Awesome Open Source
Awesome Open Source


A tensorflow implementation of Google's MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications

The official implementation is avaliable at tensorflow/model.

The official implementation of object detection is now released, see tensorflow/model/object_detection.


YellowFin optimizer has been intergrated, but I have no gpu resources to train on imagenet with it. Call for training ~_~

Official implement click here

Base Module

Accuracy on ImageNet-2012 Validation Set

Model Width Multiplier Preprocessing Accuracy-Top1 Accuracy-Top5
MobileNet 1.0 Same as Inception 66.51% 87.09%

Download the pretrained weight: OneDrive, BaiduYun


Time Benchmark

Environment: Ubuntu 16.04 LTS, Xeon E3-1231 v3, 4 Cores @ 3.40GHz, GTX1060.

TF 1.0.1(native pip install), TF 1.1.0(build from source, optimization flag '-mavx2')

Device Forward Forward-Backward Instruction set Quantized Fused-BN Remark
CPU 52ms 503ms - - - TF 1.0.1
CPU 44ms 177ms - - On TF 1.0.1
CPU 31ms - - 8-bits - TF 1.0.1
CPU 26ms 75ms AVX2 - - TF 1.1.0
CPU 128ms - AVX2 8-bits - TF 1.1.0
CPU 19ms 89ms AVX2 - On TF 1.1.0
GPU 3ms 16ms - - - TF 1.0.1, CUDA8.0, CUDNN5.1
GPU 3ms 15ms - - On TF 1.0.1, CUDA8.0, CUDNN5.1

Image Size: (224, 224, 3), Batch Size: 1


Train on Imagenet

  1. Prepare imagenet data. Please refer to Google's tutorial for training inception.

  2. Modify './script/' according to your environment.

bash ./script/

Benchmark speed

python ./scripts/

Train MobileNet Detector (Debugging)

  1. Prepare KITTI data.

After download KITTI data, you need to split it data into train/val set.

cd /path/to/kitti_root
mkdir ImageSets
cd ./ImageSets
ls ../training/image_2/ | grep ".png" | sed s/.png// > trainval.txt
python ./tools/

kitti_root floder then look like below

                  |     |-> image_2/00****.png
                  |     L-> label_2/00****.txt
                  |     L-> image_2/00****.png
                        |-> trainval.txt
                        |-> train.txt
                        L-> val.txt

Then convert it into tfrecord.

python ./tools/
  1. Mobify './script/' according to your environment.
bash ./script/

The code of this subject is largely based on SqueezeDet & SSD-Tensorflow. I would appreciated if you could feed back any bug.

Trouble Shooting

  1. About the MobileNet model size

According to the paper, MobileNet has 3.3 Million Parameters, which does not vary based on the input resolution. It means that the number of final model parameters should be larger than 3.3 Million, because of the fc layer.

When using RMSprop training strategy, the checkpoint file size should be almost 3 times as large as the model size, because of some auxiliary parameters used in RMSprop. You can use the to figure it out.

  1. Slim multi-gpu performance problems

#1390 #1428


  • [x] Train on Imagenet
  • [x] Add Width Multiplier Hyperparameters
  • [x] Report training result
  • [ ] Intergrate into object detection task(in progress)


MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications



Get A Weekly Email With Trending Projects For These Topics
No Spam. Unsubscribe easily at any time.
Tensorflow (12,571
Detection (1,181
Slim (356
Related Projects