Available with Image Server
Training samples are used to derive image chips to train deep learning models in Deep Learning Studio. Image chips are small images containing the feature or object of interest used to train the deep learning model.
Collect training samples
Create training samples by completing the Prepare training data step. When you select the Prepare training data step, you must configure the project for collecting samples.
- Select the Deep Learning Studio project and open it.
- On the step choice page, select Prepare training data to start the process of collecting training samples.
Tip:
A prompt to configure the project for preparing training data appears if it has not been configured previously.
- In the prompt, click Yes to configure the project as instructed in the Configure the project for training section of the Work with Deep Learning Studio projects topic.
- Click the Collect training samples substep.
- Use the collection tools to select all of the training samples in the selected work unit.
There are keyboard shortcuts to make the collection efficient.
Show list of shortcuts
Alt + ?
Map Tool
Rectangle
r
Circle
c
Polygon
p
Polygon from lasso
i
Circle from point and buffer
o
Polygon from line and buffer
n
Select
s
Layers
l
Basemap
b
Filter
f
Map Navigation
Home
Alt + h
Zoom in
+
Zoom out
-
Pan
Arrow keys
Toggle the crosshair cursor
m
Tip:
All keyboard shortcuts are for Windows users. For Mac users, substitute Option key for Alt.
- Click Complete or Complete and next to collect the next work unit.
Once the work unit is marked as complete, it cannot be selected in the Collect training samples substep. Once the work unit is ready for review, mark it as complete to change the status to pending review.
Caution:
If the work unit was marked as complete incorrectly, it must be set to Queued in the Review training samples substep in order to be available again.
Review training samples
Once the work unit is marked as complete, it is available for review. All the training samples must be reviewed in order to proceed to the next step.
- Click the Review training samples substep to start reviewing the samples in the next work unit that is pending review open.
The Prepare training data landing page allows you to see the status of the work units and samples to review the progress of the project. The graphics change colors based on the status and update as the work units are reviewed.
Note:
A work unit can only be reviewed when the collector has marked that work unit as complete.
Note:
To get information about a work unit, click the map or the entry in the table for specific information about that work unit.
The work unit opens with all of the collected samples listed.
- Click individual training samples in the Training samples list to review them.
When a sample is selected, the Approve selected and Reject selected options appear on the dialog box.
- Mark the work unit as Completed.
The substep returns to the Prepare training data landing page.
Tip:
More than one sample can be selected and approved without reviewing each individual sample. You can also approve or reject all of the samples in one step by choosing either the Approval all pending or Reject all pending option. By choosing an option, the selected work units, and training samples within each, will be processed.
Keyboard shortcuts for reviewing training samples
Approve selected | Alt + a |
Reject selected | Alt + r |
Complete | Alt + c |
Complete and next | Alt + n |
Once all of the training samples have been approved or rejected, a message appears indicating they have all been reviewed.
Updating work unit properties (Optional)
If you need to update work unit properties manually, there are tools within Prepare training data landing page. You can filter and select specific work units to update efficiently.
- Click the drop-down under the Show section to use a predefined filter to find specific work units.
- Optionally, you can also create a custom filter to select from the work units. Click the Filter button to open the Filter open dialog box.
- Enable one or more of the four options to filter the work units and set the filtering criteria.
Tip:
You can select any combination of filter options to select the desired work units. You will see the number of selected work units in parentheses next to the filter option.
The work units that meet the filter criteria will be shown on the map or table in the left panel.
- Enable one or more of the four options to filter the work units and set the filtering criteria.
- Select the work unit(s) to be updated and the properties that can be updated will appear on the left panel.
You can update all of the selected work units in one step rather than review them individually. This process allows you to alter the status of many work units at once.
- Choose which property to alter from the available options.
- Edit the property in the panel.
Edits made to the properties will be automatically saved.
- Click the Clear selection button to see the main left panel.
If you want to manually update the properties of the selected work units, then you can select one or more work units on the map or table and update them.
Manage image chips
After the training samples have been reviewed and approved, the last step is to export the image chips that will be used to train the model.
- Click Manage image chips when the project has enough approved training samples.
- Click the Export button to start the process of exporting the image chips based on the training samples.
- Optionally, adjust the configuration options and click Export.
Once the image chips are created, the next step is to train the deep learning model, which will be completed in the Train model substep.
Complete workflow
By following this workflow, you can create the training samples and the image chips for deep learning model training in a complete end-to-end workflow. There are several optional substeps available for use in the project. The additional substeps are provided to allow modification of existing projects as the analysis is conducted. The following are the three required substeps:
- Collect training samples
- Review training samples
- Manage image chips
When the image chips are created, they are available for the deep learning training process in the Train model step. If the trained model does not meet the expectations for the analysis, this substep can be revisited. You can modify or collect additional training samples and create image chips to use in the next model training process.