Skip to content

Profiler+ Workflow

In this workflow, you will retrain an Ultralytics model, revalidate it using edgefirst-profiler, and compare your results against the EdgeFirst Model Zoo on Hugging Face baselines. Models can be deployed in the browser or on a compatible target device.

Create an Ultralytics Training Session

  1. Click on "Model Experiments" of your created project

    Model Experiments Page
    Model Experiments Page

  2. Create a new experiment

    Model Experiments Page
    Model Experiments Page

  3. Navigate to the "Training Sessions"

    Training Sessions
    Training Sessions

  4. Create a new training session by clicking the "Actions" dropdown menu on the top right of the page and then click the "+ New" button

    New Session Button
    New Session Button

  5. Start a YOLOv8n detection model by following these settings. Once the settings are set, click on "Start Session" at the bottom of the window

    Ultralytics Training Settings
    Ultralytics Training Settings

    This will start a training session in progress.

    Ultralytics Training Progress
    Ultralytics Training Progress

    InsufficientInstanceCapacity

    InsufficientInstanceCapacity Error

    If you see this error after starting your training session, retry creating the session. This can happen when AWS reports that no EC2 instances are currently available to launch; the current workaround is to retry.

    No Training Charts

    No Training Charts

    Training sessions configured without epochs will not generate loss or metric charts, since the chart x-axis is epoch-based. This is expected behavior.

  6. The completed training session should look like the following

    Ultralytics Training Completed
    Ultralytics Training Completed

  7. We support model conversion and optimization workflows that enable trained models to be deployed on a wide range of target platforms and hardware architectures.

    Follow the model conversion workflow that corresponds to your target platform. If you are deploying to a Windows/Linux PC or macOS system, you can skip this section and proceed directly to the next step to validate the ONNX model, which is automatically generated as part of the training session outputs.

    If you are deploying to one of the supported hardware platforms listed below, follow the platform-specific conversion instructions for your target device.

    Converter Supported Targets Output Format Docs
    TFLite Converter NXP i.MX 8M Plus (VIP8000), generic CPU/NPU TFLite delegates .tflite flatbuffer TFLite Converter
    Neutron Converter NXP i.MX 95, i.MX 943/952, S32N79, MCX N54x/N94x, i.MX RT700, S32K5 .tflite flatbuffer with Neutron microcode Neutron Converter
    TensorRT Converter NVIDIA Jetson (Orin Nano Super validated; broader lineup in progress) .tensorrt.zip bundle (engine built on-device) TensorRT Converter
    Ara2 Converter NXP Ara240 DNPU .dvm Dataflow Virtual Machine binary Ara2 Converter
    Hailo Converter Hailo-8 (26 TOPS), Hailo-8L (13 TOPS) .hef Hailo Executable Format Hailo Converter

  8. All converted models should appear under the model artifacts of the training session card. Click on the training session card to expand for more details.

    YOLOv8n Detection Model Artifacts
    YOLOv8n Detection Model Artifacts

    Once you have converted your model, you can proceed towards profiling and validating the performance of your model next.

Validate on Target with the EdgeFirst Profiler

The Profiler+ workflow validates your model directly on your target hardware — not in the cloud. This gives you real inference latency, NPU utilization, and per-stage timing alongside accuracy metrics.

Cloud validation not used here

Cloud validation only supports ONNX, Keras, and TFLite artifacts running on a cloud instance. On-target profiling with the EdgeFirst Profiler supports all converted formats (.tflite, .hef, .dvm, .tensorrt.zip) and measures real hardware performance.

Install the profiler

pip install edgefirst-profiler

curl -fsSL https://raw.githubusercontent.com/EdgeFirstAI/profiler-cli/main/install.sh | bash

irm https://raw.githubusercontent.com/EdgeFirstAI/profiler-cli/main/install.ps1 | iex

Confirm the install:

edgefirst-profiler --version

For per-platform guides including NPU delegate setup, see the Profiler Installation section.

Sign in to EdgeFirst Studio

edgefirst-profiler login

The interactive prompt asks for server, username, and password. Credentials are saved locally and refresh automatically.

Run a validation session

Launch the profiler TUI:

edgefirst-profiler

Press F2 to open the Studio screen, then navigate to your training session:

Projects  →  Experiments  →  Training Sessions  →  Artifacts

Select the converted model artifact that matches your target hardware and choose Validate. The profiler creates a validation session in Studio, downloads the model and dataset to the device, and starts the run automatically.

The F4 Profiler dashboard streams iteration-level latency and per-stage timings live during the run. When complete, predictions and the timing trace upload to EdgeFirst Studio where mAP, precision-recall curves, and the trace viewer are generated.

For the full walkthrough see Profiler Quick Start and Validation from the Profiler.

Compare Results to the Model Zoo

After validation, compare your results with the EdgeFirst Model Zoo on Hugging Face to see how close your model is to the published baselines.

Focus on:

  • Task match: Use the same task type (detection, segmentation, classification) as your training run.
  • Input modality: Compare vision-only models against vision-only baselines, and fusion models against fusion baselines.
  • Metric alignment: Match the primary metric (for example, mAP or F1) to ensure an apples-to-apples comparison.
  • Latency vs accuracy: Look at accuracy and throughput together to understand the practical edge tradeoff.

Next Steps

Want to learn more?