Skip to content

Latest commit

 

History

History
 
 

mnist-keras

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
app-data
app-data
 
 
bin
bin
 
 
model
model
 
 
README.md
README.md
 
 

README.md

mnist-keras

This example runs decentralized training on MNIST 1 dataset for digit classification using TensorFlow based Swarm Learning framework.

The code for this example has been taken from 2 and modified to run on a Swarm Learning platform. The data files are in the swarm-learning-docs/examples/mnist-keras/app-data/ directory. The model program, after conversion to Swarm Learning, is in swarm-learning-docs/examples/mnist-keras/model/ and is called mnist_tf.py.

Scripts to run the example are in the swarm-learning-docs/examples/mnist-keras/bin/ directory:

• init-workspace: It creates workspace for the given example by distributing data and models in different directories It also initializes single node Swarm Network and starts spire server

• run-sl: It starts Swarm Learning node that runs training

• del-workspace: It deletes the created workspace and also stops all containers

Following environment variables are required to set in each terminal:

  • APLS_IP: IP address of the host where license server is running.
  • EXAMPLE: Name of the quick start example.
  • WORKSPACE_DIR: Path where workspace to be created. Separate model and data directories will be created for each training node inside the example workspace. Default is current directory.
  • TRAINING_NODE: Unique name of each training node prefixed with ‘node’ e.g. ‘node1’ etc.

Steps to run:

  1. Create workspace and initialize Swarm Learning First open a terminal, change directory to ‘swarm-learning-docs/examples’ and set the following environment variables. Then run ‘init-workspace’ script to create workspace for ‘mnist-keras’ example and initialize Swarm Learning environment by running Swarm Network and Spire server containers.

    APLS_IP=

    EXAMPLE=mnist-keras

    WORKSPACE_DIR=$PWD

    ./mnist-keras/bin/init-workspace -e $EXAMPLE -i $APLS_IP -d $WORKSPACE_DIR

    Once command completes, two separate folders for each training node will be created under the workspace directory. Data and model files will be distributed among them.

    mnist-keras-workspace

  2. Run decentralized training with different nodes By default two peer nodes will be created. So open two new terminals each representing individual training node.

  • Node1: Training on Terminal 1 Change directory to ‘swarm-learning-docs/examples’ and set environment variables as specified below. Then run Swarm Learning container to start training as shown below. Specify ‘--gpu ’ in ‘run-sl’ command if system has GPUs. Otherwise training will run on CPU.

    APLS_IP=

    EXAMPLE=mnist-keras

    WORKSPACE_DIR=$PWD

    TRAINING_NODE=node1

    ./mnist-keras/bin/run-sl --name $TRAINING_NODE-sl --network $EXAMPLE-net --host-ip $TRAINING_NODE-sl --sn-ip node-sn -e MAX_EPOCHS=5 --apls-ip $APLS_IP --serverAddress node-spire -genJoinToken --data-dir $WORKSPACE_DIR/ws-$EXAMPLE/$TRAINING_NODE/app-data --model-dir $WORKSPACE_DIR/ws-$EXAMPLE/$TRAINING_NODE/model --model-program mnist_tf.py --sl-platform TF

  • Node2: Training on Terminal 2 Change directory to ‘swarm-learning-docs/examples’ and set environment variables as specified below. Then run Swarm Learning container to start training as shown below. Specify ‘--gpu ’ in ‘run-sl’ command if system has GPUs. Otherwise training will run on CPU.

    APLS_IP=

    EXAMPLE=mnist-keras

    WORKSPACE_DIR=$PWD

    TRAINING_NODE=node2

    ./mnist-keras/bin/run-sl --name $TRAINING_NODE-sl --network $EXAMPLE-net --host-ip $TRAINING_NODE-sl --sn-ip node-sn -e MAX_EPOCHS=5 --apls-ip $APLS_IP --serverAddress node-spire -genJoinToken --data-dir $WORKSPACE_DIR/ws-$EXAMPLE/$TRAINING_NODE/app-data --model-dir $WORKSPACE_DIR/ws-$EXAMPLE/$TRAINING_NODE/model --model-program mnist_tf.py --sl-platform TF

    This training process will continue till training completes MAX_EPOCHS. Once training ends, the final Swarm Learning model will be saved inside the model folder for each node in workspace.

  1. Delete workspace and remove Swarm Learning containers Go back to the initial terminal where ‘init-workspace’ script was run, or open a new terminal, change directory to ‘swarm-learning-docs/examples’ and set environment variables as specified in step 1. Run ‘del-workspace’ script that will delete the workspace and remove all the containers started for this example. The script should be run as sudo.

    sudo ./mnist-keras/bin/del-workspace -e $EXAMPLE -d $WORKSPACE_DIR

References

[1] Y. LeCun, C. Cortes and C. J. Burges, "THE MNIST DATABASE," [Online]. Available: http://yann.lecun.com/exdb/mnist/

[2] https://www.tensorflow.org/tutorials/quickstart/beginner