Csmart Studio Desktop

User Guide for the Csmart Studio Desktop application

Overview

Csmart Studio Desktop is a desktop application for preparing datasets, training AI models, evaluating their performance, and exporting them for deployment on Csmart Digit devices. It provides a graphical interface that guides you through the complete machine learning workflow, from raw image data to a production-ready ONNX model.

How It Fits Into the Csmart Ecosystem

Studio Desktop is the middle link in the Csmart pipeline:

1. Csmart Digit Desktop captures images of coffee samples during operation and stores them in HDF5 format.

2. Csmart Studio Desktop uses those images to train and evaluate AI classification models.

3. The trained models are exported as ONNX files and deployed back to Csmart Digit Desktop for real-time inference.

Csmart Digit Desktop          Csmart Studio Desktop          Csmart Digit Desktop
(Image Acquisition)     -->   (Model Training & Export)  -->  (Production Inference)
                                      |
                                 (ML Engine)

Application Workflow

Studio Desktop follows a linear workflow organized into numbered screens. Each step builds on the output of the previous one:

Step	Screen	Purpose
1	Home	Create or open a project
2	Split Dataset	Divide images into training, validation, and test sets
3	Image Search	Find visually similar images across the dataset
4	Pre-Classification	Automatically classify images using an existing model
5	Outlier Detection	Identify and remove anomalous images
6	Cluster Dataset	Group similar images for balanced training
7	Train AI Model	Train a classification model using PyTorch
8	Test AI Model	Evaluate model performance with metrics and visualizations
9	Feature Extraction	Extract embeddings for similarity analysis
10	Export AI Model	Convert the trained model to ONNX format for deployment

Additional screens manage Model Settings, Hardware Settings, Runtime Environment, and License Management.

Key Concepts

Projects

All work in Studio Desktop is organized into projects. A project is a folder on your computer that contains your dataset configuration, training checkpoints, exported models, and metadata. When you create a new project, Studio Desktop sets up the following structure:

my-project/
  config.yaml          # All project settings (dataset paths, training params, export config)
  project.yaml         # Coffee metadata and export history
  Dataset/
    Original/V00/      # Link to your raw image dataset
    Split/V00/         # Train/validation/test split output
  Checkpoints/         # Training session outputs (version_00, version_01, ...)
  Models/              # Exported ONNX models

AI Models

Studio Desktop supports several model architectures for coffee classification, including ResNet, ConvNeXt, MaxViT, ViT, SegFormer, and more. Models are trained in PyTorch and exported to ONNX format, which is optimized for fast inference on the Digit hardware.

Configuration

Each project stores its configuration in a config.yaml file. Studio Desktop reads and writes this file automatically as you adjust settings through the interface. You typically do not need to edit this file manually.

System Requirements

Minimum Hardware Requirements:

• Processor: Intel Core i5 (10th generation or newer) or AMD equivalent

• RAM: 16 GB DDR4

• Storage: 500 GB SSD (models can reach several gigabytes per training session)

• GPU: NVIDIA GPU with CUDA support (recommended for training)

• OS: Windows 10 or later (64-bit)

Recommended Hardware for Training:

• Processor: Intel Core i7-13650HX (13th generation) or newer

• RAM: 32 GB DDR5

• Storage: 2 TB PCIe NVMe M.2 SSD

• GPU: NVIDIA RTX 5060 or better (16 GB+ VRAM)

• OS: Windows 11 (64-bit)

Prerequisites

Before using Studio Desktop, ensure you have:

1. A valid Csmart Studio license (.lic file) — The application requires license activation on first launch.

2. An NVIDIA GPU with up-to-date drivers if you plan to train models. CPU-only training is possible but significantly slower.

3. A dataset of classified coffee images organized in folders by class name (e.g., OK/, Black/, Sour/).

Getting StartedSplit DatasetPre-Classification