Skip to content
This repository has been archived by the owner on Aug 15, 2024. It is now read-only.

Latest commit

 

History

History
185 lines (140 loc) · 5.38 KB

Changelog.md

File metadata and controls

185 lines (140 loc) · 5.38 KB
Raw

Changelog

3.7.5

Released Wed September 12, 2018

  • Re-wrote reset_weights; just recompiles model
  • Fixed error in gridfont loader
  • All widgets/pictures are JupyterLab compatible
  • Added support for dynamic_pictures on/off; default is off
  • SVG arrows are now curves
  • New algorithm for bank layout in SVG
  • moved dataset.get() to Dataset.get() and net.get_dataset()
  • new virtual datasets API, including vmnist, H5Dataset (remote and local)
  • better cache in virtual datasets
  • Allow dropout to operate on 0, 1, 2, or 3 whole dims
  • Added cx.Layer(bidirectional=mode)
  • Show network banks as red until compiled
  • Rewrote and renamed net.test() to net.evaluate() and net.evaluate_and_label()
    • net.evaluate() for showing results
    • net.evaluate_and_label() for use in plots

3.7.4

Released Sun August 19, 2018

  • net.pp() gives standard formatting for ints and floats
  • Allow negative position in virtual dataset vectors
  • Fixed error in colors dataset that truncated the target integer to 8 bits
  • Add internal error function to net.compile(error={...})
  • New spelling: ConX
  • cx.image_to_array() removes alpha
  • vshape can be three dimensions (for color images)
  • some new image functions: image_resize(), image_remove_alpha()
  • renamed "sequence" to "raw" in utils
  • Added cx.shape(summary=False), cx.get_ranges(array, form), and get_dim(array, DIMS)
  • Use kverbose in train() for all keras activity

3.7.3

Released Mon August 13, 2017

  • Allow bool values with onehot
  • Unfix fixed crossentropy warning
  • Allow datasets to be composed of bools
  • added temperature to choice()
  • Added net.dataset.inputs.test(tolerance=0.2, index=True)

3.7.1

Released Fri August 10, 2018

  • Separate build/compile --- compile() no longer resets weights;
  • added net.add_loss()
  • Remove additional_output_banks
  • refactor build/compile
  • add LambdaLayer with size
  • add prop_from_dict[(input, output)] = model

3.7.0

Released Tue Aug 7, 2018

  • Allow additional output layers for network
  • Fix: crossentropy check
  • added indentity layer for propagating to input layers
  • Include LICENSE.txt file in wheels

3.6.10

Released Thu May 17, 2018

  • delete layers, IncepetionV3, combine networks
  • ability to delete layers
  • ability to connect two networks together
  • rewrote SVG embedded images to use standard cairosvg
  • added inceptionv3 network
  • cx.download has new verbose flag
  • fixes for minimum and maximum
  • import_keras_model now forms proper connections
  • Warn when displaying network if not compiled then activations won't be visible
  • array_to_image(colormap=) now returns RGBA image

3.6.9

Released Fri May 4, 2018

  • propagate_to_features() scales layer[0]
  • added cx.array
  • fixed (somewhat) array_to_image(colormap=)
  • added VGG16 and ImageNet notebook
  • New Network.info()
  • Updated Network.propagate_to_features(), util.image()
  • Network.info() describes predefined networks
  • new utility image(filename)
  • rewrote Network.propagate_to_features() to be faster
  • added VGG preprocessing and postprocessing
  • Picture autoscales inputs by default
  • Add net.picture(minmax=)
  • Rebuild models on import_keras
  • Added VGG19
  • Added vgg16 and idea of applications as Network.networks
  • Bug in building intermediate hidden -> output models

3.6.7

Released Tue April 17, 2018

  • Fixed bug in building hidden -> output intermediate models

3.6.6

Released Fri April 13, 2018

  • Added cx.view_image_list(pivot) - rotates list and layout
  • Added colors dataset
  • Added Dataset.delete_bank(), Dataset.append_bank()
  • Added Dataset.ITEM[V] = value

3.6.5

Released Fri April 6, 2018

  • Removed examples; use notebooks or help instead
  • cx.view_image_list() can have layout=None, (int, None), or (None, int)
  • Added cx.scale(vector, range, dtype, truncate)
  • Added cx.scatter_images(images, xys) - creates scatter plot of images
  • Fixed pca.translate(scale=SCALE) bug
  • downgrade tensorflow on readthedocs because memory hog kills build

3.6.4

Released Thur April 5, 2018

  • changed "not allowed" warning on multi-dim outputs to "are you sure?"
  • fix colormap on array_to_image; added tests
  • fix cx.view(array)
  • Allow dataset to load generators, zips, etc.

3.6.3

Released Tue April 3, 2018

  • Two fixes for array_to_image: div by float; move cmap conversion to end
  • Protection for list/array for range and shape
  • from kmader: Adding jyro to binder requirements

3.6.2

Released Tue March 6, 2018

  • added raw=False to conx.utilities image_to_array(), frange(), and reshape()

3.6.1

Released Mon March 5, 2018

  • SVG Network enhancements
    • vertical and horizontal space
    • fixed network drawing connection paths
  • save keras functions
  • don't crash in attempting to build propagate_from pathways
  • added binary_to_int
  • download() can rename file
  • fixed mislabeled MNIST image
  • better memory management when load cifar
  • Network.train(verbose=0) returns proper values
  • labels for finger dataset are now strings
  • added labels for cx.view_image_list()
  • fixed bug in len(dataset.labels)
  • added layout to net.plot_layer_weights()
  • added ImageLayer(keep_aspect_ratio)
  • fixed bugs in datavector.shape
  • added dataset.load(generator, count)
  • fixed bugs in net.get_weights()/set_weights()
  • added network.propagate_to_image(feature=NUM)

3.6.0

Released Mon Feb 12, 2018. Initial released version recommended for daily use.

  • fixed blurry activation network pictures
  • show "[layer(s) not shown]" when Layer(visible=False)
  • added fingers dataset