What is a Training?

Model training is the key step in the entire MLOps pipeline. It is the process of teaching the model to recognize patterns, objects, or specific features in your dataset and make predictions or perform tasks based on that understanding.

Training typically requires large amounts of compute resources. On Datature's platform, however, we provide the environment and compute resources to kick-start a training without having to write any setup codes.

Training Process

A typical model training goes through the following stages:

  1. Model Initialization: Typical models contain parameters that are initialized with random values or pre-trained weights. On Nexus, we use pre-trained models that are trained on large-scale datasets like ImageNet to act as an improved starting point by leveraging their learned features.
  2. Training iterations: The model is trained on the annotated dataset in multiple iterations or epochs. During each iteration, the model takes input images, makes predictions, compares the predictions with the ground truth labels, and calculates the loss or error. The model's parameters are updated using optimization algorithms (e.g., stochastic gradient descent) to minimize the loss and improve the model's performance.

You can monitor your training in real-time by analyzing the various metrics and features we provide that give a gauge of how well your model is improving over time. For more details, check out how to Monitor your Training Process.