Deeplab v2 安装及调试全过程

上期为大家带来的是从FCN到DeepLab V2的一些相关知识，今天我们就来和大家分享一些DeepLab V2的安装及调试全过程，希望可以为一些需要的科研小伙伴带来一丝丝帮助，请继续欣赏下去。把Deeplabv2的 run_pascal.sh与run_densecrf.sh成功运行，现将调试过程整理如下：

首先，安装Caffe、Ubuntu 16.04+cuda8.0等环境应该不需要再次详细说了吧，如果有不清楚的小伙伴，进点击下面的链接，也是计算机视觉平台之前推送的，可以简单方便的进行安装。

链接：Caffe（含GPU）安装与测试

---

一、安装必要的依赖库

安装 matio：

安装方法1：  sudo apt-get install libmatio-dev  

安装方法2：  下载matio (https://sourceforge.net/projects/matio/files/matio/1.5.2/)

tar zxf matio-1.5.2.tar.gz  

cd matio-1.5.2  

./configure  

make  

make check  

make install  

sudo ldconfig

安装 wget  

sudo pip install wget 

如果出错，就按照下面的命令成功：  

pip install –upgrade pip –user  

pip install –upgrade setuptools –user  

sudo pip install wget

二、下载Deeplabv2并编译  

下载代码：

git clone https://github.com/xmojiao/deeplab_v2.git  

（试过许多Deeplab代码，这个最容易编译成功，所以我用的是这个代码编译的）  

对 caffe 进行编译： 

修改deeplab_v2/deeplab-public-ver2/路径下的Makefile.config.example文件，重命名为Makefile.config；

接着修改这个文件中的内容，将第四行的 “# USE_CUDNN := 1”的 # 去掉。如果需要，因为我用的pycaffe编译，所以不需要修改python的路径，保存退出。  

编译 caffe的命令：  

cd ~/Desktop/deeplab_v2/deeplab-public-ver2  

make all -j16  

如果出现下面的错误1：  

src/caffe/net.cpp:8:18: fatal error: hdf5.h: No such file or directory  compilation terminated.  

解决办法: 修改两个make文件（Makefile.config，Makefile）  

Makefile.config：

将

INCLUDE_DIRS := $(PYTHON_INCLUDE) /usr/local/include  

LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib

修改为：  

INCLUDE_DIRS := $(PYTHON_INCLUDE) /usr/local/include /usr/include/hdf5/serial

LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib /usr/lib/x86_64-linux-gnumake

Makefile：  

将  

LIBRARIES += glog gflags protobuf boost_system boost_filesystem m hdf5_hl hdf5  

修改为：  

LIBRARIES += glog gflags protobuf boost_system boost_filesystem m hdf5_serial_hl hdf5_serial matio 

重新编译：  

make all -j16

如果出现下面的错误2：  

./include/caffe/common.cuh(9): error: function “atomicAdd(double *, double)” has already been defined  

解决方法：

打开./include/caffe/common.cuh文件，在atomicAdd前添加宏判断即可。  下面为修改后文件：

// Copyright 2014 George Papandreou    #ifndef CAFFE_COMMON_CUH_    #define CAFFE_COMMON_CUH_    #include    // CUDA: atomicAdd is not defined for doubles    #if !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 600    #else    static __inline__ __device__ double atomicAdd(double *address, double val) {      unsigned long long int* address_as_ull = (unsigned long long int*)address;      unsigned long long int old = *address_as_ull, assumed;      if (val==0.0)        return __longlong_as_double(old);      do {        assumed = old;        old = atomicCAS(address_as_ull, assumed, __double_as_longlong(val +__longlong_as_double(assumed)));      } while (assumed != old);      return __longlong_as_double(old);    }    #endif    #endif

继续编译： 

make all -j16

如果出现下面的错误3：

:.build_release/lib/libcaffe.so:undefined  reference to `cudnnConvolutionBackwardFilter_v3’  

解决方法：  

将BVLC(https://github.com/BVLC/caffe)中的下列文件copy 到相应的文件夹：

./include/caffe/util/cudnn.hpp  ./include/caffe/layers/cudnn_conv_layer.hpp  ./include/caffe/layers/cudnn_relu_layer.hpp  ./include/caffe/layers/cudnn_sigmoid_layer.hpp  ./include/caffe/layers/cudnn_tanh_layer.hpp  ./src/caffe/layers/cudnn_conv_layer.cpp  ./src/caffe/layers/cudnn_conv_layer.cu  ./src/caffe/layers/cudnn_relu_layer.cpp  ./src/caffe/layers/cudnn_relu_layer.cu  ./src/caffe/layers/cudnn_sigmoid_layer.cpp  ./src/caffe/layers/cudnn_sigmoid_layer.cu  ./src/caffe/layers/cudnn_tanh_layer.cpp  ./src/caffe/layers/cudnn_tanh_layer.cu

然后：

make clean  

make all -j16  

make pycaffe -j16  

这个时候一般都是编译成功。

---

三、对 run_pascal.sh 进行调试：

• 首先准备好数据  ：

（参考： http://blog.csdn.net/Xmo_jiao/article/details/77897109）  

cd ~/Desktop  

mkdir -p my_dataset  

# augmented PASCAL VOC  

cd my_dataset/  

wget http://www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/semantic_contours/benchmark.tgz 

# 1.3 GB 

tar -zxvf benchmark.tgz  

mv benchmark_RELEASE VOC_aug  

# original PASCAL VOC 2012  

wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar 

# 2 GB  

tar -xvf VOCtrainval_11-May-2012.tar  

mv VOCdevkit/VOC2012 VOC2012_orig && rm -r VOCdevkit

---

• 数据转换

因为pascal voc2012增强数据集的label是mat格式的文件，要把mat格式的label转为png格式的图片

~/Desktop/my_dataset/VOC_aug/dataset  

mkdir cls_png  

cd ~/Desktop/deeplab_v2/voc2012/  

./mat2png.py ~/Desktop/my_dataset/VOC_aug/dataset/cls/Desktop/my_dataset/VOC_aug/dataset/cls_png

因为pascal voc2012原始数据集的label为三通道RGB图像，但是caffe最后一层softmax loss层只能识别一通道的label，所以此处我们需要对原始数据集的label进行降维  

cd ~/Desktop/my_dataset/VOC2012_orig  mkdir SegmentationClass_1D  

cd ~/Desktop/deeplab_v2/voc2012/  

./convert_labels.py ~/Desktop/my_dataset/VOC2012_orig/SegmentationClass/ ~/Desktop/my_dataset  /VOC2012_orig/ImageSets/Segmentation/trainval.txt ~/Desktop/my_dataset/VOC2012_orig/Segmentat  ionClass_1D/

---

• 数据融合

此时已经处理好好pascal voc2012 增强数据集和pascal voc2012的原始数据集，为了便于train.txt等文件的调用，将两个文件夹数据合并到同一个文件中.现有文件目录如下：

现分别pascal voc2012增强数据集里的images和labels复制到增强数据集中，若重复则覆盖，合将并数据集的操作如下：  

cp ~/Desktop/my_dataset/VOC2012_orig/SegmentationClass_1D/* ~/Desktop/my_dataset/VOC_aug/dataset/cls_png  

cp ~/Desktop/my_dataset/VOC2012_orig/JPEGImages/* ~/Desktop/my_dataset/VOC_aug/dataset/img/

---

• 文件名修改
  

对应train.txt文件的数据集文件名，修改文件名。  

cd ~/Desktop/my_dataset/VOC_aug/dataset  

mv ./img ./JPEGImages  

那么我们这个阶段使用的数据已经整理完成

---

四、修改并运行 run_pascal.sh

• 准备必要的文件  需要的文件从这里下载 deeplabv2 有两种模型（vgg，Res102），vgg ，http://liangchiehchen.com/projects/DeepLab_Models.html  

总共需要的文件如图所示：

下载的代码中 Desktop/deeplab_v2/voc2012/list 已经有了list文件，所以不用重新下载。

/Desktop/deeplab_v2/voc2012/config/deeplab_largeFOV中也有了相应的文件，所以也无需下载。

Desktop/deeplab_v2/voc2012/model/deeplab_largeFOV 里没有model，需要把下载好的model放入文件，如图所示：

至此，所有需要的文件全部完毕。

---

五、运行 train 和 test

进入/Desktop/deeplab_v2/voc2012，修改 run_pascal.sh 文件，主要是修改路径，我的修改后的文件如下：

#!/bin/sh## MODIFY PATH for YOUR SETTINGROOT_DIR=/home/mmt/Desktop/my_dataset CAFFE_DIR=/home/mmt/Desktop/deeplab_v2/deeplab-public-ver2 CAFFE_BIN=${CAFFE_DIR}/build/tools/caffe.bin EXP=.if [ "${EXP}" = "." ]; then    NUM_LABELS=21    DATA_ROOT=${ROOT_DIR}/VOC_aug/dataset/else    NUM_LABELS=0    echo "Wrong exp name"fi## Specify which model to train########### voc12 ################NET_ID=deeplab_largeFOV## Variables used for weakly or semi-supervisedly training#TRAIN_SET_SUFFIX=TRAIN_SET_SUFFIX=_aug#TRAIN_SET_STRONG=train#TRAIN_SET_STRONG=train200#TRAIN_SET_STRONG=train500#TRAIN_SET_STRONG=train1000#TRAIN_SET_STRONG=train750#TRAIN_SET_WEAK_LEN=5000DEV_ID=0####### Create dirsCONFIG_DIR=${EXP}/config/${NET_ID}MODEL_DIR=${EXP}/model/${NET_ID}mkdir -p ${MODEL_DIR}LOG_DIR=${EXP}/log/${NET_ID}mkdir -p ${LOG_DIR}export GLOG_log_dir=${LOG_DIR}## RunRUN_TRAIN=1   #1时trainRUN_TEST=0    #1时testRUN_TRAIN2=0RUN_TEST2=0## Training #1 (on train_aug)if [ ${RUN_TRAIN} -eq 1 ]; then    #    LIST_DIR=${EXP}/list    TRAIN_SET=train${TRAIN_SET_SUFFIX}    if [ -z ${TRAIN_SET_WEAK_LEN} ]; then                TRAIN_SET_WEAK=${TRAIN_SET}_diff_${TRAIN_SET_STRONG}                comm -3 ${LIST_DIR}/${TRAIN_SET}.txt ${LIST_DIR}/${TRAIN_SET_STRONG}.txt > ${LIST_DIR}/${TRAIN_SET_WEAK}.txt    else                TRAIN_SET_WEAK=${TRAIN_SET}_diff_${TRAIN_SET_STRONG}_head${TRAIN_SET_WEAK_LEN}                comm -3 ${LIST_DIR}/${TRAIN_SET}.txt ${LIST_DIR}/${TRAIN_SET_STRONG}.txt | head -n ${TRAIN_SET_WEAK_LEN} > ${LIST_DIR}/${TRAIN_SET_WEAK}.txt    fi    #    MODEL=${EXP}/model/${NET_ID}/init.caffemodel    #    echo Training net ${EXP}/${NET_ID}    for pname in train solver; do                sed "$(eval echo $(cat sub.sed))" \                        ${CONFIG_DIR}/${pname}.prototxt > ${CONFIG_DIR}/${pname}_${TRAIN_SET}.prototxt    done        CMD="${CAFFE_BIN} train \         --solver=${CONFIG_DIR}/solver_${TRAIN_SET}.prototxt \         --gpu=${DEV_ID}"        if [ -f ${MODEL} ]; then                CMD="${CMD} --weights=${MODEL}"        fi        echo Running ${CMD} && ${CMD}fi## Test #1 specification (on val or test)if [ ${RUN_TEST} -eq 1 ]; then    #    for TEST_SET in val; do                TEST_ITER=`cat ${EXP}/list/${TEST_SET}.txt | wc -l`                MODEL=${EXP}/model/${NET_ID}/test.caffemodel                if [ ! -f ${MODEL} ]; then                        MODEL=`ls -t ${EXP}/model/${NET_ID}/train_iter_*.caffemodel | head -n 1`                fi                #                echo Testing net ${EXP}/${NET_ID}                FEATURE_DIR=${EXP}/features/${NET_ID}                mkdir -p ${FEATURE_DIR}/${TEST_SET}/fc8        mkdir -p ${FEATURE_DIR}/${TEST_SET}/fc9                mkdir -p ${FEATURE_DIR}/${TEST_SET}/seg_score                sed "$(eval echo $(cat sub.sed))" \                        ${CONFIG_DIR}/test.prototxt > ${CONFIG_DIR}/test_${TEST_SET}.prototxt                CMD="${CAFFE_BIN} test \             --model=${CONFIG_DIR}/test_${TEST_SET}.prototxt \             --weights=${MODEL} \             --gpu=${DEV_ID} \             --iterations=${TEST_ITER}"                echo Running ${CMD} && ${CMD}    donefi## Training #2 (finetune on trainval_aug)if [ ${RUN_TRAIN2} -eq 1 ]; then    #    LIST_DIR=${EXP}/list    TRAIN_SET=trainval${TRAIN_SET_SUFFIX}    if [ -z ${TRAIN_SET_WEAK_LEN} ]; then                TRAIN_SET_WEAK=${TRAIN_SET}_diff_${TRAIN_SET_STRONG}                comm -3 ${LIST_DIR}/${TRAIN_SET}.txt ${LIST_DIR}/${TRAIN_SET_STRONG}.txt > ${LIST_DIR}/${TRAIN_SET_WEAK}.txt    else                TRAIN_SET_WEAK=${TRAIN_SET}_diff_${TRAIN_SET_STRONG}_head${TRAIN_SET_WEAK_LEN}                comm -3 ${LIST_DIR}/${TRAIN_SET}.txt ${LIST_DIR}/${TRAIN_SET_STRONG}.txt | head -n ${TRAIN_SET_WEAK_LEN} > ${LIST_DIR}/${TRAIN_SET_WEAK}.txt    fi    #    MODEL=${EXP}/model/${NET_ID}/init2.caffemodel    if [ ! -f ${MODEL} ]; then                MODEL=`ls -t ${EXP}/model/${NET_ID}/train_iter_*.caffemodel | head -n 1`    fi    #    echo Training2 net ${EXP}/${NET_ID}    for pname in train solver2; do                sed "$(eval echo $(cat sub.sed))" \                        ${CONFIG_DIR}/${pname}.prototxt > ${CONFIG_DIR}/${pname}_${TRAIN_SET}.prototxt    done    CMD="${CAFFE_BIN} train \         --solver=${CONFIG_DIR}/solver2_${TRAIN_SET}.prototxt \         --weights=${MODEL} \         --gpu=${DEV_ID}"        echo Running ${CMD} && ${CMD}fi## Test #2 on official test setif [ ${RUN_TEST2} -eq 1 ]; then    #    for TEST_SET in val test; do                TEST_ITER=`cat ${EXP}/list/${TEST_SET}.txt | wc -l`                MODEL=${EXP}/model/${NET_ID}/test2.caffemodel                if [ ! -f ${MODEL} ]; then                        MODEL=`ls -t ${EXP}/model/${NET_ID}/train2_iter_*.caffemodel | head -n 1`                fi                #                echo Testing2 net ${EXP}/${NET_ID}                FEATURE_DIR=${EXP}/features2/${NET_ID}                mkdir -p ${FEATURE_DIR}/${TEST_SET}/fc8                mkdir -p ${FEATURE_DIR}/${TEST_SET}/crf                sed "$(eval echo $(cat sub.sed))" \                        ${CONFIG_DIR}/test.prototxt > ${CONFIG_DIR}/test_${TEST_SET}.prototxt                CMD="${CAFFE_BIN} test \             --model=${CONFIG_DIR}/test_${TEST_SET}.prototxt \             --weights=${MODEL} \             --gpu=${DEV_ID} \             --iterations=${TEST_ITER}"                echo Running ${CMD} && ${CMD}    donefi

接下来运行代码：  

Train：  

~/Desktop/deeplab_v2/voc2012  

sh ./run_pascal.sh  

运行结果如下：

Test：  

将相应变量改为1：  

~/Desktop/deeplab_v2/voc2012  

sh ./run_pascal.sh  

运行结果如下：

因为结果保存的是mat文件，如果想转换成png的形式，运行：  

cd ~/Desktop/deeplab_v2/voc2012  

修改create_labels_21.py的路径，然后此目录运行：  

python create_labels_21.py 

---

六、修改并运行 run_densecrf.sh

• 首先对densecrf进行编译。  

cd ~/Desktop/deeplab_v2/deeplab-public-ver2/densecrf/  make  

有许多warning，但是没出错，不用管。

• 数据整理

因为densecrf只识别ppm格式的图像，所以要转换格式。

进入/Desktop/deeplab_v2/deeplab-public-ver2/densecrf/my_script，里面有自带的修改ppm 的MATLAB程序，修改路径，直接运行即可。  

代码如下：

% save jpg images as bin file for cpp%is_server = 1; dataset = 'voc2012';  %'coco', 'voc2012'if is_server  if strcmp(dataset, 'voc2012')    img_folder  = '/home/mmt/Desktop/my_dataset/VOC_aug/dataset/JPEGImages'    save_folder = '/home/mmt/Desktop/my_dataset/VOC_aug/dataset/PPMImages';  elseif strcmp(dataset, 'coco')    img_folder  = '/rmt/data/coco/JPEGImages';    save_folder = '/rmt/data/coco/PPMImages';  endelse  img_folder = '../img';  save_folder = '../img_ppm';endif ~exist(save_folder, 'dir')    mkdir(save_folder);endimg_dir = dir(fullfile(img_folder, '*.jpg'));for i = 1 : numel(img_dir)    fprintf(1, 'processing %d (%d)...\n', i, numel(img_dir));    img = imread(fullfile(img_folder, img_dir(i).name));    img_fn = img_dir(i).name(1:end-4);    save_fn = fullfile(save_folder, [img_fn, '.ppm']);    imwrite(img, save_fn);   end

接下来，修改 run_densecrf.sh， 注意把 MODEL_NAME=deeplab_largeFOV修改了。

DATASET=voc2012  修改；SAVE_DIR=/home/mmt/Desktop/deeplab_v2/${DATASET}/res/${FEATURE_NAME}/${MODEL_NAME}/${TEST_SET} 修改；CRF_DIR=/home/mmt/Desktop/deeplab_v2/deeplab-public-ver2/densecrf  修改；if [ ${DATASET} == "voc2012" ]then    IMG_DIR_NAME=VOC_aug/dataset     修改；FEATURE_DIR=/home/mmt/Desktop/deeplab_v2/${DATASET}/${FEATURE_NAME}/${MODEL_NAME}/${TEST_SET}/${FEATURE_TYPE}  修改； 同时把一些不需要的语句都注释掉，要不然容易出错，显示找不到文件。 修改后的文件如下：#!/bin/bash ############################################ You can either use this script to generate the DenseCRF post-processed results# or use the densecrf_layer (wrapper) in Caffe###########################################DATASET=voc2012LOAD_MAT_FILE=1MODEL_NAME=deeplab_largeFOVTEST_SET=val           #val, test# the features  folder save the features computed via the model trained with the train set# the features2 folder save the features computed via the model trained with the trainval setFEATURE_NAME=features #features, features2FEATURE_TYPE=fc8# specify the parametersMAX_ITER=10Bi_W=4Bi_X_STD=49Bi_Y_STD=49Bi_R_STD=5Bi_G_STD=5 Bi_B_STD=5POS_W=3POS_X_STD=3POS_Y_STD=3######################################## MODIFY THE PATY FOR YOUR SETTING#######################################SAVE_DIR=/home/mmt/Desktop/deeplab_v2/${DATASET}/res/${FEATURE_NAME}/${MODEL_NAME}/${TEST_SET}/${FEATURE_TYPE}/post_densecrf_W${Bi_W}_XStd${Bi_X_STD}_RStd${Bi_R_STD}_PosW${POS_W}_PosXStd${POS_X_STD} echo "SAVE TO ${SAVE_DIR}"CRF_DIR=/home/mmt/Desktop/deeplab_v2/deeplab-public-ver2/densecrf#if [ ${DATASET} == "voc2012" ]#then    IMG_DIR_NAME=VOC_aug/dataset#elif [ ${DATASET} == "coco" ]#then #   IMG_DIR_NAME=coco#elif [ ${DATASET} == "voc10_part" ]#then  #  IMG_DIR_NAME=pascal/VOCdevkit/VOC2012#fi# NOTE THAT the densecrf code only loads ppm imagesIMG_DIR=/home/mmt/Desktop/my_dataset/${IMG_DIR_NAME}/PPMImages#if [ ${LOAD_MAT_FILE} == 1 ]#then    # the features are saved in .mat format    CRF_BIN=${CRF_DIR}/prog_refine_pascal_v4    FEATURE_DIR=/home/mmt/Desktop/deeplab_v2/${DATASET}/${FEATURE_NAME}/${MODEL_NAME}/${TEST_SET}/${FEATURE_TYPE}#else    # the features are saved in .bin format (has called SaveMatAsBin.m in the densecrf/my_script)   # CRF_BIN=${CRF_DIR}/prog_refine_pascal   # FEATURE_DIR=/home/mmt/Desktop/deeplab_v2/${DATASET}/${FEATURE_NAME}/${MODEL_NAME}/${TEST_SET}/${FEATURE_TYPE}/bin#fimkdir -p ${SAVE_DIR}# run the program${CRF_BIN} -id ${IMG_DIR} -fd ${FEATURE_DIR} -sd ${SAVE_DIR} -i ${MAX_ITER} -px ${POS_X_STD} -py ${POS_Y_STD} -pw ${POS_W} -bx ${Bi_X_STD} -by ${Bi_Y_STD} -br ${Bi_R_STD} -bg ${Bi_G_STD} -bb ${Bi_B_STD} -bw ${Bi_W}

进入文件路径，运行程序，结果如下图：  

cd ~/Desktop/deeplab_v2/voc2012/  

sh sh ./run_densecrf.sh

然后运行

/home/mmt/crf/deeplab-public-ver2/densecrf/my_script/GetDenseCRFResult.m

把bin生成图片格式  

注意修改文件路径（GetDenseCRFResult.m，SetupEnv在/deeplab_v2/deeplab-public-ver2/matlab/my_script中）

两个程序的代码如下：

GetDenseCRFResult.m % compute the densecrf result (.bin) to png % addpath('/home/mmt/Desktop/deeplab_v2/deeplab-public-ver2/matlab/my_script');SetupEnv;%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% You do not need to change values below%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%if is_server  if learn_crf    post_folder = sprintf('post_densecrf_W%d_XStd%d_RStd%d_PosW%d_PosXStd%d_ModelType%d_Epoch%d', bi_w, bi_x_std, bi_r_std, pos_w, pos_x_std, model_type, epoch);    map_folder = fullfile('/home/mmt/Desktop/deeplab_v2', dataset, 'densecrf', 'res', feature_name, model_name, testset, feature_type, post_folder);    save_root_folder = fullfile('/home/mmt/Desktop/deeplab_v2', dataset, 'res', feature_name, model_name, testset, feature_type, post_folder); ;  else    post_folder = sprintf('post_densecrf_W%d_XStd%d_RStd%d_PosW%d_PosXStd%d', bi_w, bi_x_std, bi_r_std, pos_w, pos_x_std);    map_folder = fullfile('/home/mmt/Desktop/deeplab_v2', dataset, 'res', feature_name, model_name, testset, feature_type, post_folder);    save_root_folder = map_folder;  endelse  map_folder = '../result';endmap_dir = dir(fullfile(map_folder, '*.bin')); fprintf(1,' saving to %s\n', save_root_folder);if strcmp(dataset, 'voc2012')  seg_res_dir = [save_root_folder '/results/VOC2012/']; elseif strcmp(dataset, 'coco')  seg_res_dir = [save_root_folder, '/results/COCO2014/'];else  error('Wrong dataset!');endsave_result_folder = fullfile(seg_res_dir, 'Segmentation', [id '_' testset '_cls']);if ~exist(save_result_folder, 'dir')    mkdir(save_result_folder);endfor i = 1 : numel(map_dir)    fprintf(1, 'processing %d (%d)...\n', i, numel(map_dir));    map = LoadBinFile(fullfile(map_folder, map_dir(i).name), 'int16');    img_fn = map_dir(i).name(1:end-4);    imwrite(uint8(map), colormap, fullfile(save_result_folder, [img_fn, '.png']));end

---

SetupEnv.m % set up the environment variables % clear all; close all; load('./pascal_seg_colormap.mat'); is_server       = 1; crf_load_mat    = 1;   % the densecrf code load MAT files directly (no call SaveMatAsBin.m)                       % used ONLY by DownSampleFeature.m learn_crf       = 0;   % NOT USED. Set to 0is_mat          = 1;   % the results to be evaluated are saved as mat (1) or png (0) has_postprocess = 0;   % has done densecrf post processing (1) or not (0) is_argmax       = 0;   % the output has been taken argmax already (e.g., coco dataset).                       % assume the argmax takes C-convention (i.e., start from 0) debug           = 0;   % if debug, show some results % vgg128_noup (not optimized well), aka DeepLab% bi_w = 5, bi_x_std = 50, bi_r_std = 10% vgg128_ms_pool3, aka DeepLab-MSc% bi_w = 3, bi_x_std = 95, bi_r_std = 3% vgg128_noup_pool3_cocomix, aka DeepLab-COCO% bi_w = 5, bi_x_std = 67, bi_r_std = 3%% these are used for the bounding box weak annotation experiments (i.e., to generate the Bbox-Seg) % erode_gt (bbox) % bi_w = 41, bi_x_std = 33, bi_r_std = 4% erode_gt/bboxErode20% bi_w = 45, bi_x_std = 37, bi_r_std = 3, pos_w = 15, pos_x_std = 3% % initial or default values for crf%% 这几个参数要修改与run_densecrf.sh中的一致。 bi_w           = 4; bi_x_std       = 49; bi_r_std       = 5; pos_w          = 3; pos_x_std      = 3; %dataset    = 'voc2012';  %'voc12', 'coco'  修改 trainset   = 'train_aug';      % not used testset    = 'val';            %'val', 'test'model_name = 'deeplab_largeFOV';  % 修改 feature_name = 'features'; feature_type = 'fc8'; % fc8 / crf id           = 'comp6';%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% used for cross-validation %%%%%%%%%%%%%%%%%%%%%%%%%%%%%% rng(10) % downsampling files for cross-validation down_sample_method = 2;      % 1: equally sample with "down_sample_rate", 2: randomly pick "num_sample" samples down_sample_rate   = 8; num_sample         = 100;    % number of samples used for cross-validation % ranges for cross-validation range_pos_w = [3]; range_pos_x_std = [3]; range_bi_w = [5]; range_bi_x_std = [49]; range_bi_r_std = [4 5];

---

• 至此，deeplabv2 程序已调试完

---

感谢 ruotianxia的分享！