Skip to content

Latest commit

 

History

History
160 lines (115 loc) · 9.42 KB

API_CHANGELOG.md

File metadata and controls

160 lines (115 loc) · 9.42 KB
Raw

Breaking changes to the NuPIC API

We try to keep the API compatible with the original numenta/nupic.core repo. That API is specified in their API Docs.

Motivation

We try to remain compatible where possible, to make it easy for the users and programmers to switch/use our implementation. And for developers to be easily able to navigate within the (known) codebase. Despite of this, sometimes changes need to happen - be it for optimization, removal/replacement of some features or implementation detail, etc.

Summary

The NetworkAPI is mostly unchanged, since it is a compatibility layer. Direct access to the algorithms APIs has changed:

  • All classes now use the SDR class where applicable instead of raw data lists.
  • All encoders have new API.
  • SDRClassifier split into two classes: Classifier and Predictor with new API.
  • Anomaly class is now built into the TemporalMemory
  • SpatialPooler & TemporalMemory have many small changes, see below.
  • String parameters are defined in the Spec as type "String" rather than as a byte array.

API breaking changes in this repo

Compared to Numenta/nupic.core; the changes here are listed in order from oldest to newest (at the bottom).

  • CapnProto serialization is replaced with binary streams in PR #62.
    Calls to read() and write() are no longer available. Use save() and load(). Network(path) is no longer used to deserialize a path. Use Network net; net.load(stream); to deserialize.
    Helpers SaveToFile(path) and LoadFromFile(path) are used to stream to and from a file using save() and load(). This was later changed to use Cereal Serialization.

  • The function Network::newRegionFromBundle() was replaced with newRegion(stream, name)` where the stream is an input stream reading a file created by region->save(steam) or region->saveToFile(path). PR#62

  • Removed methods SpatialPooler::setSynPermTrimThreshold and SpatialPooler::getSynPermTrimThreshold. Synapse trimming was an optimization which is no longer possible because of an implementation change. PR #153

  • Removed method SpatialPooler::setSynPermMax as the maximum permanence is now defined (hardcoded) as nupic::algorithms::connections::maxPermancence = 1.0f; PR #153

  • Changed callback ConnectionsEventHandler::onUpdateSynapsePermanence(). Instead of being called every time a synapses permanence changes, it is now called when a synapse changes connected state, IE: it is called when a synapses permanence crosses the connected threshold. PR #153

  • SpatialPooler now always applies boosting, even when learn=false. PR #206

  • Removed methods SpatialPooler::setSynPermConnected and TemporalMemory::setConnectedPermanence. The connected synapse permanence threshold should instead be given to the constructor or the initialize method. PR #221

  • When building with MS Visual Studio 2017, it will build bindings only for Python 3.4 and above.
    (i.e. No Python 2.7 under Windows)

  • Setting dimensions on a region is now optional. If given, it overrides any region parameters that set the width of the default output buffer.

  • The splitter maps (and the LinkPolicy that creates them) were removed. These were used as a way to re-arrange bits in the input buffer based on user defined patterns. However, thinking about how this works in biology, the neurons make connections to the synapse of other neurons. There is no order or pattern involved and in fact it needs to be fairly random as to how the neurons connect. It is the job of the Spatial Pooler to simulate that set of random connections. So in effect, any re-mapping of bits by a splitter map prior to being presented to the Spatial Pooler should have no affect on functionality. This is probably why this feature was not used anyplace except in the unit tests. As a side-effect of this change, the LinkType and LinkParam parameters in the Network.Link( ) call are ignored. PR #271

  • Removed all matrix libraries. Use the Connections class instead. PR #169

  • Removed void SpatialPooler::stripUnlearnedColumns() as unused and not useful (did not effectively remove any columns). PR #286

  • Rewrote ScalarEncoder API, all code using it needs to be rewritten. PR #314

  • Removed old TP (Temporal Pooler, Cells4.hpp) as it was not maintained, users should default to TemporalMemory, TM. With this we are also removing BacktrackingTM (and its NetworkAPI Region), which was based on TP. BackTM had slightly better anomaly scores results (+5% compared to TM), but did not have complete tests and verified (peer-reviewed) functionality. PR #356

  • Connections class must be initialized with a connectedPermanence. Methods Connections::computeActivity and Connections::raisePermanencesToThreshold no longer accept a synapse permanence threshold argument. PR #305

  • SDRClassifier class is replaced by Classifier and Predictor classes.

  • In NetworkAPI, access to a Region object was accessed using net.getRegions()->getByName('name');. This is obsolete. Use getRegion('name') instead.

  • Anomaly class removed as obsolete, use TM.anomaly which is simpler to use, and MovingAverage when you need to emulate running averages. Internally the code still uses computeRawAnomalyScore() but there's no need to call it directly. AnomalyLikelihood is still available and can be used in addition to TM.getAnomalyScore(). PR #406

  • TemporalMemory::getPredictiveCells() now returns a SDR. This ensures more convenient API and that the SDR object has correct dimensions matching TM. use TM.getPredictiveCells().getSparse() to obtain the sparse vector as before. PR #437, #442

  • TemporalMemory compute() and activateCells() now use only SDR variants, old overloads with C-style arrays removed. Bindings and tests also updated.

  • Changed all use of "nupic" to "htm". This means that C++ users must include from

    Currently Previously
    <htm/algorithms/*.hpp> <nupic/algorithms/*.hpp>
    <htm/engine/*.hpp> <nupic/engine/*.hpp>
    <htm/math/*.hpp> <nupic/math/*.hpp>
    <htm/encoders/*.hpp> <nupic/encoders/*.hpp>
    <htm/types/*.hpp> <nupic/types/*.hpp>

    We also renamed the namespaces from namespace nupic to namespace htm.

  • SpatialPooler: removed param numActiveColumnsPerInhArea, as replaced by localAreaDensity which has better properties (constant sparsity). PR #549

  • SpatialPooler: compute() now returns overlaps. SP.getOverlaps() removed. PR #552

  • Region: GetInput() and GetOutput() now return std::shared_ptr's rather than raw pointers.

  • Name changes:

    Original New
    RDSERegion RESEEncoderRegion
    ScalarSensor ScalarEncoderRegion
    VectorFileEffector FileOutputRegion
    VectorFileSensor FileInputRegion

  • Timing for data move PR# 956 Originally data is moved from output to input through the links just before executing a region. Now, the data is moved just after execution of a region. The effect is that the internal values of the Input buffers between iterations was the value just used in the previous execute. It is now the input buffer contains the value to be used by the next execution.

    There should be no effect on applications unless they are looking at the values of the internal input buffers.

Python API Changes

Changes made to the C++ Library also effect the Python Library, since python is mostly just a thin wrapper around the C++ library.

  • Serialization using canproto was removed. This was replaced with Cereal Serialization and is available via saveToFile() and loadToFile(). Pickle of a component imported from C++ will cause Cereal serialization. So Python apps should just use Pickle for serialization.

  • Changed all use of "nupic" to "htm". This means that Python users must import from

    Currently Previously
    htm.bindings.algorithms nupic.bindings.algorithms
    htm.bindings.engine_internal nupic.bindings.engine_internal
    htm.bindings.math nupic.bindings.math
    htm.bindings.encoders nupic.bindings.encoders
    htm.bindings.sdr --

  • Most algorithms now accept SDR's instead of numpy arrays. Recommend reading the documentation, see python -m pydoc htm

  • Parameters containing strings were originally defined in the NetworkAPI Spec as a byte array (i.e. type "Byte" and count 0). Byte arrays are now 8 bit integers and strings use type "String" and count 1.

  • Classifier::learn(SDR, label) The label argument can now be an unsigned integer (the label index) or it can be a vector containing a set of label indexes that relate to this pattern. This was done because syntax such as {4} passed as the label, intended to create a vector with one element, is now being rejected by at least one compiler.
    So, just pass the label index directly if there is only one.

  • AnomalyLikelihood class has been rewritten and its API simplified. PR #958