Trainable classifier

This component predicts a label for each box over the whole document using machine learning.

Note

You must train the model your model to use this classifier. See Model training for more information

Examples

The classifier is composed of the following blocks:

• a configurable box embedding layer
• a linear classification layer

In this example, we use a box-embedding layer to generate the embeddings of the boxes. It is composed of a text encoder that embeds the text features of the boxes and a layout encoder that embeds the layout features of the boxes. These two embeddings are summed and passed through an optional contextualizer, here a box-transformer.

API-basedConfiguration-based

pipeline.add_pipe(
    "trainable-classifier",
    name="classifier",
    config={
        # simple embedding computed by pooling embeddings of words in each box
        "embedding": {
            "@factory": "sub-box-cnn-pooler",
            "out_channels": 64,
            "kernel_sizes": (3, 4, 5),
            "embedding": {
                "@factory": "simple-text-embedding",
                "size": 72,
            },
        },
        "labels": ["body", "pollution"],
    },
)

[components.classifier]
@factory = "trainable-classifier"
labels = ["body", "pollution"]

[components.classifier.embedding]
@factory = "sub-box-cnn-pooler"
out_channels = 64
kernel_sizes = (3, 4, 5)

[components.classifier.embedding.embedding]
@factory = "simple-text-embedding"
size = 72

Parameters

PARAMETER	DESCRIPTION
labels	Initial labels of the classifier (will be completed during initialization) TYPE: Sequence[str] DEFAULT: ('pollution')
embedding	Embedding module to encode the PDF boxes TYPE: TrainablePipe[EmbeddingOutput]

---