Mastering Edge AI Workspaces in Fusion Studio
In our previous article, we took a guided tour through Fusion Studio, exploring how it brings data, models, and deployment into a single, connected environment.
Now, we’re going to focus on the heart of Fusion Studio: The workspace.
Everything you do in Fusion Studio begins here. A workspace is where your entire project lives. It stores your data, categories, annotations, experiments, and the models you train. It keeps them connected, versioned, and traceable. Without it, your files would be scattered across separate folders, and every minor change, such as a new dataset, a retrained model, or a revised label, would be more difficult to track.
Let’s walk through the process of creating and using a workspace within Fusion Studio.
Fusion Studio Workspace
A workspace in Fusion Studio is the complete environment where your project is defined and managed. Everything related to a single project is tied to a workspace, so you can trace every result back to its source. The Workspace has:
- Domain and Categories: The domain specifies the type of problem you’re working on, such as vision or audio. Categories are the labels you want to classify, detect, or predict — for example, Pass and Fail in a quality inspection task, or Healthy and Defective in crop monitoring.
- Datasets: You can store both raw data and annotated datasets here. Each dataset stays linked to the workspace, so you always know which version was used in which experiment.
- Annotations: Labeling information tied to each data sample. Annotations connect your categories to specific files or images, forming the ground truth your models will learn from.
- Experiments: Every training run you perform is saved here, along with its configuration, logs, and results. This makes it easy to compare models or revisit earlier runs.
- Models and Results: The trained outputs of your experiments. They stay connected to the data and parameters that produced them, so you always have context.
- Logs and History: Automatic tracking of actions taken within the workspace: data added, labels updated, experiments launched. This record helps you reproduce results or debug changes.
Next, we’ll see how to create your workspace and set the foundation for your project.
Setting up your workspace
1. Create a New Workspace
- Open Fusion Studio and click New Workspace.
- Choose either ‘Create Blank Workspace’ or ‘Get Started Quickly’ with one of our curated starter packs.
2. Configure Your Workspace
- Enter a name for your workspace.
- Select a domain (currently, only Vision is available).
- Choose your use case: Image Classification, Image Segmentation, Or Object Detection
3. Finalize Setup
- Review your selections and click Create Workspace to confirm.
Decide on a Dataset Strategy
Decide whether you’ll:
- Import pre-labeled dataset
- Upload raw data then annotate inside Fusion Studio
- Capture live samples from hardware.
Your Workspace, Ready for Action
The workspace is the foundation of every project in Fusion Studio. It keeps your data, annotations, experiments, and models connected. Once your workspace is set, you’re ready to start training.
Try MnasNet0.5 for Real-Time Workspace Object Recognition
Edge devices require models that strike a balance between precision, latency, and resource efficiency. MnasNet0.5 achieves that balance through a carefully optimized MobileNet-like architecture designed for fast inference on ARM-based processors. It delivers consistent accuracy for workspace object recognition, detecting items such as laptops, mugs, books, and keyboards, while maintaining low computational overhead.
On ModelNova, MnasNet 0.5 is available in TensorFlow Lite, PyTorch, and ONNX formats, pre-trained and ready for production. Through Fusion Studio, the model can be retrained with domain-specific datasets, fine-tuned for different lighting conditions or object classes, and deployed directly to target devices, such as Raspberry Pi.
For developers building vision-based automation, smart assistants, or inventory tracking systems, MnasNet0.5 provides a stable baseline model for real-time classification at the edge.
Meet Mnasnet05 Image Classification for Workspace
MnasNet 0.5 is a compact convolutional neural network designed for efficient image classification on constrained hardware.
Its architecture follows a depthwise separable convolution design, reducing computational cost without compromising feature extraction capability. The model is optimized for ARM-based processors and performs consistently across low-power platforms such as Raspberry Pi 4 and 5.
It is pre-trained on a curated dataset of workspace objects, including laptops, mugs, phones, notebooks, and bottles for immediate use in recognition tasks common to desk environments, office automation, and educational settings.
Key Metrics
- Model size: 1.2 million parameters
- Inference time: 7.54 milliseconds
- Energy consumption: 343.3 millijoules per inference
What You Can Build with MnasNet0.5
MnasNet0.5 enables practical edge intelligence across workspace and office environments. Its efficiency allows developers to design responsive vision systems that operate entirely on-device.
Example Applications
a) Smart Workspace Assistant: Detects and classifies tools or personal items on a desk in real time, enabling context-aware interactions and task automation.
b) Office Automation System: Monitors equipment usage by identifying objects such as projectors, laptops, or peripherals, supporting energy optimization and asset tracking.
c) Edge-Based Inventory System: Processes camera feeds locally to recognize items, update counts, and trigger workflows without transmitting data to the cloud.
Build, Adapt, and Deploy with Fusion Studio
Developers can explore MnasNet0.5 directly inside Fusion Studio, our local development environment designed for building and testing edge AI models without cloud dependencies. The model can be loaded, inspected, and deployed to supported devices, such as the Raspberry Pi, with minimal setup.
Fusion Studio provides an interactive workspace where you can:
-Load pre-trained models, such as MnasNet0.5, and test them with live or recorded image data.
-Upload your own dataset to retrain or fine-tune the model for specific objects or environments.
-Deploy the trained model back to your edge device through an integrated workflow.
All operations, from data import to quantization and packaging, run locally on your machine, ensuring full control over resources, privacy, and cost. With Fusion Studio, you can adapt and extend MnasNet0.5 into real, deployable applications.
Why Developers Love This Workflow
Fusion Studio provides a streamlined environment for building and deploying edge AI applications. It reduces the overhead of managing separate tools and frameworks, allowing developers to move from concept to deployment with clarity and speed.
i) Start Fast with Pre-Trained Models in ModelNova: Access optimized models like MnasNet0.5 immediately, ready to run on local hardware without additional setup.
ii) Customize Without Complex Pipelines: Retrain or fine-tune directly inside Fusion Studio using your own dataset, with all conversions and packaging handled automatically.
iii) Iterate Quickly with Real-Time Feedback: Evaluate model performance live, refine datasets, and redeploy updated versions in a continuous loop.
iv) Unified Platform: Manage model storage, training, and deployment from a single workspace, reducing fragmentation across tools and environments.
This workflow aligns with how modern edge AI development should operate — local, efficient, and iterative, with complete control over every stage of the process.
Get Started Today
Begin building your edge vision application with a proven foundation. Explore MnasNet0.5 Image Classification on ModelNova, download the model, and test it directly on your device.
Sign in to Fusion Studio to retrain or fine-tune the model using your own dataset, adapt it to your specific environment, and deploy it locally —all within a single workspace.
Developers are encouraged to share their projects, insights, and feedback to help expand the edge AI community.
