/tfmars

MarNASNets and CNN for sensor-based human activity recognition built in TensorFlow

Primary LanguagePythonMIT LicenseMIT

tfmars

tfmars is the TensorFlow's implementation of Mobile-aware Convolutional Neural Network for Sensor-based Human Activity Recognition, a sibling of tfgarden. In this repository, some CNN models implemented in tfgarden have been implemented as Attention insertable models. Also, MarNASNets has been implemented.

MARS means Mobile-aware Activity Recognition modelS.

Models

  • Simple CNN: used on the paper by Li et al.
  • VGG16
  • Inception v3
  • ResNet 18
  • PyramidNet 18
  • Xception
  • DenseNet 121
  • MobileNet
  • MobileNetV2
  • MobileNetV3 Small
  • NASNet Mobile
  • MnasNet
  • EfficientNet B0
  • EfficientNet lite0

MarNASNets

MarNASNets are the CNN architectures designed by using Bayesian-optimization Neural Architecture Search via Keras Tuner. MarNASNets are mobile-aware models that achieves higher accuracy with fewer parameters than existing models. There are variations with different search spaces (A - E).

Install

pip install git+https://github.com/Shakshi3104/tfmars.git

Dependency

  • tensorflow >= 2.4.1

Performance

Model Accuracy [%] 1 Size [MB] 2 MFLOPs Latency [ms] 3 CPU load 3
Simple CNN 87.71 5.31 9.22 4.37 1.59
VGG16 89.54 154.00 357.13 5.64 1.83
Inception-v3 91.85 57.22 287.16 3.69 1.51
ResNet 18 90.53 15.41 173.72 2.67 1.50
PyramidNet 18 91.48 1.63 19.49 2.12 1.65
Xception 92.31 82.69 613.98 4.09 2.10
DenseNet 121 92.55 22.31 192.97 2.84 2.11
MobileNet 91.22 23.96 155.47 2.83 1.88
MobileNetV2 90.62 26.91 147.96 2.96 1.61
MobileNetV3 Small 91.45 11.60 35.19 2.48 1.42
NASNet Mobile 86.49 16.55 147.23 3.23 2.65
MnasNet 89.75 37.44 179.77 3.12 1.66
EfficientNet B0 92.50 45.70 221.68 3.32 1.59
EfficientNet lite0 91.52 43.11 220.17 3.21 1.89
MarNASNet-A 91.68 1.31 43.29 2.30 1.68
MarNASNet-B 91.79 0.42 4.79 2.21 1.47
MarNASNet-C 92.60 3.08 46.20 2.22 1.83
MarNASNet-D 91.87 1.25 19.83 2.25 1.86
MarNASNet-E 91.70 8.16 166.26 2.86 1.46

Citation

Under construction...

Footnotes

  1. Verifying accuracy with HASC-PAC2016 (HASC). ↩

  2. Size of MLModel file. ↩

  3. Testing conducted using iPhone 12 mini with iOS 15.2. Activitybench 3.9.4 tested with MLComputeUnits=all. Performance tests are conducted using specific computer systems and reflect the approximate performance of iPhone 12 mini. ↩ ↩2