- MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications (Howard et al., 2017)
- MobileNetV2: Inverted Residuals and Linear Bottlenecks (Sandler et al., 2018)
- Installation
- Install prerequisites (Debian-specific)
- Install CK workflows (universal)
- Benchmark MobileNets via TensorFlow Lite
- Benchmark MobileNets via TensorFlow (C++)
- Benchmark MobileNets via TensorFlow (Python)
- Common tools and libraries.
- Python, pip, SciPy, Collective Knowledge (CK).
- (Optional) Android SDK, Android NDK.
$ sudo apt install autoconf autogen libtool zlib1g-dev
$ sudo apt install gcc g++ git wget
$ sudo apt install libblas-dev liblapack-dev
$ sudo apt install python3 python3-pip
$ sudo python3 -m pip install scipy
$ sudo python3 -m pip install ck
NB: CK also supports Python 2.
You can optionally target Android API 23 (v6.0 "Marshmallow") devices using the
--target_os=android23-arm64 flag
(or similar), when using
the TensorFlow Lite benchmark (recommended) and TensorFlow (C++) benchmark (not recommended).
On Debian Linux, you can install the Android SDK and the Android NDK as follows:
$ sudo apt install android-sdk
$ sudo apt install google-android-ndk-installer
$ adb version
Android Debug Bridge version 1.0.36
Revision 1:7.0.0+r33-2
$ ck pull repo:ck-tensorflow
$ ck pull repo:ck-caffe --url=https://github.com/dividiti/ck-caffe
$ ck install package:imagenet-2012-val-min
NB: ImageNet dataset descriptions are contained in CK-Caffe for historic reasons.
NB: See program:image-classification-tflite for more details.
Install TFLite from source:
$ ck install package:lib-tflite-0.1.7-src-static [--target_os=android23-arm64]
You can also install TFLite from a prebuilt binary package for your target e.g.:
$ ck list package:lib-tflite-prebuilt*
lib-tflite-prebuilt-0.1.7-linux-aarch64
lib-tflite-prebuilt-0.1.7-linux-x64
lib-tflite-prebuilt-0.1.7-android-arm64
$ ck install package:lib-tflite-prebuilt-0.1.7-android-arm64 [--target_os=android23-arm64]
Select one of the 38 MobileNets models compatible with TFLite:
$ ck install package --tags=tensorflowmodel,mobilenet,tflite
$ ck compile program:image-classification-tflite [--target_os=android23-arm64]
Run the client (if required, connect an Android device to your host machine via USB):
$ ck run program:image-classification-tflite [--target_os=android23-arm64]
...
ILSVRC2012_val_00000001.JPEG - (65) n01751748 sea snake
0.42 - (65) n01751748 sea snake
0.20 - (54) n01729322 hognose snake, puff adder, sand viper
0.14 - (58) n01737021 water snake
0.06 - (62) n01744401 rock python, rock snake, Python sebae
0.03 - (60) n01740131 night snake, Hypsiglena torquata
---------------------------------------
Summary:
-------------------------------
Graph loaded in 0.001319s
All images loaded in 0.007423s
All images classified in 0.202271s
Average classification time: 0.202271s
Accuracy top 1: 1.0 (1 of 1)
Accuracy top 5: 1.0 (1 of 1)
--------------------------------
NB: See program:image-classification-tf-cpp for more details.
Install TensorFlow (C++) from source:
$ ck install package:lib-tensorflow-1.10.1-src-static [--target_os=android23-arm64]
Select one of the 38 MobileNets models compatible with TensorFlow (C++):
$ ck install package --tags=tensorflowmodel,mobilenet,frozen
$ ck compile program:image-classification-tf-cpp [--target_os=android23-arm64]
$ ck run program:image-classification-tf-cpp [--target_os=android23-arm64]
...
*** Dependency 3 = weights (TensorFlow model and weights):
...
Resolved. CK environment UID = b4fab4037b14a0b9 (version 2_1.4_224)
...
--------------------------------
Process results in predictions
---------------------------------------
ILSVRC2012_val_00000001.JPEG - (65) n01751748 sea snake
0.17 - (62) n01744401 rock python, rock snake, Python sebae
0.17 - (54) n01729322 hognose snake, puff adder, sand viper
0.10 - (58) n01737021 water snake
0.06 - (60) n01740131 night snake, Hypsiglena torquata
0.04 - (63) n01748264 Indian cobra, Naja naja
---------------------------------------
Summary:
-------------------------------
Graph loaded in 0.108859s
All images loaded in 0.005605s
All images classified in 0.481788s
Average classification time: 0.481788s
Accuracy top 1: 0.0 (0 of 1)
Accuracy top 5: 0.0 (0 of 1)
--------------------------------
NB: See program:image-classification-tf-py for more details.
Install TensorFlow (Python) from an x86_64 binary package (requires system protobuf):
$ sudo python3 -m pip install -U protobuf
$ ck install package:lib-tensorflow-1.10.1-cpu
or from source:
$ ck install package:lib-tensorflow-1.10.1-src-cpu
Select one of the 54 MobileNets models compatible with TensorFlow (Python):
$ ck install package --tags=tensorflowmodel,mobilenet --no_tags=mobilenet-all
NB: This excludes "uber" packages which can be used to install all models in the sets v1-2018-02-22 (16 models), v1[-2018-06-14] (16 models) and v2 (22 models) in one go:
$ ck search package --tags=tensorflowmodel,mobilenet-all
ck-tensorflow:package:tensorflowmodel-mobilenet-v1-2018_02_22
ck-tensorflow:package:tensorflowmodel-mobilenet-v2
ck-tensorflow:package:tensorflowmodel-mobilenet-v1
$ ck run program:image-classification-tf-py
...
*** Dependency 4 = weights (TensorFlow-Python model and weights):
...
Resolved. CK environment UID = b4fab4037b14a0b9 (version 2_1.4_224)
...
--------------------------------
Process results in predictions
---------------------------------------
ILSVRC2012_val_00000001.JPEG - (65) n01751748 sea snake
0.38 - (65) n01751748 sea snake
0.19 - (54) n01729322 hognose snake, puff adder, sand viper
0.13 - (58) n01737021 water snake
0.12 - (62) n01744401 rock python, rock snake, Python sebae
0.03 - (60) n01740131 night snake, Hypsiglena torquata
---------------------------------------
Summary:
-------------------------------
Graph loaded in 1.933857s
All images loaded in 0.002172s
All images classified in 0.359537s
Average classification time: 0.359537s
Accuracy top 1: 1.0 (1 of 1)
Accuracy top 5: 1.0 (1 of 1)
--------------------------------