The Processing components process and transform the data:


Re-dimensions input data into a form more suitable for a subsequent component. This component's input socket is often connected to a Data component which provides the source of data, while its output socket is connected to a Deep Learning component such as a Convolution component which expects data in a certain dimension (e.g., 2D) and size. In the Basic Image Recognition tutorial for example, a Reshape component converts a one dimensional input array of 784 pixels into a 28x28x1 array which represents the pixels a 2D image.
Reshape only works if the input dimension matches up to the output dimension and it will re-order the datapoints/pixels.
  • Reshape: sets the new shape of the data by allowing you to specify the size of each dimension in up to three dimensions.
  • Transpose map (readonly): provides a visual representation of the new shape that will be applied.


Provides TensorFlow code to convert an RGB (888) image data to grayscale. It's best used in models which are provided with RGB data but don't need the colors to make predictions. Grayscaling will then reduce the complexity of the model with little processing.
This component's input socket is typically connected to an Data component.The output socket is often connected to a Dense or a Convolution component, but can go to almost any other component as well (other components for processing, merging, etc.).


Provides TensorFlow code to perform one hot encoding. One hot encoding effectively removes ordinality from classifications so that each class (e.g., type of label) has a unique ID but no inherent order in relation to other classifications. This allows a classifier to perform pure mathematical operations such as averaging classes, without any inherent ordering or relationship between those classifications.
A OneHot component's input socket is often connected to a Data component containing a source of input labels. It's output socket is then connected to a Classification component and the OneHot component removes any ordinality of the labels in preparation for the Classification component to classify data. See the Basic Image Recognition tutorial for an example.
  • Number of classes: specifies the number of classes (categories) in the data. For example, in the Basic Image Recognition tutorial, the labels represent the digits 0 through 9, so the number of classes in this case must be set to 10.


Rescales image data to the desired width and height in pixels (e.g., a 28x28 image can become 56x56).
Rescaling is performed using interpolation which allows for new sizes not based off of the original size. This component can be connected to any component which outputs an image (e.g., commonly a Reshape or Data component; in some cases a Convolutional component).
  • Width * Height: the width and height to rescale the image to.
Last modified 1yr ago