Overview

Introduction

NeoPulse® Manager is a UI tool used for the development of AI models from start-to-finish, including uploading data, training models, querying them, and more. Also possible from within NeoPulse® Manager is to upload other NeoPulse® Studio or NeoPulse® Runtime machines to split up tasks and efficiently streamline the workflow. Within NeoPulse® Manager, there are a few concepts whose language will be used throughout documentation. These components are as follows:

  • Project: a collection of experiments with the goal of trying to solve a problem. Think of them like a workspace curated towards a specific AI task.
  • Data(set): the data used to drive the training (and eventually testing/querying) of models in a project.
  • Model: the output of an AI algorithm run on a set of training data.
  • Iteration: the training of a model is typically run over many iterations, where each iteration generally improves the model's performance according to a set of metrics.
  • Portable Inference Model (PIM): contain a model's information at a specific iteration of that model, i.e. input/output requirements, hyper-parameters, and more.
  • Test: different PIMs can be fed a dataset, after which the PIMs' performance according to a set of metrics can be analyzed to help choose the optimal PIM for query.
  • Metric: mathematical functions such as Loss or Accuracy which are used to describe a model's skill. Not all metrics apply to all types of problems; regardless, they are necessary to ultimately decide which model is optimal for the task.
  • Query: once a model has been trained and fine-tuned, it's time to use it! A user can submit a query containing input data to the model, and expect a response in the form of a model's defined output(s). Queries can be submitted using the REST APIs or from the Command Line.

Workflow

For a visual representation of the workflow of NeoPulse® Manager, consider watching our Quick Start Guide, or the Features Guide.

Within NeoPulse® Manager, work begins in Datasets, where the data to be used for training can be uploaded. After a successful upload, head to Models, and create a project. Attach your uploaded data to the project, and create a view (input/output specification). This will generate the source construct of an NML File for you. Complete the rest of the NML file in the Model Editor, where the only necessary step is to complete the architecture construct. Note that the Model Editor uses an "auto block" by default, so if performing a classification or regression task, no additional work is required besides saving model design. The train construct also does not need to be edited, but this is possible through the Model Editor. Then head to DevOps and add a NeoPulse® Studio machine to handle training, after which models can be trained using the "Train" tab of models. Once model training has completed, test them using the "A/B Testing" tab and schedule the best PIM for RTQ using DevOps!