CNTK Evaluate Hidden Layers - wolfma61/CNTK GitHub Wiki

This page describes how to expose a trained model's hidden layer's values.

Overview

A CNTK model is built on interconnected layers. Some of these layers can be evaluated using the EvalDLL because they are tagged as being 'output' layers. In order to expose other layers through the EvalDLL, these layers must be tagged as output layers by adding them to the outputNodes property.

For example, the 01_OneHidden_ndl_deprecated.cntk configuration file refers to the 01_OneHidden.ndl file for the network definition. In this network description file, we have two layers defined:

h1 = DNNSigmoidLayer (featDim, hiddenDim, featScaled, 1)  
ol = DNNLayer (hiddenDim, labelDim, h1, 1)

But only one layer is marked as an output:

outputNodes = (ol)

Thus, the EvalDLL can only return values pertaining to the ol layer during evaluation.

In order to be able to evaluate the h1 hidden layer, we need to expose it first as an output node. There are three possible ways of doing this:

For models that have not been trained yet, build the model with the desired layers listed in the outputNodes property.
Modify an already trained model on the fly while loading it, creating an in-RAM copy of the trained model that exposes the layers.
Instruct EvalDLL/EvalWrapper modules to modify the set of output nodes of an already trained model while loading.

1. Training model with hidden layers exposed

To output the h1 layer, just add it as an output in the network description (01_OneHidden.bs file) when training it, and that layer would be available for reading during evaluation:

outputNodes = (h1:ol)

However this implies the model would need to be (re)trained with this configuration.

2. Modifying an already trained model

Models can be modified on the fly when being loaded using BrainScript expressions. This will be documented in a future update of this documentation.

3. Changing output-node set of an already trained model while loading it for evaluation using the `EvalDLL`/`EvalWrapper` modules

If a trained will be evaluated using the EvalDLL/EvalWrapper modules, then it is just a matter of adding the outputNodeNames property with a colon separated list of nodes to the network definition:

outputNodeNames = "h1.z:ol.z"

When loading the network, the Eval engine will recognize the outputNodeNames property and replace the model's output nodes with the list of nodes specified in the outputNodeNames property.

Looking at the code inside the CPPEvalClient example project, shows the (uncommented) line specifying the outputNodeNames property:

networkConfiguration += "outputNodeNames=\"h1.z:ol.z\"\n";
networkConfiguration += "modelPath=\"" + modelFilePath + "\"";
model->CreateNetwork(networkConfiguration);