Version 1.3.1

This is a minor release with bug fixes and general quality of life improvements.

FIles in this release

• quevedo-1.3.1-py3-none-any.whl: wheel for installing quevedo with pip.
• quevedo-1.3.1.tar.gz: sdist file for when wheel is not an option.

Version 1.3

This release adds logogram graphs and some quality of life improvements for the annotation process.

FIles in this release

• quevedo-1.3.0-py3-none-any.whl: wheel for installing quevedo with pip.
• quevedo-1.3.0.tar.gz: sdist file for when wheel is not an option.

New dataset version 2

There are four annotation schemas now, one for graphemes, one for logograms, another one for graph edges, and a final one for meta tags. The logogram annotation schema is new, but previously this information could be stored in the meta tags (now they are separate). See below for more about graphs. The new dataset version in not compatible, so existing datasets need to be migrated.

Logogram graphs

Logograms can now store a graph which connects bound graphemes with edges. These edges have their own tags and tag schema, and can be used to capture the compositional or spatial relations between bound graphemes. The dataset configuration has been updated, as well as the web interface, to allow annotating these new values.

Other

• Configuration dictionaries are now merged instead of replaced in network, pipeline and dataset configs.
• "Extend" configurations can be chained.
• In the web interface listing, annotations can be filtered according to flags.
• The color list to use in the logogram editor can now be customized.
• Other bug fixes and documentation

Version 1.2

This release adds pipelines, which allow the user to build complex systems that combine the networks and expert knowledge codified in quevedo datasets. See the documentation for more.

Files in this release

• quevedo-1.2.0-py3-none-any.whl: wheel for installing quevedo with pip.
• quevedo-1.2.0.tar.gz: sdist file for when wheel is not an option.

Changes

• New Pipeline objects.
• When creating an in-memory annotation, convert image to RGB.
• Remember last function selected in the web app edit page.
• If config key darknet.shutup is false, no stdio redirection is done.
• Allow in-memory images for annotations.
• Lambda function as pipeline step and improved test command.
• Global default config value substitutes default config per net.
• Simplified evaluation code: overall, detection and classification accuracy.
• Add an "extend" key to net configs to allow sharing config values.
• Improve user scripts init function.
• Use Graphemes and Logograms instead of dicts.
• Improved test command with more info and better algos.
• Allow training detector with "no" tag (same tag for every object).
• Allow use of relative paths for darknet library.
• Added example toy arithmetic dataset to the repo.
• Other bug fixes.

Version 1.1

Version 1.1 comes with some minor quality of life improvements, but a new an incompatible dataset schema that will need existing datasets to be migrated.

Files in this release

• quevedo-1.1.0-py3-none-any.whl: wheel for installing quevedo with pip.
• quevedo-1.1.0.tar.gz: sdist file for when wheel is not an option.
• toy_arithmetic.zip: toy dataset for demonstration purposes.

Changes

New dataset version 1

Before, Quevedo datasets were not versioned. Now, a field has been added to the config.toml file to track the Quevedo dataset schema, and a migrate command has been added to let users upgrade datasets to new versions.

Some of the following additions are incompatible changes to dataset functionality and annotation files which have made this version upgrading necessary.

Tags are now a dictionary

Grapheme tags (both free and bound) are now represented with a dictinary, with keys the names in the dataset tag_schema. This makes Annotation objects easier to use in custom code. The change affects both the library code and the files on disk, hence the migration.

Splitting now works differently

Instead of assigning annotations to either a train or test split, they are assigned to a "fold". Groups of folds can then be defined as being available for train or for test (or none). This is also an incompatible change to annotation code and file representation. Old partitions will be lost, so after migration you will need to run the split command again.

Net configuration improvements

• Detection networks can now have width and height parameters to tune network input size.
• All networks can now have a max_batches parameter to customize when to stop training the net. This can serve to prevent overfitting and shorten training times.

Annotation flags

A new option "flags" has been added to the config.toml file. These flags are matadata values just like those in meta_tags, so assigned to both Logograms and Free Graphemes. The difference is that they are presented as checkboxes in the web interface, and shown as icons in annotation listings. This can serve to quickly mark annotations for annotators, for example if some have dubious or problematic tags, need some other kind of attention, or simply you want to keep track of them.

Other

• When building the tag map for darknet, user tags are combined using the ASCII FS character instead of "", which can be problematic if tag values in the dataset contain "". This is an internal change and user code and data should not be affected.
• The dataset.get_network method now returns the same Network object if called many times with the same network name. This helps save memory, which in the case of neural networks can be crucial, without requiring the user to keep the network in their own variable.
• The web interface now can be used with touch on mobile devices.

Version 1

Quevedo is a python tool for managing datasets of images with compositional
semantics, with a focus on the training and evaluation of machine learning
algorithms on these images.

Quevedo is part of the VisSE project.

This release marks the version developed for the project, used for the datasets
and experiments being worked on.

Files in this release

• quevedo-1.0.0-py3-none-any.whl: wheel for installing quevedo with pip.
• quevedo-1.0.0.tar.gz: sdist file for when wheel is not an option.
• toy_arithmetic.zip: toy dataset for demonstration purposes.