MobileNet-Caffe

Introduction

This is a Caffe implementation of Google's MobileNets (v1 and v2). For details, please read the following papers:

[v1] MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications
[v2] Inverted Residuals and Linear Bottlenecks: Mobile Networks for Classification, Detection and Segmentation

Pretrained Models on ImageNet

We provide pretrained MobileNet models on ImageNet, which achieve slightly better accuracy rates than the original ones reported in the paper.

The top-1/5 accuracy rates by using single center crop (crop size: 224x224, image size: 256xN):

Network	Top-1	Top-5	sha256sum	Architecture
MobileNet v1	70.81	89.85	8d6edcd3 (16.2 MB)	netscope, netron
MobileNet v2	71.90	90.49	a3124ce7 (13.5 MB)	netscope, netron

Evaluate Models with a single image

Evaluate MobileNet v1:

python eval_image.py --proto mobilenet_deploy.prototxt --model mobilenet.caffemodel --image ./cat.jpg

Expected Outputs:

0.42 - 'n02123159 tiger cat'
0.08 - 'n02119022 red fox, Vulpes vulpes'
0.07 - 'n02119789 kit fox, Vulpes macrotis'
0.06 - 'n02113023 Pembroke, Pembroke Welsh corgi'
0.06 - 'n02123045 tabby, tabby cat'

Evaluate MobileNet v2:

python eval_image.py --proto mobilenet_v2_deploy.prototxt --model mobilenet_v2.caffemodel --image ./cat.jpg

Expected Outputs:

0.26 - 'n02123159 tiger cat'
0.22 - 'n02124075 Egyptian cat'
0.15 - 'n02123045 tabby, tabby cat'
0.04 - 'n02119022 red fox, Vulpes vulpes'
0.02 - 'n02326432 hare'

Finetuning on your own data

Modify deploy.prototxt and save it as your train.prototxt as follows: Remove the first 5 input/input_dim lines, and add Image Data layer in the beginning like this:

layer {
  name: "data"
  type: "ImageData"
  top: "data"
  top: "label"
  include {
    phase: TRAIN
  }
  transform_param {
    scale: 0.017
    mirror: true
    crop_size: 224
    mean_value: [103.94, 116.78, 123.68]
  }
  image_data_param {
    source: "your_list_train_txt"
    batch_size: 32 # your batch size
    new_height: 256
    new_width: 256
    root_folder: "your_path_to_training_data_folder"
  }
}

Remove the last prob layer, and add Loss and Accuracy layers in the end like this:

layer {
  name: "loss"
  type: "SoftmaxWithLoss"
  bottom: "fc7"
  bottom: "label"
  top: "loss"
}
layer {
  name: "top1/acc"
  type: "Accuracy"
  bottom: "fc7"
  bottom: "label"
  top: "top1/acc"
  include {
    phase: TEST
  }
}
layer {
  name: "top5/acc"
  type: "Accuracy"
  bottom: "fc7"
  bottom: "label"
  top: "top5/acc"
  include {
    phase: TEST
  }
  accuracy_param {
    top_k: 5
  }
}

Related Projects

MobileNet in this repo has been used in the following projects, we recommend you to take a look:

The MobileNet neural network using Apple's new CoreML framework hollance/MobileNet-CoreML
Mobile-deep-learning baidu/mobile-deep-learning
Receptive Field Block Net for Accurate and Fast Object Detection ruinmessi/RFBNet
Depthwise Convolutional Layer yonghenglh6/DepthwiseConvolution
MobileNet-MXNet KeyKy/mobilenet-mxnet
Caffe2-MobileNet camel007/caffe2-mobilenet

Updates (Feb. 5, 2018)

Add pretrained MobileNet v2 models (including deploy.prototxt and weights)
Hold pretrained weights in this repo
Add sha256sum code for pretrained weights
Add some code snippets for single image evaluation
Uncomment engine: CAFFE used in mobilenet_deploy.prototxt
Add params (lr_mult and decay_mult) for Scale layers of mobilenet_deploy.prototxt
Add prob layer for mobilenet_deploy.prototxt