Changes to cytounetΒΆ
Version 0.2.2
Fixed a bug in setting test paths when using script mode.
Versioning is now automated, as is linking to the GitHub release. Please ensure you release in the form
v#versionnumberhere
.Fixed issues with script mode
Using
tensorflow.Keras
instead ofKeras
.Added sanity checks to ensure paths actually exist.
Version 0.2.1
Extended script to handle fine-tuning and from scratch-training
Added a script only mode.
Added support for docs.
Added original a549 sample data, notebook, and pre-trained weights.
Added experimental results to the README.
Fixed issues with original images being overwritten. It is now possible to return a copy of non overwritten images.
Made
draw_contours
more flexible. Specifically, it is now possible to turn off text display as this makes the image crowded.Added
find_contours
anddraw_contours
, useful methods for area determination.Added
read_image_spec
for use only for post modeling processing. This fixes issues with incorrect shapes when usingread_images
Version 0.2.0
Kernel regularization can now be turned off via a boolean argument(use_regularizer)
Added a new data set from BBBC.
finetune
is a new function dedicated to the finetuning workflow.Regularization is now supported. It is currently limited to L1 and L2.
pretrained_weights
was dropped as an argument tounet
. Use acallback
instead. A future version wil include a fine tuning function.save_as
was removed fromtrain
. Use ModelCheckpoint instead and provide it as a callback.show_images
now shows titles. These functions will be removed later and imported frompyautocv
instead.Fixed issues with reading mixed
jpg
andpng
images.Added
reshape_images
andresize_images
. These are helper functions that may be useful when plotting or restoring original image size.show_images
andread_images
are now imported frompyautocv
>= 0.2.2Fixed issues with inconsistent image order in
show_images
when reading from a directory.Added filename printing to data generators to make it easier to show what order the files are being read in. This can be disabled by setting
show_names
toFalse
.Changes to prediction generation were made. We now use
ImageDataGenerator
for test time data generation.Fixed a bug related to
load_augmentations
that led to image flipping.Changed outputs to
sigmoid
instead ofReLU
Updated to latest API ie
predict
vspredict_generator
Added
train
to simplify model fitting.Added
predict
to reduce code repetition and make predicting easier.unet
was rewritten to increase complexity and solve issues with blank predictions. It now also usesConv2DTranspose
instead ofUpSampling2D
.Initial support for a simpler model to optimise the bias-variance trade off for small(er) datasets.
Removed
Dropout
since this is known to have no improvement over Batch Normalisation.Initial support for SGD as the default optimiser
Moved from camelCase to snake_case, now using more descriptive function names.
Fixed issues with list input to
show_images
Release 0.1.0
Renamed repository to
cytounet
to reflect the heavy focus on biological images.Initiated support for validation via
validGenerator
.Fixed issues with
show_images
failing to loadnumpy
ndarray
images.
Initiated ability to install with
pip
andsetup.py
.
show_augmented
was renamed toshow_images
and refactored as a more general method not limited
to just augmented images. A cmap
argument was also added for more flexibility. This replaces labelVisualize
which has now been dropped.
Introduced a separate save method for images and predictions. Use
saveImages
andsavePredictions
respectively.
Fixed issues with information loss following saving of predictions.
geneTrainNpY
was refactored and renamedLoadAugmented
Added
thresholdImages
to threshold masks(mostly). Please see pyautocv for a more general and flexible way to manipulate images.Added
saveImages
, a helper to save images as(by default).tif
. This is because biological images are normally tiff in nature.Removed
savePredictions
. UsesaveImages
instead.
Updated module documentation
adjustData
was removed since it had known issues. It may be restored in the future.Fixed issues that resulted in blank predictions
Added
show_augmented
to show results of data augmentationAdded
BatchNormarmalisation
stepsTraining made more flexible by allowing usage of different metrics and loss functions without editing source code(i.e change on the fly)
Saving and image reading functions made more flexible to read/save any image file format.
Made most functions compatible with Keras >= 2.0
Added
dice
loss and dice coefficient.