Model - Introduction

TensorSpace Model Introduction

The model object is one of the important components for TensorSpace. All layers are maintained in the model which is the linkage between visualized objects and computed data.

There are two types of models: Sequential() and Model(). To construct a model object, we need two parameters: the container and the config.

let model = new TSP.models.Sequential( container, config );

The container is actually the element in the HTML pages we used to render the 3D model.

The config represents the model settings which contains different properties and methods. Let's check them below.

Property

The model has several different configurations which are used as the default configurations of the model including all layers and other utility components. For example, we can configure a Sequential() model like:

let model = new TSP.models.Sequential( container, {

    layerInitStatus: "close",
    aggregationStrategy: "max",
    layerShape: "rect",
    textSystem: "enable",
    relationSystem: "enable",
    animeTime: 200,
    stats: true,

    color: {

        background: 0x000000,
        conv2d: 0xffff2E,
        pooling2d: 0x00ffff,
        dense: 0x00ff00,
        padding2d: 0x6eb6ff

    }

} );

Note:

The model configuration is a general, default configuration for the model. The detailed configurations of each element (e.g. layer) can be configured within the element, which override the model configurations.

The following shows the detailed configurations of the model:

.aggregationStrategy : String

Represents the strategy of computing data used to render layer aggregation.
"max" or "average". Default "average".
"max" to compute aggregation by using the max value of the same position in all channels; "average" to compute aggregation by using the average value of the same position in all channels.

Fig. 1 - Aggregated Layer object

.layerShape : String

Represents the layout of expanded feature maps of a layer.
"line" or "rect". Default "rect".
"line" linear arrangement; "rect" the "best rectangle" arrangement of visualization.
For more details about the "best rectangle", please check .createRectangleCenters.

Fig. 2 - Feature maps arranged as a filled rectangle

Fig. 3 - Feature maps arranged as a line

.layerInitStatus : String

Represents the initial status of an expandable layer. Either "expanded" as feature maps or "collapsed" as an aggregation
"close" or "open". Default "close".
"close" to show layer as an aggregation; "open" to show layer as expanded feature maps.

Fig. 4 - Initial layer with collapsed status

Fig. 5 - Initial layer with expanded status

.textSystem : String

Enable or disable the display text information of each layers.
"enable" or "disable". Default "enable".
"enable" to display text information; "disable" not to display text information.

Fig. 6 - Show texts of the layer dimension

Fig. 7 - Disabled text display

.relationSystem : String

Enable or disable the display of relation lines among layers.
"enable" or "disable". Default "enable".
"enable" to display relation lines; "disable" not to display relation lines.

Fig. 8 - Show relation lines between layers

Fig. 9 - Disabled relation line display

.animeTime : Double

Time spent for animation.
Default to 2000 (2 second).
The range of the value should be larger than 0.

Fig. 10 - Expansion animation with animeTime 2000 (default)

Fig. 11 - Expansion animation with animeTime 6000

.color : Color Format

All necessary color configurations of different layers and the background.
We can set the background color in the model configuration.
See color for more details.

Fig. 12 - Change background color Purple

.stats : Bool

Whether to use stats utility or not.
true or false. Default false.
true to show stats utility; false not to show stats utility.

.predictDataShapes : Array

Configure input shapes for pre-trained model if pre-trained model's input shape is dynamical.
For example, InceptionV3 from Keras, you can feed different size of input into the model, if you want to visualize this model in TensorSpace, must provide a specific input shape.
```
let model = new TSP.models.Model( modelContainer, {

    predictDataShapes: [ [ 299, 299, 3 ] ]

} );
```

.feedInputs : Array

Configure which inputs of pre-trained model to be applied into TensorSpace model's input.
For example, ACGAN's generative model, has two inputs random data ( tf.randomNormal([1, 100]) ) and label ( tf.randomNromal([1, 1]) ), if we only want to visualize the random data, we can config model like:
```
let model = TSP.models.Sequential({
    feedInputs: [ 0 ]
});
```
0 means the first input of pre-trained model to be applied into TensorSpace model. In this example, as pre-trained model has two inputs, default to: [ 0, 1 ].

Method

.init( complete_callback ) : void

The init() method use the provided configurations to create the actual objects of the model.
See Model (concept) for more details.

.load( load_config ) : void

The load() method is used to maintain the pre-trained ML model (from TensorFlow, Keras or TensorFlow.js).
We provide a callback function for extendable use.
See Load for more details.

.predict( input_data, complete_callback ) : void

The predict() method uses the provided input data and the loaded ML model to compute all necessary predictions including the intermediate outputs and final inferences for rendering visualization objects.
We provide a callback function for extendable use.
See Predict for more details.

.reset() : void

The reset() method is used to set the model back to its initialization status.
The camera will be moved to its initialized position. The view angle will be changed to initial default angle.
All existing prediction data in each layer object will be cleaned up. All visualizing objects will be re-rendered without any data.

.getAllLayers() : Layer[]

To get all layer objects within the model.
If we want to config all layers or check for specific configurations, we can use the method to catch all layer objects.

.getLayerByName( name ) : Layer

To get a specified layer object by its name (String).
If any specific layer object is desired to be coordinated, we can use the method to find and layer object.

For more detailed implementations of different TensorSpace models, please check: