Visualizing AI Models

This section describes how to use the visualization tool Netron and the command-line tool ncc-tflite to inspect model structures and tensor information.

Current Neuron Runtime versions on IoT Yocto do not support embedding or reading model metadata during the Deep Learning Archive (DLA) compilation. Consequently, developers must manually verify model details to ensure correct inference implementation.

Inspecting Models with Netron

Netron provides a graphical interface to explore model architectures. The following examples use the yolov5s_int8 model, which is available on the Model Zoo-YOLOv5s.

• Full Graph View

  Opening a model in Netron displays the entire network graph, including layer connections and flow.

  

  Full view of YOLOv5s

  Note

  Netron supports multiple frameworks, including TensorFlow Lite (TFLite), ONNX, Caffe, and Keras.

• Tensor Details

  Selecting a specific node or tensor reveals its properties. For example, the input tensor Preprocessor/sub shows a 3-channel image of size 300 x 300 with quantization parameters (zero_point: 128, scale: 0.0078).

  

  Detailed information for a specific tensor in Netron

Using ncc-tflite for Visualization

The Neuron Compiler (ncc-tflite) offers a command-line interface to extract model information directly.

• List Input and Output Tensors

  Developers can use the --show-io-info flag to view the input and output configurations of a TFLite model.

  $ ncc-tflite --show-io-info yolov5s_int8.tflite
  
  # of input tensors: 1
  [0]: Preprocessor/sub
  ├ Type: kTfLiteUInt8
  ├ Shape: {1,300,300,3}
  ├ Scale: 0.00787402
  ├ ZeroPoint: 128
  └ Bytes: 270000
  
  # of output tensors: 2
  [0]: concat_1
  ├ Type: kTfLiteUInt8
  ├ Shape: {1,1917,21}
  ├ Scale: 0.141151
  ├ ZeroPoint: 159
  └ Bytes: 40257
  [1]: Squeeze
  ├ Type: kTfLiteUInt8
  ├ Shape: {1,1917,4}
  ├ Scale: 0.06461
  ├ ZeroPoint: 173
  └ Bytes: 7668
  

• Show All Tensors and Nodes

  The --show-tflite flag displays the full list of internal tensors and operation nodes.

  $ ncc-tflite --show-tflite yolov5s_int8.tflite
  
  Tensors:
  [0]: BoxPredictor_0/BoxEncodingPredictor/BiasAdd
  ├ Type: kTfLiteUInt8
  ├ Shape: {1,19,19,12}
  ├ Scale: 0.06461
  ├ ZeroPoint: 173
  └ Bytes: 4332
  [1]: BoxPredictor_0/BoxEncodingPredictor/Conv2D_bias
  ├ Type: kTfLiteInt32
  ├ Shape: {12}
  ├ Scale: 7.66586e-05
  ├ ZeroPoint: 0
  └ Bytes: 48
  
  ...
  

Note

For comprehensive documentation on ncc-tflite parameters, please refer to the NeuroPilot Official Site - Neuron Compiler Section.