Understand models in Sentis

Sentis can import and run trained machine learning model files in Open Neural Network Exchange (ONNX) format.

To get a model that's compatible with Sentis, you can do one of the following:

Train a model using a framework like TensorFlow, PyTorch, or Keras, and subsequently export it in ONNX format.
Download a trained model file and convert to ONNX format. Refer to the ONNXXMLTools Python package for more information.
Download a trained model that's already in ONNX format, such as those available in the ONNX Model Zoo. Refer to supported models for more resources.

How Sentis optimizes a model

When you import a model, each ONNX operator in the model graph becomes a Sentis layer.

Open the Model Asset Import Settings to check the list of layers in the imported model, in the order Sentis runs them. Refer to Supported ONNX operators for more information.

Sentis optimizes models to make them smaller and more efficient. For example, Sentis might do the following to an imported model:

Remove a layer or subgraph and turn it into a constant.
Replace a layer or subgraph with a simpler layer or subgraph that works the same way.
Set a layer to run on the CPU, if the data must be read at inference time.

The optimization doesn't affect what the model inputs or outputs.

Model inputs

You can get the shape of your model inputs in one of two ways:

Inspect a model to use the inputs property of the runtime model.
Select your model from the Project window to open the Model Asset Import Settings and view the Inputs section.

The shape of a model input consists of multiple dimensions, defined as a SymbolicTensorShape.

The dimensions of a model input are either fixed or dynamic:

An int denotes a fixed dimension.
The strings ? and d0, d1 etc. denote dynamic dimensions.

Fixed dimensions

The value of the int defines the specific size of the input the model accepts.

For example, if the Inputs section displays (1, 1, 28, 28), the model only accepts a tensor of size 1 x 1 x 28 x 28.

Dynamic dimensions

When the shape of a model input contains a symbolic dimension, a ? or string, for example batch_size or height, the input dimension can be any size.

For example, if the input is (d0, 1, 28, 28), the first dimension of the input can be any size.

When you define the input tensor for this input shape, the following input tensor shapes are valid:

[1, 1, 28, 28]
[2, 1, 28, 28]
[3, 1, 28, 28] ...

If you change the size of another dimension, however, the tensor input is not valid. For example:

[1, 3, 28, 28]

Note

If a model uses inputs with dynamic shapes, Sentis might not be able to optimize the model as efficiently as a model that uses fixed input dimensions. This might slow down the model.