Faster R-CNN 和自定义 VOC 数据集
制作VOC数据
修改文件名
因为VOC文件名都是使用6位数字,为了适应代码,所以需要格式化文件名
文件改名脚本:
#!/bin/bash
index=1000000;
for f in *.png;
do
mv "$f" ${index#1}.png
((index++))
done
VOC代码处理的是jpg格式的文件,所以还需要把png转换成jpg:
bash -c 'for image in *.png; do convert "$image" "${image%.png}.jpg"; echo “image $image converted to ${image%.png}.jpg ”; done'
看了pascal_voc.py代码,可以把代码的jpg拼接改成png,这样可以不做上一步.
标记图片:
依然使用labelImg工具,生成对应的xml文件.
然后把
- 所有图片放到/tf-faster-rcnn/data/VOCdevkit2007/VOC2007/JPEGImages中
- 所有
xml文件放到/tf-faster-rcnn/data/VOCdevkit2007/VOC2007/Annotations中 - 修改tf-faster-rcnn/data/VOCdevkit2007/VOC2007/ImageSets/Main下的4个文件:
- test.txt 测试图片名称,数量占总图片数量的50%
- trainval.txt 训练/验证图片名称,数量占总图片数量的50%
- train.txt 训练图片名称,数量占 训练/验证(上一条) 总数的50%
- val.txt 验证图片名称,数量占 训练/验证总数的50%
最后修改tf-faster-rcnn/lib/datasets/pascal_voc.py,把self._classes定义的类别填入我们自己要识别的类别.
到此,自己的VOC数据集就可以使用了.
Faster RCNN使用
安装
git clone https://github.com/endernewton/tf-faster-rcnn.git
修改配置,使支持CPU选项
tf-faster-rcnn/lib/model/nms_wrapper.py 修改/注释以下行:
#from nms.gpu_nms import gpu_nms def nms(dets, thresh, force_cpu=False): #Dispatch to either CPU or GPU NMS implementations. if dets.shape[0] == 0: return [] if cfg.USE_GPU_NMS and not force_cpu: #return gpu_nms(dets, thresh, device_id=cfg.GPU_ID) return cpu_nms(dets, thresh) else: return cpu_nms(dets, thresh)tf-faster-rcnn/lib/model/config.py 修改代码为:
__C.USE_GPU_NMS = Falsetf-faster-rcnn/lib/setup.py 注释下面代码:
CUDA = locate_cuda() self.src_extensions.append('.cu') Extension('nms.gpu_nms', ['nms/nms_kernel.cu', 'nms/gpu_nms.pyx'], library_dirs=[CUDA['lib64']], libraries=['cudart'], language='c++', runtime_library_dirs=[CUDA['lib64']], # this syntax is specific to this build system # we're only going to use certain compiler args with nvcc and not with gcc # the implementation of this trick is in customize_compiler() below extra_compile_args={'gcc': ["-Wno-unused-function"], 'nvcc': ['-arch=sm_52', '--ptxas-options=-v', '-c', '--compiler-options', "'-fPIC'"]}, include_dirs = [numpy_include, CUDA['include']]
编译Cython
cd lib
make clean
make
cd ..
安装python coco API
cd data
git clone https://github.com/pdollar/coco.git
cd coco/PythonAPI
make
cd ../../..
下载预训练模型
./data/scripts/fetch_faster_rcnn_models.sh
测试
./tools/demo.py
会对/data/demo/目录下的图片进行识别.如果需要识别其他图片,修改demo.py里的im_names列表.
训练自己的数据
下载预训练的模型,目前支持VGG16和Resnet V1
mkdir -p data/imagenet_weights
cd data/imagenet_weights
wget -v http://download.tensorflow.org/models/vgg_16_2016_08_28.tar.gz
tar -xzvf vgg_16_2016_08_28.tar.gz
mv vgg_16.ckpt vgg16.ckpt
cd ../..
mkdir -p data/imagenet_weights
cd data/imagenet_weights
wget -v http://download.tensorflow.org/models/resnet_v1_101_2016_08_28.tar.gz
tar -xzvf resnet_v1_101_2016_08_28.tar.gz
mv resnet_v1_101.ckpt res101.ckpt
cd ../..
训练/验证
./experiments/scripts/train_faster_rcnn.sh [GPU_ID] [DATASET] [NET]
# GPU_ID is the GPU you want to test on
# NET in {vgg16, res50, res101, res152} is the network arch to use
# DATASET {pascal_voc, pascal_voc_0712, coco} is defined in train_faster_rcnn.sh
# Examples:
./experiments/scripts/train_faster_rcnn.sh 0 pascal_voc vgg16
./experiments/scripts/train_faster_rcnn.sh 1 coco res101
./experiments/scripts/test_faster_rcnn.sh [GPU_ID] [DATASET] [NET]
# GPU_ID is the GPU you want to test on
# NET in {vgg16, res50, res101, res152} is the network arch to use
# DATASET {pascal_voc, pascal_voc_0712, coco} is defined in test_faster_rcnn.sh
# Examples:
./experiments/scripts/test_faster_rcnn.sh 0 pascal_voc vgg16
./experiments/scripts/test_faster_rcnn.sh 1 coco res101
每次训练前需要把
output和data/cache目录删掉
训练次数需要在训练脚本中修改.它的训练不像SSD可以随时中断,然后接着之前的训练.需要一次训练好才会生成文件.后续需要添加
训练次数在
train_faster_rcnn.sh 和test_faster_rcnn.sh 里修改
错误解决:
运行test的时候:
Loading model check point from output/vgg16/voc_2007_trainval/default/vgg16_faster_rcnn_iter_300.ckpt
Loaded.
im_detect: 1/35 3.285s 0.001s
im_detect: 2/35 3.619s 0.001s
im_detect: 3/35 4.309s 0.001s
im_detect: 4/35 4.635s 0.001s
im_detect: 5/35 4.665s 0.001s
im_detect: 6/35 4.506s 0.001s
im_detect: 7/35 4.561s 0.001s
im_detect: 8/35 4.713s 0.001s
im_detect: 9/35 4.776s 0.001s
im_detect: 10/35 4.750s 0.001s
im_detect: 11/35 4.787s 0.001s
im_detect: 12/35 4.761s 0.001s
im_detect: 13/35 4.721s 0.001s
im_detect: 14/35 4.788s 0.001s
im_detect: 15/35 4.861s 0.001s
im_detect: 16/35 4.831s 0.001s
im_detect: 17/35 4.848s 0.001s
im_detect: 18/35 4.793s 0.001s
im_detect: 19/35 4.791s 0.001s
im_detect: 20/35 4.755s 0.001s
im_detect: 21/35 4.795s 0.001s
im_detect: 22/35 4.829s 0.001s
im_detect: 23/35 4.873s 0.001s
im_detect: 24/35 4.822s 0.001s
im_detect: 25/35 4.801s 0.001s
im_detect: 26/35 4.787s 0.001s
im_detect: 27/35 4.758s 0.001s
im_detect: 28/35 4.747s 0.001s
im_detect: 29/35 4.717s 0.001s
im_detect: 30/35 4.725s 0.001s
im_detect: 31/35 4.754s 0.001s
im_detect: 32/35 4.759s 0.001s
im_detect: 33/35 4.775s 0.001s
im_detect: 34/35 4.786s 0.001s
im_detect: 35/35 4.796s 0.001s
Evaluating detections
Writing knife VOC results file
Traceback (most recent call last):
File "./tools/test_net.py", line 120, in
test_net(sess, net, imdb, filename, max_per_image=args.max_per_image)
File "/home/wow/Github/tf-faster-rcnn/tools/../lib/model/test.py", line 192, in test_net
imdb.evaluate_detections(all_boxes, output_dir)
File "/home/wow/Github/tf-faster-rcnn/tools/../lib/datasets/pascal_voc.py", line 278, in evaluate_detections
self._write_voc_results_file(all_boxes)
File "/home/wow/Github/tf-faster-rcnn/tools/../lib/datasets/pascal_voc.py", line 205, in _write_voc_results_file
with open(filename, 'wt') as f:
IOError: [Errno 2] No such file or directory: '/home/wow/Github/tf-faster-rcnn/data/VOCdevkit2007/results/VOC2007/Main/comp4_c412401c-8e25-43ab-ba7f-e65b9089897c_det_test_knife.txt'
Command exited with non-zero status 1
590.78user 10.24system 2:51.10elapsed 351%CPU (0avgtext+0avgdata 2778584maxresident)k
查看pascal_voc.py代码,发现这里with open(filename, 'wt') as f代码应该会创建文件,只是因为没有results目录,所以手动创建tf-faster-rcnn/data/VOCdevkit2007/results/VOC2007/Main/目录即可.
测试通过后进行对图片的识别,识别依然使用./tools/demo.py:
./tools/demo.py --net=vgg16 --dataset=pascal_voc
得到如下错误:
2017-12-26 17:04:58.521558: I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX
begin to restore
Traceback (most recent call last):
File "./tools/demo.py", line 141, in
saver.restore(sess, tfmodel)
File "/home/wow/ML/tensorflow/local/lib/python2.7/site-packages/tensorflow/python/training/saver.py", line 1666, in restore
{self.saver_def.filename_tensor_name: save_path})
File "/home/wow/ML/tensorflow/local/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 889, in run
run_metadata_ptr)
File "/home/wow/ML/tensorflow/local/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 1120, in _run
feed_dict_tensor, options, run_metadata)
File "/home/wow/ML/tensorflow/local/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 1317, in _do_run
options, run_metadata)
File "/home/wow/ML/tensorflow/local/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 1336, in _do_call
raise type(e)(node_def, op, message)
tensorflow.python.framework.errors_impl.InvalidArgumentError: Assign requires shapes of both tensors to match. lhs shape= [84] rhs shape= [16]
[[Node: save/Assign = Assign[T=DT_FLOAT, _class=["loc:@vgg_16/bbox_pred/biases"], use_locking=true, validate_shape=true, _device="/job:localhost/replica:0/task:0/device:CPU:0"](vgg_16/bbox_pred/biases, save/RestoreV2)]]
Caused by op u'save/Assign', defined at:
File "./tools/demo.py", line 139, in
saver = tf.train.Saver()
File "/home/wow/ML/tensorflow/local/lib/python2.7/site-packages/tensorflow/python/training/saver.py", line 1218, in __init__
self.build()
File "/home/wow/ML/tensorflow/local/lib/python2.7/site-packages/tensorflow/python/training/saver.py", line 1227, in build
self._build(self._filename, build_save=True, build_restore=True)
File "/home/wow/ML/tensorflow/local/lib/python2.7/site-packages/tensorflow/python/training/saver.py", line 1263, in _build
build_save=build_save, build_restore=build_restore)
File "/home/wow/ML/tensorflow/local/lib/python2.7/site-packages/tensorflow/python/training/saver.py", line 751, in _build_internal
restore_sequentially, reshape)
File "/home/wow/ML/tensorflow/local/lib/python2.7/site-packages/tensorflow/python/training/saver.py", line 439, in _AddRestoreOps
assign_ops.append(saveable.restore(tensors, shapes))
File "/home/wow/ML/tensorflow/local/lib/python2.7/site-packages/tensorflow/python/training/saver.py", line 160, in restore
self.op.get_shape().is_fully_defined())
File "/home/wow/ML/tensorflow/local/lib/python2.7/site-packages/tensorflow/python/ops/state_ops.py", line 276, in assign
validate_shape=validate_shape)
File "/home/wow/ML/tensorflow/local/lib/python2.7/site-packages/tensorflow/python/ops/gen_state_ops.py", line 57, in assign
use_locking=use_locking, name=name)
File "/home/wow/ML/tensorflow/local/lib/python2.7/site-packages/tensorflow/python/framework/op_def_library.py", line 787, in _apply_op_helper
op_def=op_def)
File "/home/wow/ML/tensorflow/local/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 2956, in create_op
op_def=op_def)
File "/home/wow/ML/tensorflow/local/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 1470, in __init__
self._traceback = self._graph._extract_stack() # pylint: disable=protected-access
InvalidArgumentError (see above for traceback): Assign requires shapes of both tensors to match. lhs shape= [84] rhs shape= [16]
[[Node: save/Assign = Assign[T=DT_FLOAT, _class=["loc:@vgg_16/bbox_pred/biases"], use_locking=true, validate_shape=true, _device="/job:localhost/replica:0/task:0/device:CPU:0"](vgg_16/bbox_pred/biases, save/RestoreV2)]]
这个错误推测是之前训练的cache没有清空导致模型数据不匹配.删除/data/cache和/output,重新训练.
又遇到类似的错误:
tensorflow.python.framework.errors_impl.InvalidArgumentError: Assign requires shapes of both tensors to match. lhs shape= [4096,21] rhs shape= [4096,4]
这次的shape和上一次的不一样.通过阅读demo.py代码,发现这一行定义了21:
net.create_architecture("TEST", 21,
tag='default', anchor_scales=[8, 16, 32])
把21改成4,成功运行.
Reference:
![[Tensorflow] Tensorflow卷积理解](/medias/featureimages/23.jpg)
![[CS]聊聊字符编码](/medias/featureimages/20.jpg)