Categorical¶

Categorical data can take on one of a limited number of enumerated values.

In most applications, categoricals are used for ground truth and predicted class labels, either as standalone columns in classification tasks, or labels embedded within each instance in tasks like bounding box object detection.

Schema¶

Categorical labels are represented by a tlc.Schema<tlc.core.schema.Schema> with a value map, mapping class indices to tlc.MapElements, containing the name, description and color of the fields, and where the value is a numerical value with a number role set to "label". While the builtin tlc.CategoricalLabelSchema can be used in most cases, the full schema example can be used when further customization is desired.

categorical_schema = tlc.CategoricalLabelSchema(classes={0: "dog", 1: "cat"})

import tlc

value_map = {0.0: tlc.MapElement("dog"), 1.0: tlc.MapElement("cat")}
categorical_schema = tlc.Schema(
    value=tlc.Int32Value(
        value_map=value_map,
        value_min=0,
        value_max=1,
        number_role="label",
    ),
)

Visualization and Editing¶

In this example, the Label column is an editable categorical label column, with an enumerated mapping between numerical value and its string representation.

Edit a single label¶

To edit a label, click on the label and choose the label that you want to change to in the popup window.

Edit multiple labels¶

To edit multiple samples at once, select the rows, click on the label in one of the selected rows, and choose the label to be changed to.

Assign label predictions¶

To assign the model’s predictions to the labels, follow these steps:

  1. Filter on a single epoch if multiple epochs are recorded

  2. Apply filters (e.g., accuracy=0), if applicable, to narrow down to the samples that you would like to assign predictions to labels for

  3. Click and drag the Predicted_Label header to the Label header (the receiver header turns yellow while dragging)

Operation: Assigning values from one column to another

Assigning values from one column to another is not only for labels, but also may be done for any compatible column pair. Columns are compatible if they have the same data type and the receiver is an editable column.

Here is an example of assigning Loss to Weight. This can be useful when you want to assign sample weights based on a metric like Loss.

Warning: One epoch for assigning operation

The assigning operation requires that only one epoch to be filtered in if multiple epochs are recorded. This is because the column to be assigned is a part of the dataset Table and each sample can only accommodate a single value. In contrast, metrics collected in a Run can be multi-values across epochs for a given sample. Therefore, assigning metrics from multiple epochs to a single sample would violate the properties of the Table.