Visualizer

Use Visualizer to look inside a neural network. Follow how its layers connect, inspect individual operations, and compare models before and after changes such as quantization or compilation.

Open Visualizer from the sidebar in the Embedl Hub web app. You can import a model from your computer, or choose Visualize from a model artifact’s file actions on a run.

For example, compare an original model with a quantized version to see where precision changed, alongside the accuracy and latency recorded in your runs.

Get started#

Follow Visualize a model to import a file, inspect a layer, and save your workspace.

Explore the graph#

Start with Blocks for an overview of the model, then switch to Operations for individual layers. Pan and zoom to follow connections, or search for an operation by name.

Select a node to inspect its properties, inputs, outputs, and available precision information. Select a connection to inspect the tensor passed between operations. This helps you check tensor shapes and understand how a model’s data flows through the network.

Compare models#

View the model at different stages of export and compilation, such as PyTorch, ONNX, and TensorRT. Visualizer maps corresponding layers between the graphs, including cases where several layers fuse into one compiled operation. Follow Compare model stages to trace an operation back to its source and understand what changed.

You can also add an original and modified model to the same workspace to compare versions side by side.

Use this when checking what changed during quantization, reviewing a compiled model, or explaining an optimization to a teammate. Return to Experiment Tracking to compare measured accuracy and hardware performance for those candidates.

Save your inspection#

Workspaces are temporary. Before leaving, use Download .embedl bundle in the toolbar to save a .embedl file that you can reopen or share. To keep it with the experiment, upload it as a run artifact.