Working with outputs

After executing a model you can then retrieve outputs or intermediate information.

Fetching outputs

If the model has a single output, you can use worker.PeekOutput(). Otherwise, provide output names:

var output = worker.PeekOutput(outputName);
// or
var output = worker.PeekOutput(); // warning: returns outputs[0]

Note: worker.PeekOutput() does not transfer ownership of the tensor to you: it is still owned by the worker. Calling worker.PeekOutput() is preferable and reduces memory allocations. However, if you expect to use the tensor for a long period, call worker.Fetch() . Otherwise, tensor values are lost after the next call to worker.Execute() or after the call to worker.Dispose().

Introspecting intermediate nodes

You can also analyse intermediate node values by passing a list of node names that you want to query when creating the worker.

// list of nodes to querry
var additionalOutputs = new string[] {"layer0", "layer1"}
m_Worker = WorkerFactory.CreateWorker(<WorkerFactory.Type>, m_RuntimeModel, additionalOutputs);
...
var outputLayer0 = worker.PeekOutput("layer0");
var outputLayer1 = worker.PeekOutput("layer1");
// you can also querry the original output of the network
var output = worker.PeekOutput(outputName);

Introspecting the Barracuda model

The Barracuda model has simple memory representation. When the model loads you can query for inputs and outputs:

string[] inputNames = model.inputs;   // query model inputs
string[] outputNames = model.outputs; // query model outputs

Alternatively, you can directly iterate through the layers and investigate what the model is going to do:

foreach (var layer in model.layers)
    Debug.Log(layer.name + " does " + layer.type);

You can also query constants:

Tensor[] constants = worker.PeekConstants(layerName);

Verbose mode

You can enable verbose mode for different parts of Barracuda:

bool verbose = true;
var model = ModelLoader.LoadModel(onnxAsset, verbose); // verbose loader
var worker = WorkerFactory.CreateWorker(WorkerFactory.Type.ComputePrecompiled, model, verbose); // verbose execution