In this document, we will cover the common tasks involved in building a machine learning model and guide how you can use our to accomplish these tasks:
Build a baseline machine learning model with .
Optimize hyperparameters using .
Update and store models on .
Here, we will use the MNIST database to create an image classification model. All our CLI commands are one-liners but you can also select from command option prompts.
Requirements
To follow this guide, you should first have the following setup.
— a dedicated organization for you or your team
— a space for your machine learning model and mounted datasets
— Python SDK and CLI to manage ML workflows and resources on VESSL
If you have not created an Organization or a Project, first follow the instructions on the .
1. Experiment — Build a baseline model
1-1. Configure your default organization and project
Let's start by configuring the client with the default organization and project we have created earlier. This is done by executing .
vessl model create --model-name "v0.0.1"
Organization: YOUR_ORGANIZATION_NAME
Project: YOUR_PROJECT_NAME
[?] Model repository: YOUR_MODEL_REPOSITORY_NAME
> YOUR_MODEL_REPOSITORY_NAME
[?] Source: From an experiment
> From an experiment
From local files
[?] Experiment: 1-quasar-bat #1
> 1-quasar-bat #1
[?] Paths (Press -> to select and <- to unselect):
X my_model 0 B
X my_model/keras_metadata.pb 7.7 KB
X my_model/saved_model.pb 88.8 KB
X my_model/variables 0 B
X my_model/variables/variables.data-00000-of-00001 1.2 MB
> X my_model/variables/variables.index 1.4 KB
You can see that the model has been created successfully by specifying the repository name and selecting the model number.
vessl model read "tutorial-mnist"
You can get a list of model repositories and models you have created inside the project by using the following commands.
vessl model-repository list # get a list of model repositories
vessl model list # get a list of models
vessl sweep create --dataset "/input:YOUR_DATASET_NAME"
Organization: YOUR_ORGANIZATION_NAME
Project: YOUR_PROJECT_NAME
[?] Objective type: maximize
> maximize
minimize
Objective metric: val_accuracy
Objective goal: 0.99
Maximum number of experiments: 4
Number of experiments to be run in parallel: 2
Maximum number of experiments to allow to fail: 2
[?] Sweep algorithm: random
grid
> random
bayesian
Parameter #1 name: optimizer
[?] Parameter #1 type: categorical
> categorical
int
double
[?] Parameter #1 range type: list
space
> list
Parameter #1 values (space separated): adam sgd adadelta
Add another parameter (y/n): y
Parameter #2 name: batch_size
[?] Parameter #2 type: int
categorical
> int
double
[?] Parameter #2 range type: space
> space
list
Parameter #2 values ([min] [max] [step]): 64 256 8
Add another parameter (y/n): n
[?] Cluster: aws-apne2-prod1
> aws-apne2-prod1
[?] Resource: v1.cpu-4.mem-13
v1.cpu-0.mem-1
v1.cpu-2.mem-6
v1.cpu-2.mem-6.spot
> v1.cpu-4.mem-13
v1.cpu-4.mem-13.spot
v1.t4-1.mem-13
v1.t4-1.mem-13.spot
v1.t4-1.mem-54
v1.t4-1.mem-54.spot
v1.t4-4.mem-163
v1.t4-4.mem-163.spot
v1.k80-1.mem-52
v1.k80-1.mem-52.spot
[?] Image URL: public.ecr.aws/vessl/kernels:py36.full-cpu
> public.ecr.aws/vessl/kernels:py36.full-cpu
public.ecr.aws/vessl/kernels:py37.full-cpu
public.ecr.aws/vessl/kernels:py36.full-cpu.jupyter
public.ecr.aws/vessl/kernels:py37.full-cpu.jupyter
tensorflow/tensorflow:1.14.0-py3
tensorflow/tensorflow:1.15.5-py3
tensorflow/tensorflow:2.0.4-py3
tensorflow/tensorflow:2.2.1-py3
tensorflow/tensorflow:2.3.2
tensorflow/tensorflow:2.4.1
tensorflow/tensorflow:2.3.0
Start command: git clone https://github.com/vessl-ai/examples.git && pip install -r examples/mnist/keras/requirements.txt && python examples/mnist/keras/main.py --save-model --save-image
The first part of the command defines the key objective and number of experiments.
--objective-type — target object (either to minimize or maximize the metric)
--objective-goal — target metric name as defined and logged using vessl.log()
--objective-metric — target metric value
--num-experiments — total number of experiments
--num-parallel — the number of experiments to run in parallel
--num-failed — the number of failed experiments before the sweep terminates
Next, we specified the details of the parameters and which algorithm to use. In this example, the optimizer is a categorical type and the option values are listed as an array. The batch_size is an int value and the search space is set using max, min, and step.
The command is then followed by cluster, resource, image, dataset, and command options, similar to the vessl experiment create explained above.
3. Model Registry — Update and store the best model
vessl model create --model-name "v0.0.2"
Organization: YOUR_ORGANIZATION_NAME
Project: YOUR_PROJECT_NAME
[?] Model repository: YOUR_MODEL_REPOSITORY_NAME
> YOUR_MODEL_REPOSITORY_NAME
[?] Source: From an experiment
> From an experiment
From local files
[?] Experiment: 1-quasar-bat #1
> 1-quasar-bat #1
[?] Paths (Press -> to select and <- to unselect):
X my_model 0 B
X my_model/keras_metadata.pb 7.7 KB
X my_model/saved_model.pb 88.8 KB
X my_model/variables 0 B
X my_model/variables/variables.data-00000-of-00001 1.2 MB
> X my_model/variables/variables.index 1.4 KB
Finally, you can view the performance of your model by using vessl model read and specifying the model repository, followed by the model number.
vessl model read "tutorial-mnist" "2"
vessl model read
Organization: YOUR_ORGANIZATION_NAME
[?] Model repository: YOUR_MODEL_REPOSITORY_NAME
> YOUR_MODEL_REPOSITORY_NAME
[?] Model: 1
> 1
2
We covered the overall workflow of VESSL using the client CLI. We can also repeat the same process using the client SDK or through Web UI. Now, try this guide with your own code and dataset.
To create a on VESSL, run . Let's create a dataset from the public AWS S3 dataset we have prepared: s3://savvihub-public-apne2/mnist. You can check that your dataset was created successfully by clicking the output link.
To create an experiment, use . Let's run an experiment using VESSL's managed clusters. First, specify the cluster and resources options. Then, specify the image URL — in this case, we are pulling a Docker image from . Next, we are going to mount the dataset we have created previously. Finally, let's specify the start command that will be executed in the experiment container. Here, we will use the MNIST Keras example from our .
Note that you can also with your project so you don't have to git clone every time you create an experiment. For more information about these features, please refer to our doc's page.
The experiment may take a few minutes to complete. You can get the details of the experiment, including its status, using or by clicking the output link.
In VESSL, you can create a . First, let's start by creating a model repository using vessl model-repository create and specifying the repository name.
Finally, run with options including the destination repository and experiment ID. Make sure that the option value for --experiment-id is an integer, not a string.
So far, we ran a single machine learning and saved it as a inside a model repository. In this section, we will use a to find the optimal hyperparameter value. First copy and paste the following command and while the sweep is running we will explain each options.
You may find it easier to run vessl sweep create and specify the options through command prompts. For more information on sweep, refer to our
Now that we ran multiple experiments using a , let's find the optimal experiment. returns the experiment information with the best specified metric value. In this example, the command will return the details of the experiment with the maximum val_accuracy.
Let's create a v0.0.2 model with from the output of the best sweep experiment. You can get the experiment ID using the command.
