xiamenwcy / extended-caffe

An extended caffe

Extended caffe

Introduction

This repository contains an extended caffe wich is modified from caffe version of yjxiong and introduces many new features.

Features

• on-the-fly data augmentation, which is used in ImageSegData layer, including mirror, crop, scale, smooth filer, rotation, translation, please refers to caffe\src\caffe\data_transformer\TransformImgAndSeg2 an example is as follows:

layer {
  name: "data"
  type: "ImageSegData"
  top: "data"
  top: "label"
  top: "data_dim"
  include {
    phase: TRAIN
  }
  transform_param {
    mirror: true
    crop_size: 352
    mean_value: 104.008
    mean_value: 116.669
    mean_value: 122.675
    scale_factors: 0.5
    scale_factors: 0.75
    scale_factors: 1
    scale_factors: 1.25
    scale_factors: 1.5
    scale_factors: 1.75
    scale_factors: 2.0
	smooth_filtering: true
	max_smooth: 6
	apply_probability: 0.5
	max_rotation_angle: 60
	max_translation: 30
  }
  image_data_param {
    root_folder: "/data1/caiyong.wang/data/Point/CASIA/"
    source: "/data1/caiyong.wang/data/Point/CASIA/list/train_edge.txt"
    batch_size: 1
    shuffle: true
    label_type: PIXEL
  }
}

• include interp_layer used in deeplab see http://liangchiehchen.com/projects/DeepLab.html
• include balance_cross_entropy_loss_layer used in hed see Holistically-Nested Edge Detection
• include normalize_layer: L2 normalization for parsenet see ParseNet: Looking Wider to See Better
• support pooling with bin size, output_size for pspnet,segnet
• support upsample_layer used in segnet

layer {
  name: "pool4"
  type: "Pooling"
  bottom: "conv4_3"
  top: "pool4"
  top: "pool4_idx"
  top: "pool4_size"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
	output_size: true
  }
}
layer {
  name: "upsample4"
  type: "Upsample"
  bottom: "conv5_1_D"
  bottom: "pool4_idx"
  bottom: "pool4_size"
  top: "pool4_D"
}

• include dice_loss_layer
• include focal_sigmoid_loss_layer, the usage is simlar with SigmoidCrossEntropyLoss

layer {
  name: "loss_mask"
  type: "FocalSigmoidLoss"
  bottom: "mask_pred"
  bottom: "mask_label"
  top: "loss_mask"
  loss_weight: 10
  loss_param {
       ignore_label: 255
	   normalize: true
  }
  focal_sigmoid_loss_param
  {
	  alpha: 0.95
      gamma: 2
  }
}

• include focal_softmax_loss_layer, modified from https://github.com/chuanqi305/FocalLoss, the usage is similar with SoftmaxWithLoss, more details, please see Focal Loss for Dense Object Detection
• include prelu_layer
• include smooth_L1_loss_layer
• include selu_layer
• support deconvolution upsampling type: nearest

layer {
  name: "out_2_up4"
  type: "Deconvolution"
  bottom: "out_2"
  top: "out_2_up4"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  convolution_param {
    num_output: 2
    bias_term: false
    pad: 0
    kernel_size: 4
    group: 2
    stride: 4
    weight_filler {
      type: "nearest"
    }
  }
}

• include my_spp_layer for spatial pyramid pooling see SPPNet：Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition

layer {
  name: "spatial_pyramid_pooling"
  type: "MySPP"
  bottom: "conv5"
  top: "pool5"
  my_spp_param {
    pool: MAX
    bin_size: 2  
	bin_size: 3 
    bin_size: 6 
  }
} 

Installation

For installation, please follow the instructions of Caffe. For chinese users, please refers to https://blog.csdn.net/xiamentingtao/article/details/78283336, https://blog.csdn.net/xiamentingtao/article/details/78266153, https://wangcaiyong.blog.csdn.net/article/details/110262549, and https://wangcaiyong.blog.csdn.net/article/details/125335017【<---HIGHLY RECOMMENDED】.

To enable cuDNN for GPU acceleration, cuDNN v6 is needed.

The code has been tested successfully on CentOS 6.9 with CUDA 8.0.

Questions

Please contact [email protected]

---

Following is the original README of Caffe.

Caffe

Caffe is a deep learning framework made with expression, speed, and modularity in mind. It is developed by the Berkeley Vision and Learning Center (BVLC) and community contributors.

Check out the project site for all the details like

• DIY Deep Learning for Vision with Caffe
• Tutorial Documentation
• BVLC reference models and the community model zoo
• Installation instructions

and step-by-step examples.

Please join the caffe-users group or gitter chat to ask questions and talk about methods and models. Framework development discussions and thorough bug reports are collected on Issues.

Happy brewing!

License and Citation

Caffe is released under the BSD 2-Clause license. The BVLC reference models are released for unrestricted use.

Please cite Caffe in your publications if it helps your research:

@article{jia2014caffe,
  Author = {Jia, Yangqing and Shelhamer, Evan and Donahue, Jeff and Karayev, Sergey and Long, Jonathan and Girshick, Ross and Guadarrama, Sergio and Darrell, Trevor},
  Journal = {arXiv preprint arXiv:1408.5093},
  Title = {Caffe: Convolutional Architecture for Fast Feature Embedding},
  Year = {2014}
}

extended-caffe's Issues

Implement this in windows ?

I have spent a whole week to implement this ML framework to windows, but i have failed using CMake, then VS to compile. Is there any prebuild vs project(sln file) to work with? If i have sln file, it is more easy to compile and use.
I did successfully compiled using happynear `s caffe version[https://github.com/happynear/caffe-windows], but it is lack of your parameters of TransformImgAndSeg2 and so on. I really need your caffe to finish my master project.