Q7-CapsNets

Capsule Networks (Dynamic Routing) on Arm Cortex-M and RISC-V GAP8

Features

• C API (based on CMSIS-NN) for the execution of quantized Capsule Networks on Arm Cortex-M MCUs
• C API (based on PULP-NN) for the execution of quantized Capsule Networks on RISC-V GAP8 MCUs
• Python framework for the quantization of Capsule Networks developed in TensorFlow
• Python scripts to simulate the behavior of the quantized CapsNet on MCUs (fast testing of the accuracy returned by the Q7-CapsNet)
• Implementation examples for the MNIST, CIFAR-10, and SmallNORB datasets

Tests and Results

Dataset	Float-32 CapsNet ACC	Q7 CapsNet ACC (Arm)	Q7 CapsNet ACC (RISC-V)
MNIST	99.01%	98.83%	98.85%
CIFAR-10	78.54%	78.50%	78.38%
SmallNORB	92.56%	92.50%	92.49%

Directory Structure

Directory	Content
Arm_Cortex-M/API	C API for Arm Cortex-M
Arm_Cortex-M/Examples	Examples implementing Q7-CapsNets on Arm Cortex-M
RISC-V_GAP8/API	C API for RISC-V GAP8
RISC-V_GAP8/pulp-nn	Fork of pulp-nn supporting conv2d with signed int8
RISC-V_GAP8/Examples	Examples implementing Q7-CapsNets on RISC-V GAP8
Quantization_Framework	Python framework to quantize a CapsNet developed in TensorFlow and saved as .h5
Simulator	Examples on how to fast testing the accuracy of Q7-CapsNets

Help and Support

Communication

• E-mail: miguel.costa@dei.uminho.pt

Citations

If you use Q7-CapsNets in a scientific publication, we would appreciate citations: https://arxiv.org/abs/2110.02911