Model Asset Inspector
The Model Asset Inspector provides an overview of imported machine learning (ML) models. Use it to view and verify model properties after import, including structure, inputs, outputs, layers, and constants.
The Inspector supports all imported models and reflects any changes made to the import settings.
Imported object settings
After you configure the model's import settings, the Inspector displays the updated model properties in the Imported Object section.
Use the Imported Object settings section to do the following:
- Serialize the model to the
StreamingAssetsfolder. - Check the properties of an imported and optimized model.
The Imported Object section provides a structured view of an imported model, including inputs, outputs, layers, and constants.
Inputs
The Input section displays a list of model input tensors and their dimensions. The following table details the properties of each input.
| Property | Description |
|---|---|
| name | The name of the input, such as images or attention_mask. |
| index | The index of the input. |
| shape | The tensor shape of the input. A shape consisting only of numbers represents a static-shaped tensor, for example, (1, 1, 28, 28) indicates the model only accepts tensors of shape 1 × 1 × 28 × 28. If a dimension is a named string, such as batch_size or sequence_length, it's dynamic and can vary in size. For more information on dynamic inputs, refer to Model inputs. |
| dataType | The data type of the input. For example, Float or Int. |
Outputs
The Outputs section lists the model’s output tensors. The following table details the properties of each output.
| Property | Description |
|---|---|
| name | The name of the output. |
| index | The index of the output. |
| shape | The tensor shape of the output. Inference Engine tries to precalculate this shape from the model. If a dimension in the shape is a question mark (?), Inference Engine can't calculate the size of the dimension, or the size depends on the input (a dynamic output). If the entire tensor shape is Unknown, Inference Engine can't calculate the number of dimensions, or the number of dimensions is dynamic. |
| dataType | The data type of the output. |
Layers
The Layers section displays all layers in the model and the order in which Inference Engine executes them. The following table details the properties of each layer.
| Property | Description |
|---|---|
| type | The type of layer. Refer to Supported ONNX operators for more information. |
| index | The index of the layer. |
| inputs | The names of the inputs to the layer. Possible values are a model input from the Inputs section, another layer, or a constant. |
| properties | The properties of the layer. The properties will depend on the type of layer. Refer to Supported ONNX operators for more information. |
Constants
The Constants section lists fixed values used in the model. The following table shows the total number of constants and weights in the model, along with a detailed list of constants.
| Property | Description |
|---|---|
| type | The type is always Constant. |
| index | The index of the constant. |
| weights | The tensor shape of the constant. If the tensor shape is empty - () - the constant is a scalar (a zero-dimensional tensor). |
Import and update ONNX model settings
Use the Model Asset Import Settings section in the Inspector to configure dynamic input shape dimensions before Unity processes the ONNX model.
When you update a dynamic input value, such as setting batch_size from -1 (dynamic) to a static number like 4, and select Apply, Inference Engine re-imports the model with the updated configuration. These changes are reflected in the Imported Object > section, where both the Inputs list and the total number of Layers will update based on the modified input dimensions.
Note
When you modify an input dimension in an ONNX model and select Apply > Serialize To StreamingAssets, Inference Engine preserves the updated values when serializing the model to a .sentis file. The serialized model retains the optimized configuration and doesn't revert to its previous dynamic state. However, you can't modify these input dimensions directly in the .sentis file after serialization.
Model information
This section provides metadata about the imported model, including:
- Total weight size: Displays the total size of the model constants.
- Producer Name: Displays the source of the model. This might be the producer name and version from the ONNX model or
Scriptif Inference Engine generated the model.