Quick Start - TensorFlow¶
In this tutorial, we will create a TensorFlow machine learning project based on our MNIST TensorFlow example on GitHub and run a simple training with it.
2. Create a project¶
- Go to the Valohai platform front page after signing in
- Press the Create Project button
- Set a
Namefor your project, e.g. test-tensorflow
- You can leave
Descriptionblank, that is more in detail definition of your project
- Press the Create button
4. Create an execution¶
- Go to the Executions tab inside your project
- Press the Create execution button
Stepfield lists all available types of executions. Make sure Train model is selected.
- You don’t need to worry about the rest of the configuration for now.
The default inputs and parameters of the form are loaded from the
valohai.yamlconfiguration file and should be good for this example execution.
- Press Submit to start the execution.
Valohai command-line client allows creating one-off executions from local files. See Quick Start - Command-line Client for more details.
$ vh exec run --adhoc --watch name-of-your-step # sends local source code to a worker and runs commands in valohai.yaml
5. View the results¶
After you start the execution, you are forwarded to the execution page.
This page has several tabs with execution details:
The Information tab shows the basic information about the execution, most of which could’ve been modified in the previous execution creation step.
The Log tab shows real-time log output from the execution. Anything that your code prints to the standard output (stdout) or standard error (stderr) streams is shown here.
The Metadata tab shows all the metadata output from the execution. You can also plot the metadata on a line chart. Metadata is any data your execution writes to the standard output stream in JSON which we can parse. If no plottable metadata has been output, this tab is not visible.
The Output tab contains download links for all the output artifacts created by the execution.
The execution defines these outputs by writing them into
The artifacts are stored in AWS S3.
If the execution has not finished, or did not output any files, this tab will not be visible.