【caffe】标准数据层输入

caffe在构造TrainNet时&＃xff0c;首先要做的便是data和label的输入。数据层在Caffe的Data Layer中位于最底层&＃xff0c;数据可以在高效的数据库中读取&＃xff0c;也可以直接在内存中读取&＃xff0c;或者从硬盘中直接读取

1. 数据层 LevedB 和 LMDB
层类型为data&＃xff0c;主要分为LevedB 和 LMDB
其在*.prototxt文件中如下所示&＃xff1a;

layer {name: "mnist"type: "Data"top: "data"top: "label"include {phase: TRAIN}transform param{ #这个参数主要负责输入数据的预处理scale &＃61; 0.00390625 #如减均值&＃xff0c;尺度变换&＃xff0c;随机裁剪或者镜像}data_param{source: "example/mnist/mnist_train.lmdb"}batch_size:64backend: LMDB
}

这里给出一种生成这种pycaffe的生成代码&＃xff1a;

n.data, n.label &＃61; L.Data(name &＃61; &＃39;MNIST&＃39;, source &＃61; "example/mnist/mnist_train.lmdb", batch_size &＃61; 64, include &＃61; {&＃39;phase&＃39;:0}, backend &＃61; P.Data.LMDB, ntop &＃61; 2, transform_param &＃61; dict(scale &＃61; 0.00390625))

2. 内存数据

layer {name: "memory_data"type: "MemoryData"top: "data"top: "label"memory_data_param{batch_size:2height: 100width: 100channels: 1}transform param{ scale: 0.0078125mean_file: "mean.proto"mirror: false}
}

3. HDF5数据

layer{name: "data"type: "HDF5Data"top: "data"top: "label"hdf5_data_param{source: "example/mnist/mnist_train.lmdb"batch_size: 10}
}

pycaffe接口&＃xff1a;

n.data, n.label &＃61; L.HDF5Data(name &＃61; &＃39;data&＃39;, source &＃61; "example/mnist/mnist_train.lmdb", batch_size &＃61; 10, ntop &＃61; 2)

4. 图像数据Images

layer {name: "data"type: "ImageData"top: "data"top: "label"image_data_param{source: "example/mnist/mnist_train.lmdb"batch_size:50new_height: 256new_width: 256}transform param{ mirror: falsecrop_size: 227mean_file : "mean.proto"}
}

事实caffe可以自己来定义数据输入层&＃xff0c;比如FCN代码的Python Layer形式&＃xff0c;通过继承data layer的一些函数&＃xff1a;
我直接把FCN中的PythonLayer放上来&＃xff0c;这里仅仅作为事例&＃xff1a;

import caffeimport numpy as np
from PIL import Imageimport randomclass VOCSegDataLayer(caffe.Layer):"""Load (input image, label image) pairs from PASCAL VOCone-at-a-time while reshaping the net to preserve dimensions.Use this to feed data to a fully convolutional network."""def setup(self, bottom, top):"""Setup data layer according to parameters:- voc_dir: path to PASCAL VOC year dir- split: train / val / test- mean: tuple of mean values to subtract- randomize: load in random order (default: True)- seed: seed for randomization (default: None / current time)for PASCAL VOC semantic segmentation.exampleparams &＃61; dict(voc_dir&＃61;"/path/to/PASCAL/VOC2011",mean&＃61;(104.00698793, 116.66876762, 122.67891434),split&＃61;"val")"""# configparams &＃61; eval(self.param_str)self.voc_dir &＃61; params[&＃39;voc_dir&＃39;]self.split &＃61; params[&＃39;split&＃39;]self.mean &＃61; np.array(params[&＃39;mean&＃39;])self.random &＃61; params.get(&＃39;randomize&＃39;, True)self.seed &＃61; params.get(&＃39;seed&＃39;, None)# two tops: data and labelif len(top) !&＃61; 2:raise Exception("Need to define two tops: data and label.")# data layers have no bottomsif len(bottom) !&＃61; 0:raise Exception("Do not define a bottom.")# load indices for images and labelssplit_f &＃61; &＃39;{}/ImageSets/Segmentation/{}.txt&＃39;.format(self.voc_dir,self.split)self.indices &＃61; open(split_f, &＃39;r&＃39;).read().splitlines()self.idx &＃61; 0# make eval deterministicif &＃39;train&＃39; not in self.split:self.random &＃61; False# randomization: seed and pickif self.random:random.seed(self.seed)self.idx &＃61; random.randint(0, len(self.indices)-1)def reshape(self, bottom, top):# load image &＃43; label image pairself.data &＃61; self.load_image(self.indices[self.idx])self.label &＃61; self.load_label(self.indices[self.idx])# reshape tops to fit (leading 1 is for batch dimension)top[0].reshape(1, *self.data.shape)top[1].reshape(1, *self.label.shape)def forward(self, bottom, top):# assign outputtop[0].data[...] &＃61; self.datatop[1].data[...] &＃61; self.label# pick next inputif self.random:self.idx &＃61; random.randint(0, len(self.indices)-1)else:self.idx &＃43;&＃61; 1if self.idx &＃61;&＃61; len(self.indices):self.idx &＃61; 0def backward(self, top, propagate_down, bottom):passdef load_image(self, idx):"""Load input image and preprocess for Caffe:- cast to float- switch channels RGB -> BGR- subtract mean- transpose to channel x height x width order"""im &＃61; Image.open(&＃39;{}/JPEGImages/{}.jpg&＃39;.format(self.voc_dir, idx))in_ &＃61; np.array(im, dtype&＃61;np.float32)in_ &＃61; in_[:,:,::-1]in_ -&＃61; self.meanin_ &＃61; in_.transpose((2,0,1))return in_def load_label(self, idx):"""Load label image as 1 x height x width integer array of label indices.The leading singleton dimension is required by the loss."""im &＃61; Image.open(&＃39;{}/SegmentationClass/{}.png&＃39;.format(self.voc_dir, idx))label &＃61; np.array(im, dtype&＃61;np.uint8)label &＃61; label[np.newaxis, ...]return labelclass SBDDSegDataLayer(caffe.Layer):"""Load (input image, label image) pairs from the SBDD extended labelingof PASCAL VOC for semantic segmentationone-at-a-time while reshaping the net to preserve dimensions.Use this to feed data to a fully convolutional network."""def setup(self, bottom, top):"""Setup data layer according to parameters:- sbdd_dir: path to SBDD &＃96;dataset&＃96; dir- split: train / seg11valid- mean: tuple of mean values to subtract- randomize: load in random order (default: True)- seed: seed for randomization (default: None / current time)for SBDD semantic segmentation.N.B.segv11alid is the set of segval11 that does not intersect with SBDD.Find it here: https://gist.github.com/shelhamer/edb330760338892d511e.exampleparams &＃61; dict(sbdd_dir&＃61;"/path/to/SBDD/dataset",mean&＃61;(104.00698793, 116.66876762, 122.67891434),split&＃61;"valid")"""# configparams &＃61; eval(self.param_str)self.sbdd_dir &＃61; params[&＃39;sbdd_dir&＃39;]self.split &＃61; params[&＃39;split&＃39;]self.mean &＃61; np.array(params[&＃39;mean&＃39;])self.random &＃61; params.get(&＃39;randomize&＃39;, True)self.seed &＃61; params.get(&＃39;seed&＃39;, None)# two tops: data and labelif len(top) !&＃61; 2:raise Exception("Need to define two tops: data and label.")# data layers have no bottomsif len(bottom) !&＃61; 0:raise Exception("Do not define a bottom.")# load indices for images and labelssplit_f &＃61; &＃39;{}/{}.txt&＃39;.format(self.sbdd_dir,self.split)self.indices &＃61; open(split_f, &＃39;r&＃39;).read().splitlines()self.idx &＃61; 0# make eval deterministicif &＃39;train&＃39; not in self.split:self.random &＃61; False# randomization: seed and pickif self.random:random.seed(self.seed)self.idx &＃61; random.randint(0, len(self.indices)-1)def reshape(self, bottom, top):# load image &＃43; label image pairself.data &＃61; self.load_image(self.indices[self.idx])self.label &＃61; self.load_label(self.indices[self.idx])# reshape tops to fit (leading 1 is for batch dimension)top[0].reshape(1, *self.data.shape)top[1].reshape(1, *self.label.shape)def forward(self, bottom, top):# assign outputtop[0].data[...] &＃61; self.datatop[1].data[...] &＃61; self.label# pick next inputif self.random:self.idx &＃61; random.randint(0, len(self.indices)-1)else:self.idx &＃43;&＃61; 1if self.idx &＃61;&＃61; len(self.indices):self.idx &＃61; 0def backward(self, top, propagate_down, bottom):passdef load_image(self, idx):"""Load input image and preprocess for Caffe:- cast to float- switch channels RGB -> BGR- subtract mean- transpose to channel x height x width order"""im &＃61; Image.open(&＃39;{}/img/{}.jpg&＃39;.format(self.sbdd_dir, idx))in_ &＃61; np.array(im, dtype&＃61;np.float32)in_ &＃61; in_[:,:,::-1]in_ -&＃61; self.meanin_ &＃61; in_.transpose((2,0,1))return in_def load_label(self, idx):"""Load label image as 1 x height x width integer array of label indices.The leading singleton dimension is required by the loss."""import scipy.iomat &＃61; scipy.io.loadmat(&＃39;{}/cls/{}.mat&＃39;.format(self.sbdd_dir, idx))label &＃61; mat[&＃39;GTcls&＃39;][0][&＃39;Segmentation&＃39;][0].astype(np.uint8)label &＃61; label[np.newaxis, ...]return labe


在Pycaffe的Net生成中&＃xff1a;

n.data, n.label &＃61; L.Python(module&＃61;&＃39;voc_layers&＃39;, layer&＃61;pylayer,ntop&＃61;2, param_str&＃61;str(pydata_params))

Python Layer还是比较好用的&＃xff0c;可以随心所欲的定义自己的输入。还是要向大牛前辈学习一个