Akida, 2nd Generation
Overview
Installation
Supported configurations
Quick installation
Running examples
User guide
Akida user guide
Overview
Programming interface
The Akida Model
Akida layers
Model Hardware Mapping
Devices
Model mapping
Advanced Mapping Details and Hardware Devices Usage
Performance measurement
Using Akida Edge learning
Learning constraints
Compiling a layer
QuantizeML toolkit
Overview
The FixedPoint representation
Quantization flow
Compatibility constraints
Model loading
Command line interface
quantize CLI
config CLI
check CLI
insert_rescaling CLI
Supported layer types
Keras support
ONNX support
CNN2SNN toolkit
Overview
Conversion flow
Conversion compatibility
Command-line interface
Handling Akida 1.0 and Akida 2.0 specificities
Legacy quantization API
Typical quantization scenario
Design compatibility constraints
Command-line interface
Layers Considerations
Tips and Tricks
Akida models zoo
Overview
Command-line interface for model creation
Command-line interface for model training
KWS training
AkidaNet training
Command-line interface for model evaluation
Command-line interface to evaluate model MACS
Layer Blocks
Handling Akida 1.0 and Akida 2.0 specificities
Akida Engine
Overview
Engine directory structure
Engine API overview
HardwareDriver
HardwareDevice
Dense
Shape
HwVersion
Sparse and Input conversion functions
Other headers in the API
API reference
Akida runtime
Model
Layer
Layer
Mapping
Akida layers
Akida V1 layers
Akida V2 layers
Layer parameters
LayerType
Padding
PoolType
Optimizers
Sequence
Sequence
BackendType
Pass
Device
Device
HwVersion
HWDevice
HWDevice
SocDriver
ClockMode
PowerMeter
NP
Tools
Sparsity
CNN2SNN
Akida version
Conversion
Legacy quantization API
Utils
Calibration
Transforms
Constraint
Quantization
Quantizers
Quantized layers
QuantizeML
Layers
Reshaping
Activations
Attention
Normalization
Convolution
Depthwise convolution
Separable convolution
Dense
Skip connection
Pooling
Shiftmax
Transformers
Rescaling
Dropout
Quantizers
Calibration
Recording
Models
Transforms
Quantization
Quantization parameters
Calibration
Utils
Tensors
QTensor
FixedPoint
QFloat
ONNX support
Layers
Custom patterns
Akida models
Layer blocks
CNN blocks
Transformers blocks
Transposed blocks
Detection block
Helpers
Gamma constraint
Unfusing SeparableConvolutional
Extract samples
Knowledge distillation
MACS
Model I/O
Utils
Model zoo
AkidaNet
Mobilenet
DS-CNN
VGG
YOLO
PointNet++
GXNOR
CenterNet
AkidaUNet
Transformers
Examples
General examples
Global Akida workflow
1. Create and train
2. Quantize
3. Convert
4. GXNOR/MNIST
AkidaNet/ImageNet inference
1. Dataset preparation
2. Pretrained quantized model
3. Conversion to Akida
4. Hardware mapping and performance
DS-CNN/KWS inference
1. Load the preprocessed dataset
2. Load a pre-trained native Keras model
3. Load a pre-trained quantized Keras model
4. Conversion to Akida
5. Confusion matrix
Age estimation (regression) example
1. Load the UTKFace Dataset
2. Load a pre-trained native Keras model
3. Load a pre-trained quantized Keras model
4. Conversion to Akida
5. Estimate age on a single image
Transfer learning with AkidaNet for PlantVillage
Transfer learning process
1. Dataset preparation
2. Get a trained AkidaNet base model
3. Add a classification head to the model
4. Train for a few epochs
5. Quantize the model
6. Compute accuracy
YOLO/PASCAL-VOC detection tutorial
1. Introduction
2. Preprocessing tools
3. Model architecture
4. Training
5. Performance
6. Conversion to Akida
Segmentation tutorial
1. Load the dataset
2. Load a pre-trained native Keras model
3. Load a pre-trained quantized Keras model
4. Conversion to Akida
5. Segment a single image
Build Vision Transformers for Akida
1. Model selection
2. Model optimization for Akida hardware
3. Model Training
4. Model quantization
5. Conversion to Akida
6. Displaying results Attention Maps
PyTorch to Akida workflow
1. Create and train
2. Export
3. Quantize
4. Convert
Quantization
Advanced QuantizeML tutorial
1. Defining a quantization scheme
2. Calibration
Upgrading to Akida 2.0
1. Workflow differences
2. Models architecture differences
3. Using
AkidaVersion
Off-the-shelf models quantization
1. Workflow overview
2. Data preparation
3. Download and export
4. Quantize
Advanced ONNX models quantization
1. Get model and data
2. Quantize
3. Conversion
Edge examples
Akida vision edge learning
1. Dataset preparation
2. Prepare Akida model for learning
3. Edge learning with Akida
Akida edge learning for keyword spotting
1. Edge learning process
2. Dataset preparation
3. Prepare Akida model for learning
4. Learn with Akida using the training set
5. Edge learning
Tips to set Akida edge learning parameters
1. Akida learning parameters
2. Create Akida model
3. Estimate the required number of weights of the trainable layer
4. Estimate the number of neurons per class
[Deprecated] CNN2SNN tutorials
Advanced CNN2SNN tutorial
1. Design a CNN2SNN quantized model
2. Weight Quantizer Details
3. Understanding quantized activation
4. How to deal with too high scale factors
Model zoo performance
Akida 1.0 models
Image domain
Classification
Object detection
Regression
Face recognition
Audio domain
Keyword spotting
Point cloud
Classification
Akida 2.0 models
Image domain
Classification
Object detection
Regression
Face recognition
Segmentation
Audio domain
Keyword spotting
Classification
Point cloud
Classification
Changelog
Support
License
Akida Examples
User guide
User guide
Akida user guide
Overview
Programming interface
The Akida Model
Akida layers
Akida 1.0 layers
Akida 2.0 layers
Model Hardware Mapping
Devices
Discovering Hardware Devices
Virtual Devices
Model mapping
Advanced Mapping Details and Hardware Devices Usage
Performance measurement
Using Akida Edge learning
Learning constraints
Compiling a layer
QuantizeML toolkit
Overview
The FixedPoint representation
Quantization flow
Compatibility constraints
Model loading
Command line interface
quantize CLI
config CLI
check CLI
insert_rescaling CLI
Supported layer types
Keras support
ONNX support
CNN2SNN toolkit
Overview
Conversion flow
Conversion compatibility
Command-line interface
Deprecated CLI actions
Handling Akida 1.0 and Akida 2.0 specificities
Legacy quantization API
Typical quantization scenario
Design 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
KWS training
AkidaNet training
Command-line interface for model evaluation
Command-line interface to evaluate model MACS
Layer Blocks
Handling Akida 1.0 and Akida 2.0 specificities
Akida Engine
Overview
Engine directory structure
Engine API overview
HardwareDriver
HardwareDevice
Dense
Shape
HwVersion
Sparse and Input conversion functions
Other headers in the API