Skip to content

Update akida and cnn2snn to version 2.2.2

New features:

  • [akida] Dense inputs now supported for multipass models
  • [akida] Dense outputs supported for FNP
  • [akida] Dense outputs supported for CNP activations
  • [akida] new version CLI
  • [akida] new engine CLI
  • [akida] engine library included as an asset in akida package
  • [akida] added support for multiple devices on the same host
  • [engine] fit is now supported (single-pass only)

API changes:

  • [akida] forward CLI is removed
  • [cnn2snn] Minimum Tensorflow version is now 2.8.0

Update akida_models to 1.1.3

  • updated CNN2SNN minimal required version to 2.2.2
  • ImageNet preprocessing refactored and made public
  • included a calibration step in YOLO detection models training pipelines
  • retrained VGG/ImageNet model with a gamma constraint on batch normalization layers to be hardware compatible

Documentation update

  • examples can now be run on RPi

Update akida and cnn2snn to version 2.2.1

New features:

  • [akida] statistics are averaged by batch
  • [akida] add support for new PCIe driver
  • [cnn2snn] new PTQ calibration CLI

API changes:

  • [akida] Concat layer removed
  • [akida] metrics are now directly related to the Model
  • [akida] predict and evaluate are now interverted (Keras naming)
  • [akida] Model.predict_classes replaces legacy predict

Update akida_models to 1.1.2

  • updated CNN2SNN minimal required version to 2.2.1
  • fixed imagenette pretrained helper SHA256

Documentation update

  • updated examples with the predict/evaluate swap
  • removed mentions to Concat layer
  • fixed broken link in examples
  • fixed a MetaTF installation issue that was due to imageio dependency
  • fixed an issue in KWS example following an update of scikit-learn

Update akida and cnn2snn to version 2.1.6

BREAKING CHANGE:

  • akida models starting with an 8-bit convolution do not invert their output spatial dimensions anymore.

New features:

  • [cnn2snn] Post-Training-Quantization calibration tools
  • [akida] added a tool to transpose legacy fbz weights

Bug Fixes:

  • [akida] increase reliability of INA power measurements

Update akida_models to 1.1.1

  • updated CNN2SNN minimal required version to 2.1.5
  • pruning tools can now handle quantized models
  • pruning tools refactored to prune layers in reverse order and fixed an issue on how acceptance is computed
  • CIFAR10 DS-CNN and VGG models removed in favor of an AkidaNet 0.5 model obtained with transfer learning
  • added a Visual Wake Words (VWW) model based on AkidaNet 0.25 R96 transfer learning
  • reworked GXNOR/MNIST model to be hardware compatible
  • updated the DVS models and training scripts for better Akida performances
  • 'dv' dependency package is now checked at runtime to prevent an installation issue on windows
  • DVS preprocessing script now saves images instead of events

Documentation update

  • added new elements to API documentation
  • updated HRC pooling constraint
  • updated references to CIFAR10 models in akida_models user guide
  • updated zoo performance page with existing models changes and new models
  • removed the CIFAR10 inference example (AkidaNet/ImageNet inference being the reference example)
  • added extra information on YOLO transfer learning step in the dedicated example

Update akida and cnn2snn to version 2.1.3

New features:

  • [akida] Allow the retrieval of the last layer weights using the akida engine API
  • [cnn2snn] Support ReLU with a max_value higher than 6

Bug Fixes

  • [akida] Discard null power measures (INA231 cold boot issue)

Documentation update

  • user guide/akida/hardware mapping now has additional information about ClockMode usage
  • advanced CNN2SNN tutorial updated with changes on quantized activation

Update akida and cnn2snn to version 2.1.2

New features:

  • [akida] HardwareDevice Statistics
  • [akida] power measurements for rev5 boards
  • [cnn2snn] Reduce Depthwise weights Post-Training quantization errors
  • [cnn2snn] New reshape transformation

Documentation update

  • CNN2SNN tutorial will now produce a model that is HW compatible
  • edge tutorial using KWS dataset improved to use akida model when computing average spikes
  • added a 'transforms' section to CNN2SNN API documentation

Documentation update

  • added an hardware performance retrieval section in user_guide/akida
  • rebased the AkidaNet/Imagenet inference example on AkidaNet 0.5
  • AkidaNet/ImageNet inference example now has a last section with power information retrieval
  • cats vs. dogs transfer learning tutorial rebased on PlantVillage and reworked to make it more concise and more straightforward

Update akida and cnn2snn to version 2.1.1

New features:

  • [akida] add support for multi-pass hardware programs
  • [akida] add support for dense inputs in hardware
  • [akida] NSoC floor power
  • [akida] New batch size inference parameter (implicit pipelining)
  • [cnn2snn] StdPerAxisQuantizer

Removed features:

  • [akida] NSoC_v1 is no longer supported
  • [akida] python 3.6 is no longer supported
  • [akida] hardware 8-bit convolutions cannot return potentials
  • [akida] pipelining removed from python API
  • [akida] DMA clock-count removed from python API
  • [akida] Max-pooling size 3 is not supported on NSoC_v2
  • [cnn2snn] only single-branch models can be quantized and converted
  • [cnn2snn] Trainable quantizer is removed
  • [cnn2snn] BatchNorm folding only possible with per-axis quantizer

API changes:

  • [akida] NP mapping available through layer, including NP type
  • [akida] NSoC features (clock, power) available through device.soc
  • [akida] New PowerMeter API
  • [akida] New power statistics
  • [akida] Metrics now exposed by Sequence, and simplified
  • [engine] Simplified HardwareDriver API

Bug Fixes:

  • [akida] Calling fit without having compiled doesn't core dump
  • [akida] Mapping failed for some SepConv layer sizes
  • [akida] Minimum input size for valid convolutions is kernel_size + 1
  • [akida] NSoC power retrieval over I2C more reliable
  • [akida] Wrong memory usage when using dense inputs

Update akida_models to 1.1.0

  • updated CNN2SNN minimal required version to 2.1.1
  • python 3.6 is no longer supported
  • introduced AkidaNet NSoC v2 optimized architecture trained on ImageNet
  • introduced AkidaNet edge architecture trained on ImageNet
  • rebased MobileNet based models on AkidaNet backbone: cats vs. dogs, imagette, YOLO VOC, YOLO WiderFace, face identification and verification
  • rebased models obtained from VGG11/ImageNet transfer learning on AkidaNet/ImageNet transfer learning: melanoma, ODIR-5K, retinal OCT and ECG classification
  • updated ConvTiny/CWRU model head for NSoC v2 efficiency
  • updated ConvTiny/DVS_Gesture model for NSoC v2
  • introduced an AkidaNet/PlantVillage model obtained from AkidaNet/ImageNet transfer learning
  • cats vs. dogs transfer learning pipeline revisited

Documentation update

  • added missing elements to engine API
  • removed AEE references
  • updated overview block diagrams
  • updated hardware limitation page for production NSoC
  • fixed broken link for VGG/CIFAR10 in zoo performances page
  • updated zoo performances page with AkidaNet numbers
  • rebased MobileNet/ImageNet inference, YOLO/VOC detection and cats vs. dogs transfer learning examples on AkidaNet

Update akida and cnn2snn to version 2.0.5

API changes:

  • [akida] layer.mapping now replaces program.config(layer)
  • [akida] sequence.program is now directly a bytebuffer
  • [akida] LayerStatistics replaced by evaluate_sparsity

Bug fixes:

  • [akida] split filters on multiples of 8 on NSoC_v1
  • [akida] fix power measures toggling
  • [akida] avoid numpy warning in summary()
  • [akida] model.statistics fails if no power during inference

Documentation update

  • improved model hardware mapping documentation in Akida user guide
  • updated sparsity usage in examples
  • updated ContvTiny/CWRU model performances in zoo performances page

Update akida and cnn2snn to version 2.0.4

Bug fixes:

  • [akida] Fix out of memory for big inputs
  • [akida] Reset memory even after an exception
  • [akida] Perform output memory check only in Debug (performance)

Documentation update

  • updated mesh topology schematics

Update akida and cnn2snn to version 2.0.3

New features:

  • [akida] Power meter available for mini-PCIE boards

API changes:

  • [akida] added device.soc.power_meter
  • [akida] PowerMeter.flush() renamed to PowerMeter.events()

Bug fixes:

  • [akida] Power events are not flushed after every inference
  • [akida] Mesh mapper now correctly prioritizes FNP2

Documentation update

  • updated PowerMeter in AEE API