In this post, we will describe the common practices that could be useful for improving your experience with Robotika. We will show how to access the output of the pipeline, how to take advantage of logging, and what to do in case you are stuck.
As soon as you save the Scheduler, Robotika runs the pipeline. The status of the pipeline run can be “Completed,” “Failed,” or “Processing.”
You can click on the blue arrow to visualize each run regardless of its status. In the image below, you can see the example of the run where the Transform component failed. You can easily see what component has failed by looking at the color.
By clicking on this component, we can investigate what went wrong by looking in the logs. Please note that we enable access to the logs for Robotika users. It has been done to increase the speed of the development. In our particular case, we have intentionally made a transform script empty. As a result, the algorithm failed precisely on the transform component step.
Let us briefly describe the information you can extract from the pipeline’s logs and components information.
2. Logging and component analysis.
The logging in Robotika enables data scientists to get additional insight into the pipeline run and the data. Each component of the pipeline contains some information that could be useful either for debugging or for additional statistics about the data. Let us go briefly what type of insight can the logs bring you for each component.
- The ExampleGen contains the links to the training and validation files (if they were created in the pipeline).
- SchemeGen contains the description of the columns in your dataset and their distribution.
- Transform component does not contain any useful information.
- By clicking on Trainer, component user will be headed to the TensorBoard, where the process of the training is visualized. To know more about Tensorboard, please look here (https://www.tensorflow.org/tensorboard)
- The Evaluator component indicates whether your model has been blessed or not.
3. Technical support.
Sometimes the process of debugging could be frustrating. Just a piece of advice can change things upside down. For this reason, we created a special ticket system that enables us to simulate the pipeline run you are currently developing. Thus, we will not only answer a specific question, but also we will try to find a bug for you. In order to activate this function, you can click on the “Get Support” button either in “Edit pipeline” or in “Pipeline run details”, and we will try to help you deploy your algorithm.