Run the DL Workbench in the Intel® DevCloud for the Edge

Intel® DevCloud for the Edge is a computing resource to develop, test, and run your workloads across a range of Intel® CPUs, GPUs, and Movidius™ Myriad™ Vision Processing Units (VPUs). Running the DL Workbench in the DevCloud helps you optimize your neural network and analyse its performance on various Intel® hardware configurations. DevCloud is available for enterprise developers, independent developers, students, and faculty members.

When running the DL Workbench in the DevCloud, you can use the following DL Workbench features:

Feature Supported
Single and group inference Yes
(HDDL plugin is not supported)
INT8 calibration Yes
Winograd algorithmic tuning No
Accuracy measurements No
Model output visualization Yes
Visualization of runtime and Intermediate Representation (IR) graph Yes
Connecting to your remote machine** No
Downloading models from the Intel Open Model Zoo Yes
Deployment package creation Yes

** In the DevCloud, you are connecting only to remote machines that are available in it. You cannot work with your local workstation or connect to other machines in your local network.

NOTES:

  • Profiling and calibration on DevCloud machines take more time than on a local machine due to the exchange of models, datasets, the job script, and performance data.
  • Inference results may insignificantly vary for identical environment configurations. This happens because the same environment in the DevCloud does not mean the same physical machine.

Start the DL Workbench in the Intel® DevCloud for the Edge

  1. Register in the DevCloud. Shortly after the registration, you receive a mail with a link to the home page.
  2. Follow the link from the mail. If you agree to the Terms and Conditions, click Accept.
  3. In the Advanced tab on the Home page, click Deep Learning Workbench.
  4. The JupyterHub* tab opens indicating that the server is not running. Press Launch Server. Make sure your browser does not block pop-up windows as it prevents the tab from opening.
  5. The Jupyter* notebook called DLWorkbench_Launcher.ipynb opens. In the notebook, run the code cell.
  6. The Start Application button appears.
  7. Once you click the button, the DevCloud starts initializing and loading the DL Workbench.
  8. In about a minute, two links appear under the cell:

Click the first link. The DL Workbench opens:

Congratulations, you have started the DL Workbench in the DevCloud. Now follow the general DL Workbench workflow to create a project:

  1. Select a model
  2. Select a dataset
  3. Select a platform

Then run a baseline inference and proceed to the experiments listed below, then prepare the model for deployment:

A DL Workbench session in the DevCloud is limited to four hours. Remaining session time is indicated in the upper-right corner of the interface:

After four hours, the Docker container with the DL Workbench stops, but your data is autosaved in the DevCloud. To continue working with the DL Workbench, start a new session.

Start a New DL Workbench Session in the Intel® DevCloud for the Edge

To start a new session, return to the Jupyter notebook.

  1. Stop the current session by pressing the Stop Application button under the code cell in the notebook.
  2. Once the DL Workbench is stopped, the Start Application button appears. Press it.
  3. Once you click the button, the DevCloud starts initializing and loading the DL Workbench.
  4. In about a minute, two links appear under the cell:

Click the first link. The DL Workbench opens:

Obtain Optimized Model

After optimizing your model in the DL Workbench, you can download it from the tool directly. There are two ways to do this:

  • From the Configurations table by clicking the download icon in the Actions column.
  • From the Pack tab.
    1. Select the configuration and go to the Pack tab.
    2. Check Yes in the Include Model to Package field. Configure other fields based on your goal. Refer to Build Your Application with Deployment Package for details.
    3. Press the Pack button.

Then you can upload to the DevCloud filesystem like any other asset and integrate it into your application in the same Jupyter notebook.


See Also