🚀 Features

What

• Update the adjacency matrix data from to a sparse tensor to significantly reduce memory footprint.
• Move Track and concat_tracks functions from deepcell-tracking to deepcell.data.tracking. They are really just .trk preprocessing and are unnecessary outside of prepare_data.
• Update temporal_slice to not slice into padded frames.
• Bump deepcell-tracking to 0.5.0.

Why

• Continue to upgrade the tracking model to make it more usable.
• The temporal_slice fix should improve the precision metrics of the model by not training on padded data.

🐛 Bug Fixes

What

• Update numpy version constraints to match setup.py.

Why

• Fixes #559.

What

• Revert the change to prepare_dataset that allows it to process .trks files.

Why

• This change caused complications downstream (primarily in using different train and val files) and is not useful.

What

• Fix the filter_and_flatten function.
• Add the graph_layer argument to the model call.

Why

• This function was not filtering the padded data properly, fixing it resolves the training issues.
• Improve overall clarity of options for tracking models.

🧰 Maintenance

What

• Update the package version to 0.11.0.

Why

• Getting ready for the next release.

What

• Pin docutils to 0.16 to resolve readthedocs build failures
• Add a monthly downloads badge
• Clean up grammar and whitespace.

Why

• General improvement for the README

What

• Retrain tracking models using SparseTensors

Why

• Provide the latest and greatest models in the next release.

🐛 Bug Fixes

What

• Update numpy version constraints to match setup.py.

Why

• Fixes Python 3.9 compatibility issue (#559)

0.8.8

Bugfixes

• Upgrade tensorflow to v2.3.4 to fix several CVEs.

0.9.2

Bugfixes

• Upgrade tensorflow to v2.4.3 to fix several CVEs.

0.9.1

Bugfixes

• Upgrade tensorflow to v2.4.2 to fix several CVEs.

0.9.0

This release updates TensorFlow to 2.4.1 which drops support for Python 3.5.x.

Features

• Update Layers and Losses for mixed-precision training. (#490)
• Use a dict of names and num_classes to name PanOpticNet semantic heads. (#498)

Bugfixes

• Fix architectures for ScaleDetection and LabelDetection models. (#491)
• Fix SemanticMovieIterator to prevent saving duplicate images for every batch. (#492)
• Update SiameseDataGenerator and siamese_model for multichannel data. (#495)
• Deprecate in_shape argument for Location layers. (#497)
• Pin deepcell-toolbox to 0.9.x and deepcell-tracking to 0.3.x. (#500)

Breaking Changes

• Update tensorflow version to 2.4.1, drop support for Python 3.5. (#476)
• Migrate RetinaMask models to deepcell-retinamask. (#486)

0.8.7

Bugfixes

• Pin deepcell-toolbox to <0.9.0 to prevent breaking changes with multiplex_utils.py.
• Pin deepcell-tracking to <0.4.0.

0.8.6

Bugfixes

• Update sed target to fix package name in deepcell-cpu deployment script. (#485)

0.8.5

Features

• Enable deepcell.__version__. (#482)

Bugfixes

• Add logging to deepcell.applications, fix NuclearSegmentation preprocessing, and update CytoplasmSegmentation with TensorFlow 2 SavedModel. (#483)

0.8.4

Features

• Updated the MultiplexSegmentation model (#467)
• Publish deepcell-cpu on new releases (#472)
• Update tracking model, data generators, and application for TensorFlow 2 (#470, #473)

Bugfixes

• Application._resize_output handles lists of tensors and single tensor outputs (#468)

0.8.3

Features

• Migrate to opencv to opencv-headless to remove GUI dependencies. (#466)
• Update deepcell-toolbox to 0.8.3.
• Update deepcell-tracking to 0.2.7.

0.8.2

Bugfixes

• Update NuclearSegmentation model file and preprocessing.
• Removed num_semantic_heads from PanopticNet to prevent any conflicts with num_semantic_classes.
• Remove use_pretrained_weights from Application docstrings.

0.8.1

Bugfixes

• Fix Docker builds by including README.md in the Docker image. (#462)

0.8.0

This release supports TensorFlow 2.3.1+ and drops support for Python 2.7.

Features

• Migrate to TensorFlow 2.3.1 (#442)
• pip install deepcell is now supported. (#461)
• Applications load SavedModels and are decoupled from the current model architecture. (#460)

Bugfixes

• Reduce the number of installed packages in the Dockerfile (#442)
• Removed training.py from all reference notebooks (#458).
• Fixed a dtype bug for disc transforms (#442).

Breaking Changes

• Python 2.7 is no longer supported
• TensorFlow < 2.3.1 is no longer supported.

0.7.0

This will be the final release that supports Python 2.7 and TensorFlow 1.X.

Features

• Visualize RGB data with outlines (#441)
• Remove the Resize2D layer. (#443)
• Reduce ImageDataGenerator memory footprint. (#436)
• Create CellTracking application object, various application bugfixes. (#444)
• Use keras.utils.custom_object_scope and keras.testing_utils.layer_test for testing layers. (#447)
• Migrate CI/CD from TravisCI to GitHub Actions. (#449, #451, #453)
• Improve docstring RST formatting and include the RTD config file (#448)
• Update weights and post-processing parameteres for deepcell.applications.MultiplexSegmentation. (#434, #452)

Bugfixes

• Do not always save the final model weights in training.py. (#435)
• Fix typo in README. (#440)
• Add K.epsilon to the denominator in whole_image normalization. (#446)

0.6.0

Features

• Add CroppingDataGenerator to randomly crop form the source images. (#413)
• SemanticDataGenerator supports multiple labels. (#344)
• Update the MultiplexSegmentation application with new weights and post-processing (#407, #413, #422, #424, #425, #429, #431)
• Added the SemanticMovieDataGenerator for 3D data. (#419)

Bugfixes

• Updated README typos and URLs. (#401, #402)
• Move compute_overlap to deepcell_toolbox. (#409)
• Improve performance of pixelwise transform. (#415)
• Fix PanopticNet model for 3D data. (#416)
• Pin version of opencv for better python compatibility (#422)

0.5.0

This release supports TensorFlow 1.14.x and 1.15.x as well as python 2.7 and 3.6+.

Features

• Updated match_nodes to return IoU directly instead of indices. (#267)

• Added get_anchor_parameters to automatically determine feature pyramid parameters (#269)

• Added new custom layer, ConvGRU2D (#278)

• Updated layer_test testing routine (#279)

• Travis will now tag and push a latest-gpu image with every release. (#281)

• Speed up the pixel-wise transform (#286, #295)

• Add temporal information options to the featurenets. (#282)

• Improve docstrings for sphinx compatibility. (#310)

• Improved data quality for cytoplasm and phase data. (#318)

• Added new model_zoo.PanopticNet and SemanticDataGenerator to generalize a model for learning multiple tasks simultaneously, both regression and classification. (#319)

• Added Application objects to easily use models with a simple API. (#341)

• Simplified transform names (#376).

Bugfixes

• deepcell-tracking has been updated to 0.2.4, which resolves some ISBI function bugs. (#267)

• Updated tf.image.resize_images to tf.image.resize as the former is deprecated. (#268)

• Correct upper limit for clipping boxes (#277)

• Fixed broken data links and README links (#292, #300, #307)

• Migrate general utility functions into a new package deepcell-toolbox (#319).

• Fixed RetinaNet interpolation bug (#357)

Breaking Changes

• Support for TensorFlow 1.10.x - 1.13.x has been dropped.

• The default TF_VERSION in the Dockerfile has been updated to 1.14.0-gpu, as many users were expecting this. (#281, #311)

• MaskRCNN has been refactored to RetinaMask (#360).

• /scripts has been migrated to /notebooks (#374).

• deepcell.notebooks has been removed (#390).

0.4.0

This release fully supports Tensorflow 1.10.x through 1.14.x, and Python 2.7, 3.5, 3.6. Future releases will drop support for TensorFlow 1.10.x, 1.11.x 1.12.x, 1.13.x as well as dropped support for Python 3.5.

Features

• Replaces tracking.py and tracking_utils.py with a dependency on the pip package deepcell_tracking. (PR #254)

• Break image_generators.py into a submodule with each family of generators in a different file. (PR #258)

Bugfixes

• Updated reshape_matrix to work with non-square matrices. (PR #257)

• Fixed namespace imports. (PR #262)

• 3D FeatureNet models can now be re-instantiated with a new frames_per_batch value. (PR #250)

Breaking Changes

• Updates the ImageNormalization layers to not have trainable weights (PR #250), which unfortunately means that models trained in versions <0.4.0 cannot be loaded with version 0.4.0+.

Known Issues

• While this release supports TensorFlow 1.12.x and 1.13.x, there have been compatibility issues with these versions in the past (#244 and #245).

frames_per_batch was added to both RetinaNet and RetinaMask which have both been adapted for 3D data if frames_per_batch is greater than 1.

Other bugs were fixed:

• Issue #225 was resolved, enabling compatibility testing with TensorFlow 1.14.0.
• Issue #219 was also resolved and tests were improved to ensure sequential labels for reshaped data.
• Tests for deepcell.model_zoo and deepcell.applications were also significantly improved by parameterizing tests instead of using a for-loop.

Compatible with TensorFlow versions 1.10+, this release forms the basis for the cell tracking and benchmarking algorithms put forth in the publication "Accurate cell tracking and lineage construction in live-cell imaging experiments with deep learning" (https://doi.org/10.1101/803205).