Akida 1.8.7
Overview
Installation
Requirements
Quick installation
Running examples
User guide
Getting started
For beginners
For users familiar with deep-learning
Akida user guide
The Akida Execution Engine
1. The Spiking Neural Network model
2. Input data format
3. Determine training mode
4. Interpreting outputs
Neural Network model
Specifying the Neural Network model
Saving and loading
Input layer types
Data-Processing layer types
Using Akida Unsupervised Learning
Learning constraints
Compiling a layer
Learning parameters
CNN2SNN toolkit
Overview
Conversion workflow
Typical training scenario
Design compatibility constraints
Quantization compatibility constraints
Command-line interface
Layers Considerations
Supported layer types
CNN2SNN Quantization-aware layers
Training-Only Layers
First Layers
Final Layers
Tips and Tricks
Akida models zoo
Overview
Command-line interface for model creation
Command-line interface for model training
CIFAR10 training and tuning
Layer Blocks
conv_block
dense_block
separable_conv_block
Hardware constraints
Input layer
Data-Processing layers
Convolutional layer
Fully connected layer
API reference
Akida Execution Engine
Model
Layer
LayerStatistics
Observer
InputData
InputConvolutional
FullyConnected
Convolutional
SeparableConvolutional
Dense
Sparse
coords_to_sparse
dense_to_sparse
packetize
Backend
ConvolutionMode
PoolingType
LearningType
Compatibility
CNN2SNN
quantize
convert
check_model_compatibility
WeightQuantizer
TrainableWeightQuantizer
MaxQuantizer
MaxPerAxisQuantizer
WeightFloat
QuantizedConv2D
QuantizedDepthwiseConv2D
QuantizedDense
QuantizedSeparableConv2D
ActivationDiscreteRelu
QuantizedReLU
Akida models
Quantization blocks
conv_block
separable_conv_block
dense_block
Layer blocks
conv_block
separable_conv_block
dense_block
Model zoo
Mobilenet
DS-CNN
VGG
YOLO
Examples
GXNOR/MNIST inference
1. Loading the MNIST dataset
2. Look at some images from the test dataset
3. Load the pre-trained Akida model
4. Classify a single image
5. Check performance across a number of samples
Regression tutorial
1. Load dependencies
2. Load the dataset
3. Create a Keras model satisfying Akida NSoC requirements
4. Check performance
5. Conversion to Akida
5.1 Convert the trained Keras model to Akida
5.2 Check Akida model accuracy
6. Estimate age on a single image
CNN conversion flow tutorial
1. Load and reshape MNIST dataset
2. Model definition
3. Model training
4. Model quantization
5. Model fine tuning (quantization-aware training)
6. Model conversion
DS-CNN/KWS inference
1. Load CNN2SNN tool dependencies
2. Load the preprocessed dataset
3. Create a Keras model satisfying Akida NSoC requirements
4. Check performance
5. Conversion to Akida
5.1 Convert the trained Keras model to Akida
5.2 Check prediction accuracy
5.3 Confusion matrix
DS-CNN CIFAR10 inference
1. Dataset preparation
2. Create a Keras DS-CNN model
3. Quantized model
4. Pretrained quantized model
5. Conversion to Akida
5.1 Convert to Akida model
5.2 Check hardware compliancy
5.3 Check performance
5.4 Show predictions for a random image
Transfer learning with MobileNet for cats vs. dogs
1. Transfer learning process
2. Load and preprocess data
2.A - Load and split data
2.B - Preprocess the test set
2.C - Get labels
3. Convert a quantized Keras model to Akida
3.A - Instantiate a Keras base model
3.B - Modify the network and load pre-trained weights
3.C - Convert to Akida
4. Classify test images
4.A Classify test images
4.B Compare results
MobileNet/ImageNet inference
1. Load CNN2SNN tool dependencies
2. Load test images from ImageNet
2.1 Load test images and preprocess test images
2.2 Load labels
3. Create a quantized Keras model
3.1 Instantiate Keras model
3.2 Check performance of the Keras model
4. Convert Keras model for Akida NSoC
4.1 Convert Keras model to an Akida compatible model
4.2 Test performance of the Akida model
4.3 Show predictions for a random test image
Changelog
Support
License
Akida Examples
Docs
»
User guide
User guide
ΒΆ
Getting started
For beginners
For users familiar with deep-learning
Akida user guide
The Akida Execution Engine
1. The Spiking Neural Network model
2. Input data format
Data format
Input events buffering
Data frames parallelization
3. Determine training mode
Akida Unsupervised learning
CNN to SNN conversion
4. Interpreting outputs
Neural Network model
Specifying the Neural Network model
Saving and loading
Input layer types
Data-Processing layer types
Activation parameters
Pooling parameters
Using Akida Unsupervised Learning
Learning constraints
Compiling a layer
Learning parameters
CNN2SNN toolkit
Overview
Conversion workflow
Typical training scenario
Design compatibility constraints
Quantization compatibility constraints
Command-line interface
Layers Considerations
Supported layer types
CNN2SNN Quantization-aware layers
Training-Only Layers
First Layers
Input Scaling
Final Layers
Tips and Tricks
Akida models zoo
Overview
Command-line interface for model creation
Command-line interface for model training
CIFAR10 training and tuning
Layer Blocks
conv_block
dense_block
separable_conv_block
Hardware constraints
Input layer
Data-Processing layers
Convolutional layer
Fully connected layer