Caffe 训练命令行（solverstate开始继续训练） 测试---＞ 直接返回acc和loss lmdb数据库制作

进入caffe根目录

./build/tools/caffe.bin test -model ../deploy.prototxt -weights ../resnet-18.caffemodel -gpu 0

一、train.prototxt

• 进入caffe根目录执行，自带的test可以测出top1 top5 acc。
• 直接从头训练或者从预训练的模型开始训练。
• 从solverstate继续中断的训练。

#####test
./build/tools/caffe.bin test -model ../compression/train_cp.prototxt -weights ../models/resnet-18_cp_val_trained_iter_10000.caffemodel -gpu 0
/home/allen/Documents/lck/Resnet-FPGA/caffe-resnet18/models/resnet-18_cp_val_trained_iter_10000.caffemodel

####
# acc1=0.6544  
# acc5=0.8592    
# loss=1.4521

# train 直接从头开始:
./build/tools/caffe.bin train -solver ../compression/solver_cp.prototxt -weights /home/allen/Documents/lck/Resnet-FPGA/caffe-resnet18/compression/models/resnet18-compress-1.caffemodel -gpu all
从weights继续训练，loss不变，并不work
# 用伪逆 可以收敛

# 继续训练
./build/tools/caffe.bin train -solver ../compression/solver_cp.prototxt -weights /home/allen/Documents/lck/Resnet-FPGA/caffe-resnet18/compression/models/resnet18-compress-1.caffemodel -gpu all
# solverstate
./build/tools/caffe.bin train -solver ../compression/solver_cp.prototxt -snapshot ../models/resnet-18_cp_val_compressed_init_iter_45889.solverstate -gpu all

# epoch
总样本数/batchsize (iteration) = 1 (epoch)

二 ·Imdb格式文件获取

得到listfile：文件清单 找到train.txt

bash examples/images/create_filelist.sh

找到tools/create_imagenet.sh文件

# 修改路径
EXAMPLE=/home/allen/Documents/lck/dataset/ILSVRC2012
DATA=/home/allen/Documents/lck/dataset/ILSVRC2012
TOOLS=/home/allen/Documents/lck/Resnet-FPGA/resnet18-caffe/caffe-master/build/tools

TRAIN_DATA_ROOT=/home/allen/Documents/lck/dataset/ILSVRC2012/train/
VAL_DATA_ROOT=/home/allen/Documents/lck/dataset/ILSVRC2012/val/
# 修改名字
$DATA/train.txt \
    $EXAMPLE/ilsvrc12_train_lmdb

若数据集需要resize，直接运行会报错

Check failed: width ＜= datum_width (28 vs. 17)

RESIZE=true
size=224

# 运行 create lmdb
sh create_imagenet.sh