Ray Tutorial

Setup

1. Install Ray by following the Ray installation instructions. Verify that the following works.

python ray/test/runtest.py

2. Clone the tutorial repository with

git clone https://github.com/ray-project/tutorial.git

3. Install the tutorial package with

cd tutorial
python setup.py install

4. Install the following additional dependencies.

pip install tensorflow
pip install gym

Verify that you can run import tensorflow and import gym in a Python interpreter.

Exercises

Each file exercises/exercise*.py is a separate exercise. Instructions are written in each file. Each file should run without raising any exceptions. Throughout these exercises, you may find the Ray documentation helpful.

Exercise 1: Define a remote function, and execute multiple remote functions in parallel.

Exercise 2: Execute remote functions in parallel with some dependencies.

Exercise 3: Pass object IDs into tasks to construct dependencies between tasks.

Exercise 4: Call remote functions from within remote functions.

Exercise 5: Use ray.wait to ignore stragglers. See the documentation for wait.

Exercise 6: Use ray.wait to process tasks in the order that they finish. See the documentation for wait.

Exercise 7: Use actors to share state between tasks. See the documentation on using actors.

Exercise 8: Use ray.put to avoid serializing and copying the same object into shared memory multiple times.

Exercise 9: Parallelize a serial example that uses a neural net to perform rollouts in a gym environment.

Exercise 10: Extract neural network weights from an actor on one process, and set them in another actor. You may want to read the documentation on using Ray with TensorFlow.

Exercise 11: Specify that an actor requires some GPUs. For a complete example that does something similar, you may want to see the ResNet example.

Longer Examples

Now that you've completed the basic exercises, some more involved exercises are provided under examples/.

Block Distributed Matrix Algorithms: Implement a distributed matrix multiplication algorithm for block distributed matrices. For this example, run python examples/block_linear_algebra.py.

Hyperparameter Search: Here, a serial implementation of random hyperparameter search is provided. Modify this example to parallelize the hyperparameter search and to start a new experiment whenever one finishes. For this example, run python examples/hyperparameter_search.py.

Evolutionary Algorithms: Here, a serial implementation of an evolutionary algorithm for solving reinforcement learning problems is provided. Modify this example to parallelize the implementation. For this example, run python examples/evolution_strategies.py.