SigOpt Orchestrate Reference
SigOpt Orchestrate is a command-line tool for managing training clusters and running optimization experiments.
Cluster Configuration File Back to Top
The cluster configuration file is commonly referred to as
cluster.yml, but you can name yours anything you like. The file is used when we create a SigOpt Orchestrate cluster, with
orchestrate cluster create -f cluster.yml. You can update your cluster configuration file after the cluster has been created to change the number of nodes in your cluster or change instance types. These changes can be applied by running
orchestrate cluster update -f cluster.yml. Some updates might not be supported, for example introducing GPU nodes to your cluster in some regions. If the update is not supported then you will need to destroy the cluster and create it again.
The available fields are:
|Y||You must provide at least one of either |
|Y||You must provide a name for your cluster. You will share this with anyone else who wants to connect to your cluster.|
|N||Override environment-provided values for |
|N||The version of Kubernetes to use for your cluster. Currently supports Kubernetes 1.16, 1.17, 1.18, and 1.19. Defaults to the latest stable version supported by SigOpt Orchestrate, which is currently 1.18.|
|N||Currently, AWS is our only supported provider for creating clusters. You can, however, use a custom provider to connect to your own Kubernetes cluster with the |
|N||System nodes are required to run the autoscaler. You can specify the number and type of system nodes with |
The example YAML file below defines a CPU cluster named
tiny-cluster with two
t2.small AWS instances.
# cluster.yml # AWS is currently our only supported provider for cluster create # You can connect to custom clusters via orchestrate connect provider: aws # We have provided a name that is short and descriptive cluster_name: tiny-cluster # Your cluster config can have CPU nodes, GPU nodes, or both. # The configuration of your nodes is defined in the sections below. # (Optional) Define CPU compute here cpu: # AWS instance type instance_type: t2.small max_nodes: 2 min_nodes: 0 # # (Optional) Define GPU compute here # gpu: # # AWS GPU-enabled instance type # # This can be any p* instance type # instance_type: p2.xlarge # max_nodes: 2 # min_nodes: 0 kubernetes_version: '1.18'
Configure training orchestration Back to Top
The SigOpt Orchestrate configuration file tells SigOpt Orchestrate how to setup and run the model, which metrics to track, as well as details about which hyperparameters to tune.
You can use a SigOpt Orchestrate config YAML file you've already created, or SigOpt Orchestrate will auto-generate an
orchestrate.yml template file for you if you run the following:
The available fields are:
|Y||Name of Docker container SigOpt Orchestrate creates for you. You can also point this to an existing Docker container to use for SigOpt Orchestrate.|
|Y||Name for your training run or HPO experiment|
|Y||Model file to execute|
|N||Resources required by your training run or HPO experiment. Can specify: |
Enables an HPO experiment. Requires at least one metric
When specifying CPUs, valid amounts are whole numbers (1, 2), and fractional numbers or millis (1.5 and 1500m both represent 1.5 CPU). When specifying memory, valid amounts are shown in the Kubernetes documentation for memory resources, but some examples are 1e9, 1Gi, 500M. For gpus, only whole numbers are valid.
When choosing the resources for a single model training run, it's important to keep in mind that some resources on your cluster will be auto-reserved for Kubernetes processes. For this reason, you must specify fewer resources for your model than are available on each node. A good rule of thumb is to assume that your node will have 0.5 CPU less than the total to run your model.
For example, if your nodes have 8 CPUs then you must specify fewer than 8 CPUs in the requests section of your
resources_per_model in order for your model to run. Keep in mind that you can specify fractional amounts of CPU, e.g. 7.5 or 7500m.
Here's an example SigOpt Orchestrate configuration file.
# orchestrate.yml resources_per_model: requests: cpu: 0.5 memory: 512Mi limits: cpu: 1 memory: 512Mi # We don't need any GPUs for this example, so we'll leave this commented out # gpus: 1 # Choose a descriptive name for your model name: Orchestrate SGD Classifier (python) # Model run command run: python model.py # Optimization details optimization: # Every experiment needs at least one named optimized metric. metrics: - name: accuracy strategy: optimize objective: maximize # Parameter that are defined here are available to SigOpt Orchestrate parameters: - name: l1_ratio type: double bounds: min: 0 max: 1.0 - name: log_alpha type: double bounds: min: -5 max: 2 # Our example cluster has two machines, so we have enough compute power # to execute two models in parallel. parallel_bandwidth: 2 # We want to evaluate our model on sixty different sets of hyperparameters observation_budget: 60 # SigOpt Orchestrate creates a container for your model. Since we're using an AWS # cluster, it's easy to securely store the model in the Amazon Elastic Container Registry. # Choose a descriptive and unique name for each new experiment configuration file. image: orchestrate/sgd-classifier
SigOpt Orchestrate Commands Back to Top
The best way to learn the most up to date information about cluster commands is from the command line interface (CLI) itself! Append any command with
--help to learn about sub commands, arguments, and flags.
For example, to learn more about all SigOpt Orchestrate commands, run:
To learn more about the specific
orchestrate optimize command, run:
orchestrate optimize --help
|Clean up SigOpt Orchestrate Docker images to free up disk space.|
|SigOpt Orchestrate will connect to an existing Kubernetes cluster on the specified cloud provider.|
|Optional. Defaults to checking |
|Valid inputs are: |
|Optional. Defaults to docker.io on a custom cluster, or to ECR for clusters created with SigOpt Orchestrate.|
|AWS only. Creates a cluster using the specifications from the config YAML to launch a Kubernetes cluster.|
|Destroy a cluster. At this time, only AWS clusters can be destroyed.|
|Specify the cluster to destroy.|
|Currently only supports: |
|Disconnect from a cluster. Requires one of the following flags.|
|Specify the cluster to disconnect.|
|Disconnect from all clusters.|
|See the status of all your training runs and HPO experiments.|
|Checks SigOpt Orchestrate cluster connection.|
|AWS only. Updates the cluster configuration with the changes in the config yaml. Not supported for all operations, ex. introducing GPU nodes to your cluster in some regions.|
|Used to configure the SigOpt Orchestrate CLI. You will need your|
|Creates a SigOpt Orchestrate configuration YAML file and Dockerfile that can be modified as needed.|
|Executes an HPO experiment on your Kubernetes cluster using SigOpt Orchestrate.|
|Executes a training run on your Kubernetes cluster using SigOpt Orchestrate.|
|See the status of all your training runs and HPO experiments. Requires at least one argument. A run ID is formatted as |
|Stops and archives an experiment or a run on the cluster. If an experiment is stopped, all in-progress runs will be halted. Objects will still exist on sigopt.com.|
|Stops an HPO experiment.|
|Stops a run. Valid inputs for the id argument are: |
|Tests the execution of a single training run on your Kubernetes cluster using SigOpt Orchestrate. Outputs debugging information and logs from your training run.|
Adding Additional AWS Policies Back to Top
Users creating AWS clusters with SigOpt Orchestrate can easily interface with different AWS services. To allow your cluster permission to access different AWS services, provide additional AWS policies in the
aws.additional_policies section of the cluster configuration file.
SigOpt Orchestrate Logging Back to Top
SigOpt Orchestrate integrates seamlessly with the SigOpt API to optimize the hyperparameters of your model. SigOpt Orchestrate is built to handle communication with theSigOpt API under the hood, so that you only need to focus on your model, some lightweight installation requirements, and your experiment configuration file.
As you write your model, use a few lines of code from the
sigopt package to read hyperparameters and write your model's metric(s).
Below is a side-by-side comparison of two nearly-identical
Multilayer Perceptron models. The right model is written for SigOpt Orchestrate, the left model is not. As you can see, the right model uses
sigopt.get_parameter to read assignments from SigOpt Orchestrate, as well as
sigopt.log_metric to send its metric value back to SigOpt Orchestrate.
import numpy import keras from keras.models import Sequential from keras.layers import Dense, Dropout, Activation from keras.optimizers import SGD x_train = numpy.random.random((1000, 20)) y_train = keras.utils.to_categorical( numpy.random.randint(10, size=(1000, 1)), num_classes=10, ) x_test = numpy.random.random((100, 20)) y_test = keras.utils.to_categorical( numpy.random.randint(10, size=(100, 1)), num_classes=10, ) dropout_rate = 0.5 model = Sequential() model.add(Dense( units=64, activation='relu', input_dim=20, )) model.add(Dropout(dropout_rate)) model.add(Dense( units=64, activation='relu', )) model.add(Dropout(dropout_rate)) model.add(Dense(10, activation='softmax')) sgd = SGD( lr=0.01, decay=1e-6, momentum=0.9, nesterov=True, ) model.compile( loss='categorical_crossentropy', optimizer=sgd, metrics=['accuracy'], ) model.fit( x=x_train, y=y_train, epochs=20, batch_size=128, ) evaluation_loss, accuracy = model.evaluate( x=x_test, y=y_test, batch_size=128, ) print('evaluation_loss:', evaluation_loss) print('accuracy:', accuracy)
import numpy import keras from keras.models import Sequential from keras.layers import Dense, Dropout, Activation from keras.optimizers import SGD import sigopt x_train = numpy.random.random((1000, 20)) y_train = keras.utils.to_categorical( numpy.random.randint(10, size=(1000, 1)), num_classes=10, ) x_test = numpy.random.random((100, 20)) y_test = keras.utils.to_categorical( numpy.random.randint(10, size=(100, 1)), num_classes=10, ) dropout_rate = sigopt.get_parameter('dropout_rate', default=0.5) model = Sequential() model.add(Dense( units=sigopt.get_parameter('hidden_1', default=64), activation=sigopt.get_parameter('activation_1', default='relu'), input_dim=20, )) model.add(Dropout(dropout_rate)) model.add(Dense( units=sigopt.get_parameter('hidden_2', default=64), activation=sigopt.get_parameter('activation_2', default='relu'), )) model.add(Dropout(dropout_rate)) model.add(Dense(10, activation='softmax')) sgd = SGD( lr=10**sigopt.get_parameter('log_learning_rate', default=-2), decay=10**sigopt.get_parameter('log_decay', default=-6), momentum=sigopt.get_parameter('momentum', default=0.9), nesterov=True, ) model.compile( loss=sigopt.get_parameter('loss', default='categorical_crossentropy'), optimizer=sgd, metrics=['accuracy'], ) model.fit( x=x_train, y=y_train, epochs=sigopt.get_parameter('epochs', default=20), batch_size=sigopt.get_parameter('batch_size', default=128), ) evaluation_loss, accuracy = model.evaluate( x=x_test, y=y_test, batch_size=128, ) sigopt.log_metric('evaluation_loss', evaluation_loss) sigopt.log_metric('accuracy', accuracy)
sigopt.get_parameter(name, default=None): Read a single assignment for a parameter by name. If you setup your parameter in the parameters section of your experiment configuration file, this value will be a new assignment to try. If the parameter is not included in the experiment configuration file, the default will be returned. The default values will also be used if you execute your model file directly, without using orchestrate.
sigopt.get_suggestion: The latest Suggestion from the SigOpt API. This is automatically created for you, and is available to read.
sigopt.log_metric(name, value, stddev=None): Log a metric value for your model, like "accuracy". You may include an optional standard deviation. Any metrics not defined in the experiment configuration file will be recorded as metadata (see below).
sigopt.log_metadata(key, value): Log a piece of metadata.
sigopt.log_failure(): Indicate that this model is a failure.
In the backend, the information from the
sigopt.log_* calls will be automatically bundled into a SigOpt Observation. The JSON payload for every SigOpt Observation create call is logged to standard out. You will see log warnings if:
- You use
parameter_nameis not a parameter in your
- You use
metric_nameis not a metric in your
- You use
sigopt.log_failure()during the evaluation of your model.
- Any of your
runcommands failed. This is also interpreted as a failure.
- At least one of your metrics was not logged with
sigopt.log_metric. This is also interpreted as a failure.
SigOpt Orchestrate Compute Resources Back to Top
If you're training a model that needs a GPU you will want to use
resources_per_model to ensure that your model has access to GPUs. Requests and limits are optional, but may be helpful if your model is having trouble running with enough memory or CPU resources.
Requests are resource guarantees and will cause your model to wait until the cluster has available resources before running. Limits prevent your model from using additional resources. These map directly to Kubernetes requests and limits.
Note: If you only set a limit it will also set a request of the same value. See the Kubernetes documentation for details.
- CPU resources are measured in number of "logical cores" and can be decimal values. This is generally a vCPU in the cloud and a hyperthread on a custom cluster. See Meaning of CPU on the Kubernetes documentation for cloud specific and formatting details.
- Memory is measured in number of bytes but can be postfixed by "Mi, Gi" for megabytes and gigabytes respectively. See Meaning of Memory on the Kubernetes documentation for details and below for a simple example.
- The gpus field is currently specific to Nvidia gpus tagged as "nvidia.com/gpu". Alternatives can used by adding them to the limits field.
The below example will guarantee 20%(.2) of a logical core, 200 megabytes of memory, and a gpu are available for your model to run. If the cluster you are running on does not have enough free compute resources it will wait until they become available before running your model. This example will also limit your model so that it does not use more than 1 logical core and 2 gigabytes of memory.
name: My Orchestrate Experiment install: - pip install -r requirements.txt run: python model.py image: example/foobar resources_per_model: gpus: 1 requests: cpu: .2 memory: 200Mi limits: cpu: 1 memory: 2Gi optimization: ...
Docker Back to Top
Orchestrate uses Docker to build and upload your model environment. If you find that
orchestrate optimize is taking a long time, then you may want to try some of the following tips to reduce the build and upload time of your model:
Keep your model directory free of extra files
Omit files like logs, saved models, tests, and virtual environments. Changes to these extra files will cause Orchestrate to re-build your model environment.
Reduce the complexity of your install commands
If you can, use one of our pre-built frameworks as a starting point.
Omit your training data from your model directory
You can try downloading or streaming your training data in your run commands instead.
.dockerignore file in your model directory
This file should contain a list of the files that you want to omit from your model environment.
# python bytecode **/*.pyc **/__pycache__/ # virtual environment venv/ # training data data/ # tests tests/ # anything else .git/ saved_models/ logs/
See the official Docker documentation for more information.
(Advanced) Custom Image Registries Back to Top
To use a custom image registry, provide the
registry argument when you connect to your cluster:
cluster connect \ --cluster-name tiny-cluster \ --provider custom \ --kubeconfig /path/to/kubeconfig \ --registry myregistrydomain:port