vultr Today with

$50 Free on Us!

Want to contribute?

You could earn up to $600 by adding new articles.

How to Build Recommendation System using BERT and NVIDIA NGC

Author: Mirdul Swarup

Last Updated: Mon, Oct 16, 2023
Machine Learning Vultr Cloud GPU

Header Image


Bidirectional Encoder Representations from Transformers (BERT) is a pre-trained Natural Language Processing Model (NLP) used to understand and generate human language text. Its bidirectional approach allows it to capture the meaning of words more comprehensively. NVIDIA GPU Cloud (NGC) offers a variety of GPU-accelerated containers that take maximum advantage of NVIDIA GPUs and come pre-installed with GPU drivers, libraries, and frameworks that improve development efficiency.

This article explains how to use the NGC PyTorch container with the BERT model to make a movie recommendation system by generating embeddings on a Vultr Cloud GPU Server.


Before you begin:

Install the PyTorch GPU Container and Access Jupyter Notebook

  1. Install and run the PyTorch GPU container

    $ sudo docker run --gpus all -p 9000:8888 -it

    The above command installs and runs the GPU-accelerated docker container with the following options:

    • --gpus all: Allocates all available host GPUs to the container, this makes sure that all GPU resources handle GPU-accelerated tasks

    • -p 9000:8888: Maps the host port 9000 to the container port 8888. This creates a separate Jupyter Notebook instance difference from the Jupyter Lab service available on the host machine.

    • --it: Run the container in interactive mode

    When successful, verify that you can access the container root shell:

    root@workspace #
  2. Start a new Jupyter Notebook instance

    # jupyter notebook --ip=


        To access the notebook, open this file in a browser:
    Or copy and paste this URL:

    Copy the generated access token to securely access Jupyter Notebook in your browser session.

  3. Using a web browser such as Chrome, access Jupyter Notebook using your generated access token


Run the BERT Model

To run the BERT model on your Cloud GPU server, initialize the BERT model. Then, import the necessary modules, prepare the movie data and user data, create embeddings, and calculate similarity scores to retrieve the top movie suggestions based on user preferences as described in the steps below.

  1. Access your Jupyter Notebook web interface

  2. In the middle top right corner, click the New dropdown to reveal a list of options

    Create a new Jupyter Notebook

  3. Click Notebook, and select Python 3 (ipykernel) to open a new file

  4. Within the new Notebook file, install transformers dependency package

    !pip install transformers

    The above command installs transformers package that simplifies the process of working with NLP models.

  5. Click Run or press CTRL + ENTER to run the code cell

  6. Import the necessary modules

    import torch
    from transformers import BertTokenizer, BertModel
    import numpy as np

    The above command imports the following modules:

    • BertTokenizer for tokenization

    • BertModel for inference

    • numpy for mathematical calculations

  7. Initialize the model

    model_name = "bert-base-uncased"
    tokenizer = BertTokenizer.from_pretrained(model_name)
    model = BertModel.from_pretrained(model_name)

    The above command initializes the bert-base-uncased model and its tokenizer

  8. Prepare the movies data

    movie_data = [
        {"name": "Avengers", "genres": ["Action", "Adventure"]},
        {"name": "Social Network", "genres": ["Suspense", "Drama"]},
        {"name": "Intersteller", "genres": ["Sci-Fi", "Suspense", "Thriller"]},
        {"name": "Iron Man", "genres": ["Action", "Drama", "Adventure"]},
        {"name": "Lights Out", "genres": ["Horror", "Mystery"]},

    The above command prepares a dataset of movies along with the genres the movies belong to. You can add as many movies as you require and genre tags for a particular movie to match your needs.

  9. Prepare the user profiles

    user_preferences = {
    "user1": ["Action", "Adventure", "Sci-Fi"],
    "user2": ["Drama", "Suspense", "Mystery"],
    "user3": ["Adventure", "Horror"],

    The above command creates three user profiles along with preferences for each user. This assists the model in matching which movies are best suitable to suggest by calculating the similarity of the embeddings you created earlier.

  10. Encode the movie names

    encoded_data = [tokenizer.encode(movie["name"], add_special_tokens=True) for movie in movie_data]

    The above command processes each movie name through a tokenizer to represent it numerically. Then, it stores the numerical data in a list of encoded movie names encoded_data.

  11. Generate embeddings from the encoded data

    embeddings = []
    for input_ids in encoded_data:
        input_ids = torch.tensor(input_ids).unsqueeze(0)
        with torch.no_grad():
            last_hidden_states = model(input_ids)[0]
        avg_embedding = torch.mean(last_hidden_states, dim=1)

    The above command processes each encoded movie name through a PyTorch model for computing average embeddings and stores them in the embeddings list. The list is then used in the similarity calculations.

  12. Calculate the similarity score

    def genre_similarity(user_preferences, movie_genres):
        intersection = set(user_preferences).intersection(set(movie_genres))
        return len(intersection) / (len(user_preferences) + len(movie_genres) - len(intersection))

    The above command defines a genre_similarity function that calculates the similarity score between a user's genre preference and a movie's genre. The score represents the overlap between a user's preference and the movie genres.

  13. Select the user

    target_user = input("Specify User: ")
    user_genre_preferences = user_preferences[target_user]
    recommendations = []

    The above command sets a condition that prompts you to enter the user's name you want to retrieve the movie recommendations.

  14. Generate movie recommendations

    for movie_index, movie_info in enumerate(movie_data):
        movie_genres = movie_info["genres"]
        similarity = genre_similarity(user_genre_preferences, movie_genres)
        recommendations.append((movie_info["name"], similarity))

    The above command generates movie recommendations for the specified user based on the similarity between the user's genre preference and the genre for each movie. Then, all the recommendations get stored in the recommendations list.

  15. Print the results

    recommendations.sort(key=lambda x: x[1], reverse=True)
    top_n = int(input("# of Recommendations: "))
    print(f"\nTop {top_n} recommendations for '{target_user}' based on genre similarity:\n")
    for recommendation, similarity in recommendations[:top_n]:
        print(f"Name: {recommendation}, Similarity: {similarity:.4f}")

    The above command sorts the movie recommendations in an ascending list based on similarity scores. Then it prompts you to enter the number of recommendations you want and prints the top_n number of movies along with similarity scores.


You have built a BERT-based movie recommendation system on a Vultr Cloud GPU server. You created the system using embeddings and calculating similarity scores between different movie genres and user preferences within the PyTorch NGC GPU accelerated container. For more information, visit the following documentation resources:

Want to contribute?

You could earn up to $600 by adding new articles.