一、im2rec用法简介
首先看文档:
usage: im2rec.py [-h] [--list] [--exts EXTS [EXTS ...]] [--chunks CHUNKS] [--train-ratio TRAIN_RATIO] [--test-ratio TEST_RATIO] [--recursive] [--no-shuffle] [--pass-through] [--resize RESIZE] [--center-crop] [--quality QUALITY] [--num-thread NUM_THREAD] [--color {-1,0,1}] [--encoding {.jpg,.png}] [--pack-label] prefix root Create an image list or make a record database by reading from an image list positional arguments: prefix prefix of input/output lst and rec files. root path to folder containing images. optional arguments: -h, --help show this help message and exit Options for creating image lists: --list If this is set im2rec will create image list(s) by traversing root folder and output to <prefix>.lst. Otherwise im2rec will read <prefix>.lst and create a database at <prefix>.rec (default: False) --exts EXTS [EXTS ...] list of acceptable image extensions. (default: ['.jpeg', '.jpg', '.png']) --chunks CHUNKS number of chunks. (default: 1) --train-ratio TRAIN_RATIO Ratio of images to use for training. (default: 1.0) --test-ratio TEST_RATIO Ratio of images to use for testing. (default: 0) --recursive If true recursively walk through subdirs and assign an unique label to images in each folder. Otherwise only include images in the root folder and give them label 0. (default: False) --no-shuffle If this is passed, im2rec will not randomize the image order in <prefix>.lst (default: True) Options for creating database: --pass-through whether to skip transformation and save image as is (default: False) --resize RESIZE resize the shorter edge of image to the newsize, original images will be packed by default. (default: 0) --center-crop specify whether to crop the center image to make it rectangular. (default: False) --quality QUALITY JPEG quality for encoding, 1-100; or PNG compression for encoding, 1-9 (default: 95) --num-thread NUM_THREAD number of thread to use for encoding. order of images will be different from the input list if >1. the input list will be modified to match the resulting order. (default: 1) --color {-1,0,1} specify the color mode of the loaded image. 1: Loads a color image. Any transparency of image will be neglected. It is the default flag. 0: Loads image in grayscale mode. -1:Loads image as such including alpha channel. (default: 1) --encoding {.jpg,.png} specify the encoding of the images. (default: .jpg) --pack-label Whether to also pack multi dimensional label in the record file (default: False)
必须要填写的参数有prefix、和root两个路径参数,
prefix:生成文件文件夹目录
可以指定为.lst的路径,这样生成文件会和.lst同一级别,且会根据.lst中的条目生成二进制文件
root:图片文件目录,默认的话里面是类别文件夹,类别名做label,每个文件夹存储图像
如果指定了.lst,则每个图片路径变为root路径+.lst中每个图片的路径
--pack-label:在指定了.lst后很有用,此时允许label为高维度(主要是label可以设置为数组)
实际上我们之前也介绍过,.lst文件并不是必须的,仅有.rec和.idx就可以满足需要(标签存储在.rec中),它是一个辅助(人工生成.lst指导.rec生成),或者作为一个结果展示(自动生成.rec时选择同时生成.lst)。
【注意】,最新版本的pack-label等bool参数已经变成了开关型参数,即输入--pack-label表示原意True。
二、通用图像数据存储以及迭代读取方式
1、生成.lst文件
具体格式如下,各列之间使用' '分隔,第一列为索引,最后一列为文件路径,这两列是固定的,中间为标签列,列数不固定。
0 2 5 0 0.0 0.0 0.3 0.3 2007_000129.jpg
1 2 5 1 0.1 0.1 0.4 0.4 2007_000027.jpg
2 2 5 2 0.2 0.2 0.5 0.5 2007_000123.jpg
3 2 5 3 0.3 0.3 0.6 0.6 2007_000063.jpg
4 2 5 4 0.4 0.4 0.7 0.7 2007_000033.jpg
5 2 5 5 0.5 0.5 0.8 0.8 2007_000121.jpg
6 2 5 6 0.6 0.6 0.9 0.9 2007_000042.jpg
7 2 5 7 0.7 0.7 1.0 1.0 2007_000039.jpg
8 2 5 8 0.8 0.8 1.1 1.1 2007_000032.jpg
9 2 5 9 0.9 0.9 1.2 1.2 2007_000061.jpg
10 2 5 10 1.0 1.0 1.3 1.3 2007_000068.jpg
11 2 5 11 1.1 1.1 1.4 1.4 2007_000170.jpg
代码如下,
import os name_list = [f for f in os.listdir('./') if f.endswith('jpg')] with open('my_test.lst', 'w+') as f: for i, n in enumerate(name_list): f.write( str(i) + ' ' + # idx '2' + ' ' + '5' + ' ' + # 头信息长度(不含索引), 每个obj长度 str(i) + ' ' + # class str((i / 10)) + ' ' + str((i / 10)) + ' ' + str(((i + 3) / 10)) + ' ' +str(((i + 3) / 10)) + ' ' + # xmin, ymin, xmax, ymax n + ' ' # image path )
2、创建.rec
调用子进程创建rec文件:
import subprocess import mxnet as mx im2rec_path = os.path.join(mx.__path__[0], 'tools/im2rec.py') # 寻找im2rec.py路径 # final validation - sometimes __path__ (or __file__) gives 'mxnet/python/mxnet' instead of 'mxnet' if not os.path.exists(im2rec_path): im2rec_path = os.path.join(os.path.dirname(os.path.dirname(mx.__path__[0])), 'tools/im2rec.py') subprocess.check_call(["python", im2rec_path, os.path.abspath('my_test.lst'), os.path.abspath('./'), "--pack-label"])
返回0表示进程顺利结束。
3、读取.rec
这里我门介绍一下几种读取API使用方式,ImageIter要求标签列数固定,label_width=7表示中间7列为标签列,我们打印了标签作示范:
data_iter = mx.image.ImageIter(batch_size=4, resize=30, label_width=7, data_shape=(3, 30, 60),# depth,height,width path_imgrec="./my_test.rec", path_imgidx="./my_test.idx" ) data_iter.reset() batch = data_iter.next() img, labels = batch.data[0], batch.label[0] print(labels)
[[ 2. 5. 0. 0. 0. 0.30000001 0.30000001] [ 2. 5. 1. 0.1 0.1 0.40000001 0.40000001] [ 2. 5. 2. 0.2 0.2 0.5 0.5 ] [ 2. 5. 3. 0.30000001 0.30000001 0.60000002 0.60000002]] <NDArray 4x7 @cpu(0)>
简明易懂,不过对于很多目标检测任务来说,object数目并不一致,中间的label列也就不一致,此时下面的API泛用性更好:
rec = mx.image.ImageDetIter( path_imgrec = './my_test.rec', path_imglist = '', batch_size = 4, data_shape = (3, 300, 300)) rec.next().label[0]
InageDetIter没有label_width参数,其扣除首末两列,中间都作为待定标签列:0列-索引列,1列-头信息列数,2列-每个对象信息列数,3~(n-1)列-标签列,n列-图像路径。
输出时会在1~n-1列中先扣除第二列数字的列数(本例中1、2两列被扣除),之后的列数才是标签:
[[[ 0. 0. 0. 0.30000001 0.30000001]] [[ 1. 0.1 0.1 0.40000001 0.40000001]] [[ 2. 0.2 0.2 0.5 0.5 ]] [[ 3. 0.30000001 0.30000001 0.60000002 0.60000002]]] <NDArray 4x1x5 @cpu(0)>
下面我们测试一下各张图片label长度不等的情况:
im2rec_path = os.path.join(mx.__path__[0], 'tools/im2rec.py') # 寻找im2rec.py路径 # final validation - sometimes __path__ (or __file__) gives 'mxnet/python/mxnet' instead of 'mxnet' if not os.path.exists(im2rec_path): im2rec_path = os.path.join(os.path.dirname(os.path.dirname(mx.__path__[0])), 'tools/im2rec.py') subprocess.check_call(["python", im2rec_path, os.path.abspath('my_test.lst'), os.path.abspath('./'), "--pack-label"]) rec = mx.image.ImageDetIter( path_imgrec = './my_test.rec', path_imglist = '', batch_size = 4, data_shape = (3, 300, 300)) rec.next().label[0]
label长度不足的图片使用-1进行了补齐:
[[[ 0. 0. 0. 0.30000001 0.30000001] [ 1. 2. 3. 4. 5. ]] [[ 1. 0.1 0.1 0.40000001 0.40000001] [-1. -1. -1. -1. -1. ]] [[ 2. 0.2 0.2 0.5 0.5 ] [-1. -1. -1. -1. -1. ]] [[ 3. 0.30000001 0.30000001 0.60000002 0.60000002] [-1. -1. -1. -1. -1. ]]] <NDArray 4x2x5 @cpu(0)>
这一点上反倒是TensorFlow宽松一点,如果是None位置,在同一个会话中也可以通过不同的形状:
import tensorflow as tf input_ = tf.placeholder(dtype=tf.int8, shape=(None,)) with tf.Session() as sess: print(sess.run(input_, feed_dict={input_:(10,1)})) print(sess.run(input_, feed_dict={input_:(10,1,5)}))
[10 1]
[10 1 5]
也即是说在标签读取中,TensorFlow不需要补全-1,不同图片的标签形状不需要进行统一(见前面的TensorFlow-SSD标签处理一节)。