Image Classification with SambaStudio

Overview:

This document explains how to perform image classification using the Vit_B_Classification model in SambaStudio. Image classification is the process of assigning labels to images from a set of predefined classes.

Data Preparation:

To train a classification model, you need:

  1. Images:
  • Supported formats: .jpg, .png
  • Must be 3-channel RGB encoded as uint8
  • Images with an alpha channel must have the alpha removed during preprocessing
  1. labels.csv:

    • A CSV file mapping image paths to class labels and splits.
  2. class_to_idx.json (optional):

  3. Maps readable class names to numerical indices

{code:json}
{
  "cat": 0,
  "dog": 1
}
{code}

Using SambaStudio:

  • Log in to your on-prem SambaStudio instance
  • Create a new project and choose Vit_B_Classification
  • Upload your dataset (including labels.csv and optionally class_to_idx.json)
  • Train, validate, and evaluate your model

Tips & Troubleshooting:

  • Ensure all images are RGB and correctly formatted
  • Double-check paths in labels.csv
  • Use consistent class labels and file naming

Note - “This requires a purchased on-prem subscription to SambaNova Suite”

1 Like