/ML-zoo

Primary LanguagePythonApache License 2.0Apache-2.0

Model Zoo

version

A collection of machine learning models optimized for Arm IP.

Anomaly Detection

Network Type Framework Cortex-A Cortex-M Mali GPU Ethos U Score (AUC)
MicroNet Large INT8 INT8 TensorFlow Lite ✖️ ✔️ ✔️ ✔️ 0.968
MicroNet Medium INT8 INT8 TensorFlow Lite ✖️ ✔️ ✔️ ✔️ 0.963
MicroNet Small INT8 INT8 TensorFlow Lite ✖️ ✔️ ✔️ ✔️ 0.955

Dataset: Dcase 2020 Task 2 Slide Rail

Image Classification

Network Type Framework Cortex-A Cortex-M Mali GPU Ethos U Score (Top 1 Accuracy)
MobileNet v2 1.0 224 INT8 * INT8 TensorFlow Lite ✔️ ✔️ ✔️ ✔️ 0.697
MobileNet v2 1.0 224 UINT8 UINT8 TensorFlow Lite ✖️ ✖️ ✔️ ✔️ 0.708

Dataset: ILSVRC 2012

Keyword Spotting

Network Type Framework Cortex-A Cortex-M Mali GPU Ethos U Score (Accuracy)
CNN Large INT8 * INT8 TensorFlow Lite ✔️ ✔️ ✔️ ✔️ 0.923
CNN Medium INT8 * INT8 TensorFlow Lite ✔️ ✔️ ✔️ ✔️ 0.905
CNN Small INT8 * INT8 TensorFlow Lite ✔️ ✔️ ✔️ ✔️ 0.902
DNN Large INT8 * INT8 TensorFlow Lite ✔️ ✔️ ✔️ ✔️ 0.860
DNN Medium INT8 * INT8 TensorFlow Lite ✔️ ✔️ ✔️ ✔️ 0.839
DNN Small INT8 * INT8 TensorFlow Lite ✔️ ✔️ ✔️ ✔️ 0.821
DS-CNN Large Clustered FP32 * FP32 TensorFlow Lite ✔️ ✔️ ✔️ ✖️ 0.948
DS-CNN Large Clustered INT8 * INT8 TensorFlow Lite ✔️ ✔️ ✔️ ✔️ 0.939
DS-CNN Large INT8 * INT8 TensorFlow Lite ✔️ ✔️ HERO ✔️ ✔️ 0.945
DS-CNN Medium INT8 * INT8 TensorFlow Lite ✔️ ✔️ HERO ✔️ ✔️ 0.939
DS-CNN Small INT8 * INT8 TensorFlow Lite ✔️ ✔️ HERO ✔️ ✔️ 0.931
DS-CNN Small INT16 * INT16 TensorFlow Lite ✔️ ✔️ HERO ✔️ ✔️ 0.934
CNN Large FP32 * FP32 TensorFlow Lite ✔️ ✔️ ✔️ ✖️ 0.934
CNN Medium FP32 * FP32 TensorFlow Lite ✔️ ✔️ ✔️ ✖️ 0.918
CNN Small FP32 * FP32 TensorFlow Lite ✔️ ✔️ ✔️ ✖️ 0.922
DNN Large FP32 * FP32 TensorFlow Lite ✔️ ✔️ ✔️ ✖️ 0.867
DNN Medium FP32 * FP32 TensorFlow Lite ✔️ ✔️ ✔️ ✖️ 0.850
DNN Small FP32 * FP32 TensorFlow Lite ✔️ ✔️ ✔️ ✖️ 0.836
DS-CNN Large FP32 * FP32 TensorFlow Lite ✔️ ✔️ HERO ✔️ ✖️ 0.950
DS-CNN Medium FP32 * FP32 TensorFlow Lite ✔️ ✔️ HERO ✔️ ✖️ 0.943
DS-CNN Small FP32 * FP32 TensorFlow Lite ✔️ ✔️ HERO ✔️ ✖️ 0.939
MicroNet Large INT8 INT8 TensorFlow Lite ✖️ ✔️ ✔️ ✔️ 0.965
MicroNet Medium INT8 INT8 TensorFlow Lite ✖️ ✔️ ✔️ ✔️ 0.958
MicroNet Small INT8 INT8 TensorFlow Lite ✖️ ✔️ ✔️ ✔️ 0.953

Dataset: Google Speech Commands Test Set

Noise Suppression

Network Type Framework Cortex-A Cortex-M Mali GPU Ethos U Score (Average Pesq)
RNNoise INT8 * INT8 TensorFlow Lite ✔️ ✔️ ✔️ ✔️ 2.945

Dataset: Noisy Speech Database For Training Speech Enhancement Algorithms And Tts Models

Object Detection

Network Type Framework Cortex-A Cortex-M Mali GPU Ethos U Score (mAP)
SSD MobileNet v1 FP32 * FP32 TensorFlow Lite ✔️ ✖️ ✔️ ✖️ 0.210
SSD MobileNet v1 INT8 * INT8 TensorFlow Lite ✔️ ✖️ ✔️ ✖️ 0.234
SSD MobileNet v1 UINT8 * UINT8 TensorFlow Lite ✖️ ✖️ ✔️ ✖️ 0.180
YOLO v3 Tiny FP32 * FP32 TensorFlow Lite ✔️ ✖️ ✔️ ✖️ 0.331

Dataset: COCO Validation 2017

Speech Recognition

Network Type Framework Cortex-A Cortex-M Mali GPU Ethos U Score (LER)
Wav2letter INT8 INT8 TensorFlow Lite ✔️ ✔️ ✔️ ✔️ 0.0877
Wav2letter Pruned INT8 * INT8 TensorFlow Lite ✔️ ✔️ ✔️ ✔️ 0.0783
Tiny Wav2letter INT8 * INT8 TensorFlow Lite ✔️ ✔️ ✖️ ✔️ 0.0348
Tiny Wav2letter Pruned INT8 * INT8 TensorFlow Lite ✔️ ✔️ ✖️ ✔️ 0.0283

Dataset: LibriSpeech, Fluent Speech

Superresolution

Network Type Framework Cortex-A Cortex-M Mali GPU Ethos U Score (PSNR)
SESR INT8 ** INT8 TensorFlow Lite ✔️ ✖️ ✔️ HERO ✖️ 35.00dB

Dataset: DIV2K

Visual Wake Words

Network Type Framework Cortex-A Cortex-M Mali GPU Ethos U Score (Accuracy)
MicroNet VWW-2 INT8 INT8 TensorFlow Lite ✖️ ✔️ ✔️ ✔️ 0.768
MicroNet VWW-3 INT8 INT8 TensorFlow Lite ✖️ ✔️ ✔️ ✔️ 0.855
MicroNet VWW-4 INT8 INT8 TensorFlow Lite ✖️ ✔️ ✔️ ✔️ 0.822

Dataset: Visual Wake Words

Key

  • ✔️ - Will run on this platform.
  • ✖️ - Will not run on this platform.
  • * - Code to recreate model available.
  • ** - This model has a large memory footprint – it will not run on all platforms.

License

Apache-2.0 unless otherwise explicitly stated.