Changelog

0.6.4 (current github version)

Install with:

pip install git+https://github.com/jahuth/convis.git

Changes:

• fixed bug in convolution filter alignment in Conv2d and Conv3d: odd sized filters are now centered correctly

• fixed two bugs in LeakyIntegrateAndFireNeuron and RefractoryLeakyIntegrateAndFireNeuron:

  • The default time step from convis.default_resolution was not copied to the internal variable tau. Now layers will use the value from the time of their creation.
  • The input was not normalized with the step size, leading to different results when changing the resolution and leak simultaneously.

• added flip argument (default: True) to convis.filters.Conv3d.set_weight() to keep the filter and impulse response aligned.
• fixed bug in plot_impulse (impulse was too long due to padding)
• added Difference Layer convis.filters.Diff
• added spiking layers convis.filters.spiking.Poisson and convis.filters.spiking.IntegrativeMotionSensor, which is a DVS like Layer
• added convis.streams.ProcessingStream and convis.streams.MNISTStream (and neuromorphic versions: PoissonMNISTStream and PseudoNMNIST)
• fixed bug in variable describe()
• added usage docs about inputs and outputs

All older versions are available from PyPI. Install the most recent stable version (currently 0.6.3) with:

pip install convis

0.6.3

• GanglionSpiking had a bug in the refractory period
• the different implementations of OPL Layers now produce roughly the same output
• FullConvolutionOPLFilter now creates a filter from configuration options

• run() now always returns an Output object

  • fixed a bug where it returned an Output object wrapped in a Output object

• run() can now process infinite streams if it gets an argument max_t
• added warnings to convis.streams classes that are unstable
• fixed convis.streams.ImageSequence
• convis.utils.plot_tensor() is now available as convis.plot_tensor

0.6.2

A small amount of bug fixes:

• Fixed a bug in convis.filters.retina.GanglionInput:

  • the convolution filters now initialize without having to reapply a configuration to the layer/parameters

• made padding flags in convis.filters.Conv3d default to True

  • in contrast to PyTorch Conv3d the output will now have the same shape as the input by default. To disable, use time_pad=False and autopad=False as arguments.

• fixed a bug when calling array() on convis.base.Output.
• fixed a bug where resolution was not a valid argument to convis.filters.Conv3d.gaussian()

0.6.1

• It is now possible to disable the computational graph see these examples

• The documentation now has a section about changing the global configuration parameters.

  • default_resolution now lives directly in convis, no longer in convis.variables
  • other submodules now use convis._get_default_resolution() to get the current default_resolution to avoid copies that don’t update

• some issues with creating inline plots in Python 3 were fixed

• feeding an Output object to a run() function will now take the first output and process it as input

  • this way, it is now easier to continue a graph through multiple models

• a new submodule convis.layers now contains all Layers from convis.filters and convis.models! If you are unsure whether something is only a “filter” or a already a “model”, they can all be found in the same module now. Still, convis.filters and convis.models will continue to be available separately.

• added an unfinished class convis.models.Dict which will be similar to convis.models.List, but can name its layers arbitrarily.

  • possibly both classes will be merged and replace the convis.models.Sequential and convis.models.Parallel classes

0.6

• This version fixed an issue with the new PyTorch version (>0.4)