Changelog

2.4.0 (24/01/2025)

Qualia-Core

New features

  • deployment.qualia_codegen.Linux: use CMake project.

  • Add NucleoH7S3L8 Qualia-CodeGen deployment target.

  • Add NucleoU575ZIQ Qualia-CodeGen deployment target.

  • evaluation.host.Qualia: collect all metrics.

  • evaluation.target.Qualia: get latency measurement from target.

  • Add CMSISMFCC DataAugmentation.

  • Add Cutout1D DataAugmentation.

  • Add EZBirds Dataset.

  • learningframework.PyTorch: implement save_graph_plot() to plot model topology to SVG file using torch.fx.

  • Add SlopeMetric PyTorch metric: linear regression slope.

  • learningmodel.pytorch.QuantizedCNN: add support for quantized GSP.

  • Add ResNetSampleNorm PyTorch model: ResNet with SampleNorm layer after input.

  • Add ResNetStride PyTorch model: ResNet with strides parameter for strided convolutions and pool_sizes for max-pooling.

  • Add SampleNorm and QuantizedSampleNorm PyTorch layers.

  • postprocessing.QualiaCodeGen: custom metrics instanciation with metrics param.

Bug fixes

  • Make local git repository and git dependency optional.

  • Make deploy.optimize param optional.

  • learningframework.Keras: save_graph_plot() outputs to out/learningmodel subdirectory instead of current directory.

  • learningframework.PyTorch: do not attempt to compute predictions if confusion matrix is disabled.

  • preprocessing, postprocessing: make Matplotlib optional when visualization modules are not used.

  • postprocessing.QuantizationAwareTraining: testacc metric is optional.

Breaking changes

  • learningmodel.pytorch.CNN: GSP makes last Conv layer implicit without BatchNorm.

Qualia-Plugin-SNN

New features

  • Add SpikingJellyMultiStepTimeStepsInData LearningFramework.

  • Add Add and QuantizedAdd layers with SpikingJelly support.

  • Add NucleoL452REP Qualia-CodeGen deployment target.

  • postprocessing.EnergyEstimationMetric: input/output spike count.

  • postprocessing.EnergyEstimationMetric: handle multi-step mode and hybrid layers.

  • postprocessing.EnergyEstimationMetric: add support for ResNet.

Bug fixes

  • postprocessing.EnergyEstimationMetric: take timsteps into account for biases.

  • postprocessing.EnergyEstimationMetric: take potential reads into account.

  • postprocessing.EnergyEstimationMetric: do not count reset and leak in case no IF activation.

  • postprocessing.EnergyEstimationMetric: compute Conv/FC as dense with MACs if input is non-binary.

  • postprocessing.EnergyEstimationMetric: use e_rdram() for e_rdpot_conv_snn().

Breaking changes

  • learningmodel.pytorch.SCNN: GSP makes last Conv layer implicit without BatchNorm.

Qualia-CodeGen-Core

New features

  • Add Concatenate layer for Keras.

  • Add Slice layer for Keras 2.x SlicingOpLambda and Keras 3.x GetItem.

  • Add Permute layer for torch.permute().

  • Add SampleNorm layer for Keras custom layer.

  • Add MetricsConverter: generate instanciation of C++ metrics.

  • assets/include/model.hh: add output quantization information.

  • examples/Linux: CMake project.

  • examples/Linux: use libqualia-neuralnetwork.

  • Add examples/NucleoH7S3L8.

  • Add examples/NucleoU575ZIQ.

  • examples/NucleoL452REP: measure latency on target.

  • libqualia-neuralnetwork/Metrics: add Accuracy, MAE, MSE, PCC and linear regression slope metrics in C++.

  • libqualia-neuralnetwork/NeuralNetwork: quantize targets to match outputs.

  • libqualia-neuralnetwork: install as Python package.

  • libqualia-neuralnetwork/NeuralNetwork: add metrics computation.

Bug fixes

  • graph.TorchModelGraph: fix AdaptiveAvgPool1d argument.

  • graph.TorchModelGraph: fix handling of method before any module.

  • graph.TorchModelGraph: filter args of methods.

  • graph.TorchModelGraph: fix input shape of methods.

  • graph.KerasModelGraph: fix input/output shape access for some Keras 3.x layers.

  • graph.KerasModelGraph: fix for Keras 3.x layers without dtype.

  • graph.KerasModelGraph: handle improperly ordered Keras graphs.

  • assets/model.cc: support multiple input layers.

  • assets/layers/conv1d.cc: fix usage of arm_convolve_HWC_q15_fast_nonsquare() due to undocumented constraint.

  • examples/Linux: disable -floop-nest-optimize, broken with libisl 0.27.

Breaking changes

  • learningmodel.pytorch.SCNN: GSP makes last Conv layer implicit without BatchNorm.

  • examples/NucleoL452REP: default to 80 MHz core clock, optional 48 MHz setting.

Qualia-CodeGen-Plugin-SNN

New features

  • assets/include/model.hh: add output quantization information.

  • examples/Linux: CMake project.

  • examples/Linux: use libqualia-neuralnetwork with SNN support.

  • Add examples/NucleoL452REP.

Bug fixes

  • graph.TorchModelGraph: fix handling of method before any module.

2.3.0 (15/07/2024)

Qualia-Core

New features

  • Bump maximum Python version to 3.12.

  • learningframework.PyTorch: custom metric name for checkpoints with checkpoint_metric param.

  • learningframework.PyTorch: optionally disable confusion matrix with enable_confusion_matrix param.

  • learningframework.PyTorch: custom loss with loss param.

  • learningframework.PyTorch: custom metric selection with metrics param.

  • Add BrainMIX dataset.

  • Add Amplitude DataAugmentation.

  • Add CopySet PreProcessing.

  • postprocessing.Keras2TFLite, postprocessing.RemoveKerasSoftmax, postprocessing.Torch2Keras: add support for Keras 3.x.

  • Add VisualizeFeatureMaps PostProcessing.

  • postprocessing.FuseBatchNorm: add evaluate param to optionally disable evaluation.

  • learningmodel.pytorch.Quantizer: add v tensor type for more flexibility with Qualia-Plugin-SNN.

Bug fixes

  • preprocessing.DatasetSplitterBySubjects: add dest set to available sets.

  • preprocessing.DatasetSplitter: update dest info set.

  • learningframework.PyTorch: fix default metric for checkpoint (valavgclsacc).

  • learningframework.PyTorch: fix seeding after several training (integer overflow).

  • learningframework.PyTorch: micro computation for prec, rec, and f1 metrics.

  • Root logger stdout stream can output DEBUG level so that child logger can log debug messages. Default level is still INFO.

  • learningmodel.pytorch.layers.QuantizedLayer: fix multiple inheritance Protocol on Python 3.10.

  • Fix parsing of [[parameter_research]] section in configuration file. Actual behaviour of parameter_research is still untested.

Breaking changes

  • Some metrics previously computed as macro (prec, rec, f1) are now computed as micro, results will be different on unbalanced datasets.

Qualia-Plugin-SNN

New features

  • Bump maximum Python version to 3.12.

  • postprocessing.EnergyEstimationMetric: print ModelGraph.

  • postprocessing.EnergyEstimationMetric: trace GlobalSumPooling (not counted in metric).

  • learningmodel.pytorch.layers.quantized_SNN_layers.py: membrane potential uses new tensor type v to enable quantizer configuration decoupled from activations.

Bug fixes

  • learningmodel.pytorch.SNN: inherit from SpikingJelly’s StepModule to avoid warning when using set_step_mode() on the network.

  • postprocessing.FuseBatchNorm: copy step_mode and inherit returned GraphModule from SpikingJelly’s StepModule.

  • pyproject.toml: require pytorch dependency group of qualia-core.

Qualia-CodeGen-Core

New features

  • Bump maximum Python version to 3.12.

  • graph.KerasModelGraph: Add support for Keras 3.x.

  • graph.KerasModelGraph: add support for padding='same' param of Conv layers.

Bug fixes

  • graph.TorchModelGraph: fix ouput shape for methods.`

  • graph.TorchModelGraph: keep first dim as-is when generating dummy input (propagate timestep dim with Qualia-CodeGen-Plugin-SNN).

Qualia-CodeGen-Plugin-SNN

New features

  • Bump maximum Python version to 3.12.

Bug fixes

  • graph.TorchModelGraph: reset network to make sure potential are not initialized with wrong shape.

2.2.0 (01/03/2024)

Qualia-Core

New features

  • Pass separate scale factor for bias to Qualia-CodeGen.

  • bench.use_test_as_valid configuration option to use the test dataset for validation metrics when no validation dataset is available.

  • Add QuantizationAwareFX PostProcessing module as an alternative to QuantizationAwareTraining using torch.fx to build the quantized model, replacing layers with their quantized alternatives when possible.

  • QuantizedLayer: add from_module() method to build a quantized layer from a non-quantized one with same configuration and weights.

  • TorchVisionModel LearningModel: allow using original classifier.

  • TorchVisionModel LearningModel: allow choosing the last layer for the feature extractor.

  • Add CIFAR-10 TorchVision’s MobileNetv2 configuration for float32 training and int16 quantization.

  • Add Normalize DataAugmentation using torchvision.transforms.Normalize.

  • Add TorchVisionModelTransforms DataAugmentation for use with TorchVisionModel to adapt input data.

  • PyTorch LearningFramework: show loss in progress bar.

  • Colored console logging to stderr for warnings and errors.

  • qualia_codegen.NucleoL452REP: use CMake project instead of STM32CubeIDE.

Bug fixes

  • Fix detection of seperate quantizer for bias.

  • Fix for no [[preprocessing]] section in configuration file.

  • Fix TorchVisionModel LearningModel construction.

  • qualia_codegen.Linux Deployer: try to fix overflow detection.

Breaking changes

  • activations_range.txt file: introduce bias_q column. Existing models will have to be re-quantized in order to be deployed using Qualia-CodeGen, this does not change the classification results if seperate quantizer for bias was not used.

  • Symmetric quantization uses the full range down to -2^{b-1} instead of restricting lower bound to -2^{b-1} - 1. Existing models will have to be re-quantized and this may slightly change the classification results.

Qualia-Plugin-SNN

Bug fixes

  • QuantizedIFNode/QuantizedLIFNode: return correct round mode for membrane potential and hyperparameters for Qualia-CodeGen.

Qualia-CodeGen-Core

New features

  • Add support for separate scale factor for bias.

  • Add support for 1D/2D grouped convolutions (incl. depthwise convolutions).

  • Add left shift overflow detection when TRAPV_SHIFT is defined and not using CMSIS-NN.

  • TorchModelGraph: add support for ReLU6 activation layer.

  • TorchModelGraph: add support for operator.add() function in addition to Add layer.

  • TorchModelGraph: add support for torch.nn.functional.adaptive_avg_pool2d function.

  • TorchModelGraph: add support for torch.flatten() function in addition to Flatten layer.

  • Add NucleoL452REP example project using CMake instead of STM32CubeIDE.

  • Add common libqualia-neuralnetwork for string to float decoding, quantization of inputs and classification task results, used only for NucleoL452REP CMake project for now.

  • Abort compilation in case of unsupported activation function for layer.

Bug fixes

  • TorchModelGraph: fix AdaptiveAvgPool1d layer.

  • Linux example: fix single.cpp build.

  • SparkFunEdge example: fix round-to-nearest support.

2.1.0 (16/01/2024)

Qualia-Core

New features

  • Deployment Qualia-CodeGen: Add support for nearest rounding mode (in addition to floor rounding mode) on Linux, NucleoL452REP (incl. CMSIS-NN), SparkFun Edge (incl. CMSIS-NN) and Longan Nano (incl. NMSIS-NN).

Bug fixes

  • Fix importing qualia_core packages after plugins initialization.

  • Fix some Python 3.9 compatibility issues.

  • LearningModel MLP/QuantizedMLP: Fix layers instanciation.

  • PostProcessing QuantizationAwareTraining: Use validation set instead of test set for validation with.

Other changes

  • Various refactor, cleanup and typing fixes (quantized layer inheritance, qualia_core.deployment, qualia_core.postprocessing.QualiaCodeGen, qualia_core.evaluation).

Breaking changes

  • activations_range.txt file: remove unused global_max columns and introduced round_mode columns. Existing models will have to be re-quantized in order to be deployed using Qualia-CodeGen, this does not change the classification results.

  • Nearest rounding mode for quantization with PyTorch now rounds upwards for half tie-breaker instead of round half to even in order to match Qualia-CodeGen. Existing models using nearest rounding mode will have to be re-quantized and this may slightly change the classification results.

Qualia-Plugin-SNN

Bug fixes

  • LearningModel SResNet/QuantizedSResNet Fix using 2D layers.

  • PostProcessing EnergyEstimationMetric: Fix display of message when non-binary inputs/outputs are included in the total.

Other changes

  • Some refactor for quantized layer inheritance.

Qualia-CodeGen-Core

New features

  • Add support for nearest rounding mode for weights and activations in all supported layers and for inputs in all included example targets, with and without CMSIS-NN/NMSIS-NN.

  • Generate a new include/defines.h file to be used as a global pre-processor macros definition file (used to set global rounding mode for CMSIS-NN/NMSIS-NN for now).

  • Add an exclude argument to Quantizer.quantize_weights() to exclude some weights type by name in a custom plugin.

Bug fixes

  • Longan Nano: build NMSIS-NN from upstream sources instead of using pre-built binary in order to apply custom optimizations and support floor rounding mode instead of nearest.

  • graph.layers: fix weights property inheritance

Qualia-CodeGen-Plugin-SNN

New features

  • Add support for nearest rounding mode for potentials in if and lif layers and for inputs in the included Linux target.

Bug fixes

  • graph.layers: fix weights property inheritance

2.0.0 (24/11/2023)

Initial release of Qualia

1.0.0 (15/09/2021)

Initial release of MicroAI