Datasets

Description

Use the teklia-layout-reader inference command to predict on a dataset.

Parameter Description Type Default

--dataset

Path to the local LayoutReader dataset directory. The directory must contain .jsonl.gz files.

string

--split

Dataset split to use. Must match the name of a corresponding archive in the --dataset directory (e.g., train → train.jsonl.gz).

string (choices: ["train", "dev", "val", "test"])

--model

Name of the LayoutReader checkpoint dataset.

string

--output-dir

Output directory where results will be saved. If --visualize, it will also plot the predicted reading order on each image.

pathlib.Path

--with-classes

Whether to use the zone classes for the prediction.

bool

false

--with-separators

Whether to use the separators for the prediction.

bool

false

--images

Path to the images (required to visualize predictions).

pathlib.Path

None

--visualize

Whether to visualize the predicted reading order. Plots will be saved in output_dir.

bool

false

Examples

Predict reading order on bounding boxes

This command will predict the reading order on the test set of the given dataset, using only bounding box coordinates.

teklia-layout-reader inference --dataset finlam_dataset/ \
    --split test \
    --model checkpoint_finlam_best/ \
    --output output_finlam \

The prediction file (predictions.json) will be saved in the --output directory.

It contains the following information for each sample:

  • "boxes": the list of input boxes.

  • "classes": the list of input classes.

  • "separators": the list of input separators.

  • "predicted_order": the predicted reading order.

  • "target_order": the target reading order (only when predicting on an annotated dataset).

  • "average_relative_distance": the average absolute difference between the predicted index and the annotated index for each box (only when predicting on an annotated dataset).

  • "visualization": the path to the visualization image (only when --visualize is used).

Predict reading order on bounding boxes using additional information (classes / separators)

By default, LayoutReader makes decision based on box coordinates, but it can also use additional information:

  • --with-classes - LayoutReader will also use the class information for each zone (ex: title, subtitle, body, advertisement…​).

  • --with-separators - LayoutReader will also use the localization of vertical and horizontal separators.

teklia-layout-reader inference --dataset finlam_dataset/ \
    --split test \
    --model checkpoint_finlam_best/ \
    --output output_finlam \
    --with-classes \
    --with-separators

Predict reading order on pre-ordered bounding boxes

The input boxes can be pre-sorted to reduce processing errors. Four methods are available via --sort-method:

  • sortxy_by_columns

  • sortxy

  • sortyx

  • random

By default, the boxes are not ordered, and will be taken as they appear in the dataset.

teklia-layout-reader inference --dataset finlam_dataset/ \
    --split test \
    --model checkpoint_finlam_best/ \
    --output output_finlam \
    --sort-method random

Predict and visualize

Use the --visualize option to plot the predicted reading order on the images.

teklia-layout-reader inference --dataset finlam_dataset/ \
    --split test \
    --model checkpoint_finlam_best/ \
    --output output_finlam \
    --with-classes \
    --with-separators \
    --images finlam_dataset/images/test/ \
    --visualize

This will create visualizations in your output directory:

example