AIoTBench on Android

AIoTbenchmark program on Android; For different frameworks, models and data sets;

install

To install the app on an Android device and run, see install.md

Framework

PyTorch
Caffe2
Tensorflow lite

Model

MobileNet
ResNet
etc.

Data set

demo image
ImageNet

compile env

android studio: 3.6.1(most recent);
ndk: r21(most recent);

PyTorch

code from: PyTorch android demo

Caffe2

code from: pytorch android
other code: AICamera, only for old ndk(like r12b).
as pytorch android /README.md mentioned, this program need libcaffe2.a;
generate by: PyTorch
in dir: app/src/main/jniLibs/armeabi-v7a/, see app/.gitignore;

Tensorflow lite

code from: Tensorflow lite android

aiot dir

aiot
  models
    20200109-1
      models.json desc
    20200109-2
      models.json desc
    20200110-1
      models.json desc
  datasets
    20200109-1
      datasets.json desc

models.json

{
  "caffe2": [
    {
      "name": "inception_v3",
      "dir": "inception_v3",
      "init_net_pb": "init_net.pb",
      "predict_net_pb": "predict_net.pb",
      "norm_mean": [0,0,0],
      "norm_std_dev": [255.0,255.0,255.0],
      "labels": "synset2015.txt",
      "bitmap_rgb_type": "bgr",
      "bitmap_convert_size": [299,299]
    }
  ],
  "pytorch": [
    {
      "name": "inception_v3",
      "dir": "inception_v3",
      "net_pt": "inception_v3.pt",
      "bitmap_convert_size": [299,299]
    }
  ],
  "tflite": [
    {
      "name": "inception_v3",
      "dir": "inception_v3",
      "net_tflite": "inception_v3.tflite",
      "quantization": "float",
      "norm_mean": [127.5],
      "norm_std_dev": [127.5],
      "labels": "labels.txt",
      "bitmap_convert_method": "copy",
      "bitmap_rgb_type": "default",
      "bitmap_convert_size": [299,299]
    }
  ]
}

datasets.json

{
  "imagenet": {
    "classes_info": "ILSVRC2012_img_val_classes_with_name",
    "datasets": [
      {
        "name": "imagenet_2x2",
        "dir": "ILSVRC2012_img_val_sample/validation"
      }
    ]
  }
}

data set

ImageNet

image files

ILSVRC2012_img_val/validation
  eg:
    n01440764/
      eg
        ILSVRC2012_val_00000293.JPEG
        ILSVRC2012_val_00002138.JPEG
        ILSVRC2012_val_00003014.JPEG
      50 ILSVRC2012_val_xxxxxxxx.JPEG files;
    n01443537/
    n01484850/
  1000 nxxxxxxxx/ dirs;

class info

ILSVRC2012_img_val_classes_with_name
  eg:
    n01440764	tench, Tinca tinca
    n01443537	goldfish, Carassius auratus
    n01484850	great white shark, white shark, man-eater, man-eating
  1000 nxxxxxxxx classes;
  index of each class in this array is the classification output data of model;