darknect/caffe/train_test_val/readme.md
需要提供 一个 预训练的 模型权重文件
bvlc_googlenet.caffemodel
下载： http://dl.caffe.berkeleyvision.org/bvlc_googlenet.caffemodel
###### train.sh ##############
#!/usr/bin/env sh
# caffe main path
CAFFE_HOME=../..
# learning rate paramter
SOLVER=./gnet_solver.prototxt
# init weights
WEIGHTS=/home/wanyouwen/ewenwan/software/caffe-yolo/weights/bvlc_googlenet.caffemodel
# training --gpu=0,1,2,3,4,5,6,7
$CAFFE_HOME/build/tools/caffe train \
--solver=$SOLVER --weights=$WEIGHTS \
--gpu=0,1,2,3 2>&1 | tee train_yolov1.log
###### gnet_solver.prototxt #######
net: "examples/yolo/gnet_train.prototxt" # 网络配置文件
# 在这四个train_net_param, train_net, net_param, net字段中至少需要出现一个，
# 当出现多个时，就会按着(1) test_net_param, (2) test_net, (3) net_param/net 的顺序依次求解。
test_iter: 4952 # 迭代器
test_interval: 32000 # 是指测试间隔，每训练test_interval次，进行一次测试。
test_initialization: false # 是指在第一次迭代前，计算初始的loss以确保内存可用
display: 20 # 是信息显示间隔
average_loss: 100 # 用于显示在上次average_loss迭代中的平均损失
lr_policy: "multifixed"
#stagelr: 0.001
#stagelr: 0.01
#stagelr: 0.001
#stagelr: 0.0001
#stageiter: 520
#stageiter: 16000
#stageiter: 24000
#stageiter: 32000
max_iter: 32000
momentum: 0.9 # 动量 上次梯度更新的权重
weight_decay: 0.0005 # 权重衰减，防止过拟合
snapshot: 2000 # 快照 将训练出来的model和solver状态进行保存
snapshot_prefix: "./models/gnet_yolo"# 保存路径
solver_mode: GPU# 运行模式
使用 ./convert.sh 转换 图片数据 到 lmdb/leveldb
cd cafe/data/yolo
ln -s /your/path/to/VOCdevkit/ .
python ./get_list.py
# change related path in script convert.sh
./convert.sh
问题：
执行shell脚本时提示bad interpreter:No such file or directory的解决办法
convert.sh 格式不正确 dos 转成 unix notepad就可以转换
/////////////////////////////////
convert.sh
//////////// #!/usr/bin/env sh # cafe main path CAFFE_ROOT=../.. #/your/path/to/vocroot/ ROOT_DIR=/home/wanyouwen/ewenwan/software/darknet/data/voc/VOCdevkit # string label Map to int label LABEL_FILE=$CAFFE_ROOT/data/yolo/label_map.txt
# 2007 + 2012 trainval
# source img
LIST_FILE=$CAFFE_ROOT/data/yolo/trainval.txt
# date base file
LMDB_DIR=./lmdb/trainval_lmdb
# shuff;e
SHUFFLE=true
# 2007 test
# LIST_FILE=$CAFFE_ROOT/data/yolo/test_2007.txt
# LMDB_DIR=./lmdb/test2007_lmdb
# SHUFFLE=false
# resize size
RESIZE_W=448
RESIZE_H=448
# execute
$CAFFE_ROOT/build/tools/convert_box_data --resize_width=$RESIZE_W --resize_height=$RESIZE_H \
--label_file=$LABEL_FILE $ROOT_DIR $LIST_FILE $LMDB_DIR --encoded=true --encode_type=jpg --shuffle=$SHUFFLE
///////////////////////////////
见 https://github.com/yeahkun/caffe-yolo/blob/master/tools/convert_box_data.cpp
需要对
caffe/util/io.hpp
caffe/src/caffe/util/io.cpp 做一定修改
./train.sh
数据 添加了 框标签 以下需要修改 base_data_layer.cpp base_data_layer.cu box_data_layer.cpp
出现的问题 0519 02:13:01.861333 9687 layer_factory.hpp:77] Creating layer data I0519 02:13:01.861886 9687 net.cpp:84] Creating Layer data I0519 02:13:01.861901 9687 net.cpp:380] data -> data I0519 02:13:01.861933 9687 net.cpp:380] data -> label I0519 02:13:01.866948 9704 db_lmdb.cpp:35] Opened lmdb ../../data/yolo/trainval_lmdb I0519 02:13:01.873008 9687 box_data_layer.cpp:46] output data size: 64,3,448,448 F0519 02:13:01.963363 9687 blob.cpp:133] Check failed: data_ *** Check failure stack trace: *** @ 0x7f85615eddaa (unknown) @ 0x7f85615edce4 (unknown) @ 0x7f85615ed6e6 (unknown) @ 0x7f85615f0687 (unknown) @ 0x7f8561db39de caffe::Blob<>::mutable_cpu_data() @ 0x7f8561ca4dae caffe::BasePrefetchingDataLayer<>::LayerSetUp() @ 0x7f8561dacc35 caffe::Net<>::Init() @ 0x7f8561daeb52 caffe::Net<>::Net() @ 0x7f8561c7c92e caffe::Solver<>::InitTrainNet() @ 0x7f8561c7d973 caffe::Solver<>::Init() @ 0x7f8561c7dc4f caffe::Solver<>::Solver() @ 0x7f8561c26fe1 caffe::Creator_SGDSolver<>() @ 0x40dc6e caffe::SolverRegistry<>::CreateSolver() @ 0x40848d train() @ 0x405fec main @ 0x7f8560467f45 (unknown) @ 0x40685b (unknown) @ (nil) (unknown) Aborted
未解决