Fully convolutional reduced VGGNet

readme.md

name	caffemodel	caffemodel_url	sha1	gist_id
Fully convolutional reduced VGGNet	VGG_ILSVRC_16_layers_fc_reduced.caffemodel	http://cs.unc.edu/~wliu/projects/ParseNet/VGG_ILSVRC_16_layers_fc_reduced.caffemodel	97eb7c469c5097f51a0f9a944f4a5731f470eee2

This is a model used in the paper

ParseNet: Looking Wider to See Better
Wei Liu, Andrew Rabinovich, Alexander C. Berg
arXiv:1506.04579

This is a network modified from VGGNet by making it fully convolutional and also by subsampling parameters from fc6 and fc7 layers. This is useful when using it to finetune for segmentation. For example, ParseNet shows how to use it to finetune for semantic segmentation task.

VGG_ILSVRC_16_layers_fc_reduced_deploy.prototxt

name: "VGG_ILSVRC_16_layers_fc_reduced"
input: "data"
input_dim: 10
input_dim: 3
input_dim: 500
input_dim: 500
layers {
  bottom: "data"
  top: "conv1_1"
  name: "conv1_1"
  type: CONVOLUTION
  convolution_param {
    num_output: 64
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv1_1"
  top: "conv1_1"
  name: "relu1_1"
  type: RELU
}
layers {
  bottom: "conv1_1"
  top: "conv1_2"
  name: "conv1_2"
  type: CONVOLUTION
  convolution_param {
    num_output: 64
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv1_2"
  top: "conv1_2"
  name: "relu1_2"
  type: RELU
}
layers {
  bottom: "conv1_2"
  top: "pool1"
  name: "pool1"
  type: POOLING
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layers {
  bottom: "pool1"
  top: "conv2_1"
  name: "conv2_1"
  type: CONVOLUTION
  convolution_param {
    num_output: 128
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv2_1"
  top: "conv2_1"
  name: "relu2_1"
  type: RELU
}
layers {
  bottom: "conv2_1"
  top: "conv2_2"
  name: "conv2_2"
  type: CONVOLUTION
  convolution_param {
    num_output: 128
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv2_2"
  top: "conv2_2"
  name: "relu2_2"
  type: RELU
}
layers {
  bottom: "conv2_2"
  top: "pool2"
  name: "pool2"
  type: POOLING
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layers {
  bottom: "pool2"
  top: "conv3_1"
  name: "conv3_1"
  type: CONVOLUTION
  convolution_param {
    num_output: 256
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv3_1"
  top: "conv3_1"
  name: "relu3_1"
  type: RELU
}
layers {
  bottom: "conv3_1"
  top: "conv3_2"
  name: "conv3_2"
  type: CONVOLUTION
  convolution_param {
    num_output: 256
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv3_2"
  top: "conv3_2"
  name: "relu3_2"
  type: RELU
}
layers {
  bottom: "conv3_2"
  top: "conv3_3"
  name: "conv3_3"
  type: CONVOLUTION
  convolution_param {
    num_output: 256
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv3_3"
  top: "conv3_3"
  name: "relu3_3"
  type: RELU
}
layers {
  bottom: "conv3_3"
  top: "pool3"
  name: "pool3"
  type: POOLING
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layers {
  bottom: "pool3"
  top: "conv4_1"
  name: "conv4_1"
  type: CONVOLUTION
  convolution_param {
    num_output: 512
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv4_1"
  top: "conv4_1"
  name: "relu4_1"
  type: RELU
}
layers {
  bottom: "conv4_1"
  top: "conv4_2"
  name: "conv4_2"
  type: CONVOLUTION
  convolution_param {
    num_output: 512
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv4_2"
  top: "conv4_2"
  name: "relu4_2"
  type: RELU
}
layers {
  bottom: "conv4_2"
  top: "conv4_3"
  name: "conv4_3"
  type: CONVOLUTION
  convolution_param {
    num_output: 512
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv4_3"
  top: "conv4_3"
  name: "relu4_3"
  type: RELU
}
layers {
  bottom: "conv4_3"
  top: "pool4"
  name: "pool4"
  type: POOLING
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layers {
  bottom: "pool4"
  top: "conv5_1"
  name: "conv5_1"
  type: CONVOLUTION
  convolution_param {
    num_output: 512
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv5_1"
  top: "conv5_1"
  name: "relu5_1"
  type: RELU
}
layers {
  bottom: "conv5_1"
  top: "conv5_2"
  name: "conv5_2"
  type: CONVOLUTION
  convolution_param {
    num_output: 512
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv5_2"
  top: "conv5_2"
  name: "relu5_2"
  type: RELU
}
layers {
  bottom: "conv5_2"
  top: "conv5_3"
  name: "conv5_3"
  type: CONVOLUTION
  convolution_param {
    num_output: 512
    pad: 1
    kernel_size: 3
  }
}
layers {
  bottom: "conv5_3"
  top: "conv5_3"
  name: "relu5_3"
  type: RELU
}
layers {
  bottom: "conv5_3"
  top: "pool5"
  name: "pool5"
  type: POOLING
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layers {
  bottom: "pool5"
  top: "fc6"
  name: "fc6"
  type: CONVOLUTION
  convolution_param {
    num_output: 1024
    kernel_size: 3
    dilation: 3
    pad: 3
  }
}
layers {
  bottom: "fc6"
  top: "fc6"
  name: "relu6"
  type: RELU
}
layers {
  bottom: "fc6"
  top: "fc6"
  name: "drop6"
  type: DROPOUT
  dropout_param {
    dropout_ratio: 0.5
  }
}
layers {
  bottom: "fc6"
  top: "fc7"
  name: "fc7"
  type: CONVOLUTION
  convolution_param {
    num_output: 1024
    kernel_size: 1
  }
}
layers {
  bottom: "fc7"
  top: "fc7"
  name: "relu7"
  type: RELU
}
layers {
  bottom: "fc7"
  top: "fc7"
  name: "drop7"
  type: DROPOUT
  dropout_param {
    dropout_ratio: 0.5
  }
}
layers {
  bottom: "fc7"
  top: "fc8"
  name: "fc8"
  type: CONVOLUTION
  convolution_param {
    num_output: 1000
    kernel_size: 1
  }
}
layers {
  bottom: "fc8"
  top: "prob"
  name: "prob"
  type: SOFTMAX
}

This is to make VGG faster and also implements the "atrous" algorithm described in DeepLab paper. Here is how the reduced VGG is configured for that: https://gist.github.com/weiliu89/45e9e8de2c13af6476ca#file-vgg_voc2012ext-prototxt-L385

I missed the DeepLab paper -- sorry! ("atrous" algoritm is great!!)

@weiliu89
How can I train this VGG_ILSVRC_16_layers_fc_reduced model?
Any examples or source codes?
Thank you！

Could you provide SSD train.prototxt, deploy.prototxt and solver.txt of InceptionV3?
Thank you.

why the dilation = 3 in fc6?

@shesung, dilation=3, then you can get a true convolutional kernel with size dilation_(kernel_size-1)+1=3_(3-1)+1=7, which is the same with the kernel_size of fc6 in the original VGGnet.

Can i treat dilation equal to the filter_stride in your code of caffe-fcn?

@tron19920125, dilation is not filter_stride.dilation is the hole between your single convolution unit

@weiliu89,What dataset do you use to train ,then get this VGG_ILSVRC_16_layers_fc_reduced.model,because I don't see any data illustrations from VGG_ILSVRC_16_layers_fc_reduced_deploy.prototxt, thank you!

@mjssat7 , using ILSVRC dataset which has 1000 classes.

@weiliu89, I'm trying to understand how do I parse the network output?
I'm using python and get dict with the prob inside. there is a (10,1000,16,16) array.
please any clue.. :)

We fine-tune the basic reduced VGG on ImageNet-1000 dataset or VOC dataset? If ImageNet-1000, Could you report the result of the reduced VGG on ImageNet?

How to create custon dataSet and mark up in on caffe?

@idanusher I guess an average global pooling is needed for the final probability. For 16 x 16 feature maps, avg. global pooling will average all entries in the feature map and return one single value. You can check this in the Caffe Pooling layer.

I can't download the .caffemodel.

I still had a copy of the caffemodel running around. I uploaded it to
https://owncloud.gwdg.de/index.php/s/SjXmX0Uqh1zaYgI/download
checksum should match the one in the table above:
sha1: 97eb7c469c5097f51a0f9a944f4a5731f470eee2
md5: f544332b79d78c838978ce2782b0c196
size: 83 MB

I can't download the .caffemodel.
The URL given in the previous version of this document seems to work, too:

http://vision.cs.unc.edu/wliu/projects/ParseNet/VGG_ILSVRC_16_layers_fc_reduced.caffemodel