movie-recs

Movie poster object detection with Ray and Anyscale

The following workload has been run on a ml.g4dn.12xlarge Sagemaker instance.

Requirements

# Local commit with changes needed, will be merged soon
# In case you're not using py3.7, get wheel links in https://buildkite.com/ray-project/ray-builders-pr/builds/41075#01825f93-238f-4376-8df2-0f60d5df468b
pip install https://ray-ci-artifact-pr-public.s3.amazonaws.com/c58874ae8545eef0b5c7632418eba3da0b5015c9/tmp/artifacts/.whl/ray-3.0.0.dev0-cp37-cp37m-manylinux2014_x86_64.whl
pip install ray[tune]
pip install torch
pip install torchvision
pip install tqdm # to visualize progress bar

PyTorch Serial

First download the images from s3

cd /home/ec2-user/SageMaker/movie-recs/movie_posters
mkdir images
aws s3 cp --recursive s3://waleed-movies ./images/

Then run python pytorch_serial.py

Running Ray locally on Sagemaker

python movie_poster_rec.py

This will

1. Read the Dataset from S3
2. Perform batch prediction with Ray AIR
3. Take the first 100 prediction results to visualize in the ./object_detections folder.

Running Ray on Anyscale via Workspaces

1. Set up Anyscale git access so you can clone the repo.
   • You can either create a Personal Access token on Github.
   • Or you can generate an ssh key on the Sagemaker instance and add it to your Github account
2. Configure the Personal Access Token/SSH key with Anyscale SSO
3. Clone the product repo git clone git@github.com:anyscale/product.git
4. Install Anyscale CLI

cd product/frontend/cli
pip install -e .

4. Create a workspace on Anyscale staging. Use the sagemaker-cluster-env:1 cluster env. Use 4 g4dn.12xlarge nodes.
5. Add Anyscale credentials to your Sagemaker instance
   Go to https://console.anyscale-staging.com/o/anyscale-internal/credentials and follow the instructions for setting environment variables. Seems like anyscale auth does not work for staging.
   For example

export ANYSCALE_HOST=https://console.anyscale-staging.com
export ANYSCALE_CLI_TOKEN=<INSERT_YOUR_CLI_TOKEN>

6. Clone the workspace in your Sagemaker Notebook. For example, if your workspace is called sagemaker-demo, then do anyscale workspace clone -n sagemaker-demo
7. Copy files to your workspace cp *.py workspace-project-sagemaker-demo
8. Run batch prediction

cd workspace-project-sagemaker-demo
anyscale workspace run "python movie-poster-rec.py"

9. If you access your workspace from Anyscale, you should see the first 100 object detections saved on the head node.

Running Ray on Anyscale via Anyscale Connect

Follow instructions 1-5 from the previous section.

Now do the following

1. Add the appropriate Ray address to your script.
   For example, add the following line to movie_poster_rec.py: ray.init("anyscale://workspace-project-sagemaker-demo/workspace-cluster-sagemaker-demo", runtime_env={"working_dir": "."})
2. Run python movie_poster_rec.py
3. You should see the 100 object detections saved on the locally on the Sagemaker notebook.