py-faster-rcnn：windows10 + vs2013 + caffe(python版)

一. 准备工作

1. 开发环境

• Windows10 x64
• VS2013
• CUDA 8.0 for windows10
• cuDNN 5 for windows10
• Anaconda3(python2.7)
• MS Caffe
• py-faster-rcnn
• py-faster-rcnn-windows

2. 版本说明

• Visual Studio: 因项目开发需求笔者所用电脑装有vs2013和vs2015两个版本(vs2015为后安装的，打开工程时只能用vs2015打开)，MS Caffe编译必须采用vs2013。在打开caffe.sln时，会有图1提示，选择"取消"即可，打开后如图2，即虽然用的是vs2015，但工程配置仍为2013版.
  

• Anaconda：caffe支持python2.7。笔者安装的是Anaconda3(默认python3.7)，因此需要在Anaconda Navigator中创建新环境。如图3-4，创建的新环境基于python2.7，名为"caffe"，后续配置python路径需要配置该环境下python2.7的路径。
  

二. caffe编译

下载MS Caffe，将caffe-master/windows/CommonSettings.props.example复制一份，存为caffe-master/windows/CommonSettings.prop。打开CommonSettings.props进行配置，如图2-1，注意修改CuDNN和Python2.7的路径。

配置完成后，用vs打开caffe-master/windows/Caffe.sln。由于faster-rcnn中使用了roi-pooling-layer层，而微软版本编译时未添加roi_pooling_layer，因此需要将roi_pooling_layer的.cu/.hpp/.cpp分别添加libcaffe项目的cu/include/src文件下的layer中，如图2-2。
添加完成后，配置为Release模式先编译libcaffe，成功后再编译整个解决方案即可。

三. Faster R-CNN配置

1. win10下配置py-faster-rcnn

(1)下载py-faster-rcnn和py-faster-rcnn-windows，将后者中的文件复制到py-faster-rcnn中替换。

修改1：lib/rpn/proposal_layer.py line 64:
修改前：cfg_key = str(self.phase)
修改后：cfg_key = str('TRAIN' if self.phase == 0 else 'TEST')

修改2：将以下文件中的“param_str_”替换为“param_str”
lib/rpn/anchor_target_layer.py line 27
lib/rpn/proposal_layer.py line 26
lib/rpn/proposal_target_layer.py line 25
lib/roi_data_layer/layer.py line 87

(2)将编译好的Caffe根目录下 caffe-master\Build\x64\Release\pycaffe中的caffe文件夹复制到py-faster-rcnn根目录py-faster-rcnn\caffe-fast-rcnn\python中，如图
(3)修改py-faster-rcnn根目录下py-faster-rcnn/lib/setup_cuda.py line 33，配置为本机的CUDA/include路径。

2. 运行demo

(1)运行py-faster-rcnn/data/scripts下的fetch_faster_rcnn_models.sh获取model，或者直接从网站下载faster_rcnn_models.tgz，下载后解压到py-faster-rcnn/data下，如图

(2)打开Anaconda Prompt，激活caffe环境，切换到py-faster-rcnn根目录下，执行

python ./tools/demo.py

具体操作如图

运行结果如图

四. 问题汇总

错误1：db_lmdb.hpp Check failed: mdb_status == 0 (3 vs. 0) 系统找不到指定的路径
解决：路径问题，mnist数据集下载后压缩包名称和解压缩后的文件名称有差别。如压缩包名：train-images-idx3-ubyte.gz，内部文件名：train-images.idx3-ubyte。数据转换时数据集名称一定是解压缩后内部的文件名。

错误2：Microsoft\Visual C++ for Python\9.0\VC\Bin\amd64\cl.exe’ failed with exit status 2
解决：下载并安装VCForPython27 Microsoft Visual C++ Compiler for Python 2.7

错误3：安装VCForPython27后，Cannot open include file: ‘stdbool.h’
解决：环境变量添加 VS90COMNTOOLS，并将其指给 D:\Program Files\Microsoft Visual Studio 12.0\Common7\Tools

错误4：layer_factory.hpp Check failed: registry.count(type) == 1 (0 vs. 1) Unknown layer type: ROIPooling
解决：faster-rcnn中使用了roi-pooling-layer层，而微软版本caffe编译时未添加roi_pooling_layer，需要将roi_pooling_layer的.cu/.hpp/.cpp分别添加libcaffe项目的cu/include/src文件下的layer中。三个文件都要添加！如果未添加.hpp，亲测也会出现上述错误。
错误5：IndentationError
解决：python代码对齐问题。

错误6： No module named xxx
解决：安装相应包，pip install xxx

错误7：has no attribute ‘param_str_’
解决：将’param_str_‘更改为’param_str’，'param_str_'存在于以下4个位置：lib/rpn/anchor_target_layer.py、lib/rpn/proposal_layer.py、lib/rpn/proposal_target_layer.py、lib/roi_data_layer/layer.py

错误8：TypeError: slice indices must be integers or None or have an _ index_ method

Solving...
Traceback (most recent call last):
  File "./tools/train_net.py", line 140, in 
    max_iters=args.max_iters)
  File "G:\DL\py-faster-rcnn-sp\FRCNN\tools\..\lib\fast_rcnn\train.py", line 160, in train_net
    model_paths = sw.train_model(max_iters)
  File "G:\DL\py-faster-rcnn-sp\FRCNN\tools\..\lib\fast_rcnn\train.py", line 101, in train_model
    self.solver.step(1)
  File "G:\DL\py-faster-rcnn-sp\FRCNN\tools\..\lib\rpn\proposal_target_layer.py", line 67, in forward
    rois_per_image, self._num_classes)
  File "G:\DL\py-faster-rcnn-sp\FRCNN\tools\..\lib\rpn\proposal_target_layer.py", line 192, in _sample_rois
    _get_bbox_regression_labels(bbox_target_data, num_classes)
  File "G:\DL\py-faster-rcnn-sp\FRCNN\tools\..\lib\rpn\proposal_target_layer.py", line 128, in _get_bbox_regression_labels
    bbox_targets[ind, start:end] = bbox_target_data[ind, 1:]
TypeError: slice indices must be integers or None or have an __index__ method

解决：修改lib/rpn/proposal_target_layer.py line 123

更改前：
    for ind in inds:
        cls = clss[ind]
        start = 4 * cls
        end = start + 4
        bbox_targets[ind, start:end] = bbox_target_data[ind, 1:]
        bbox_inside_weights[ind, start:end] = cfg.TRAIN.BBOX_INSIDE_WEIGHTS
    return bbox_targets, bbox_inside_weights
    
更改后：
    for ind in inds:
        ind = int(ind)  #cwr
        cls = clss[ind]
        start = int(4 * cls)  #cwr
        end = int(start + 4)  #cwr
        bbox_targets[ind, start:end] = bbox_target_data[ind, 1:]
        bbox_inside_weights[ind, start:end] = cfg.TRAIN.BBOX_INSIDE_WEIGHTS
    return bbox_targets, bbox_inside_weights

错误9：训练过程中出现CURAND_STATUS_LAUNCH_FAILURE，如下

Solving...
F0509 13:49:04.570956 14628 math_functions.cu:375] Check failed: status == CURAND_STATUS_SUCCESS (201 vs. 0)  CURAND_STATUS_LAUNCH_FAILURE
*** Check failure stack trace: ***

解决：笔者最初环境配置为win10 + CUDA7.5，出现该问题，看有人说是CUDA7.5对于10系列的显卡支持不是很完善。尝试用CUDA8.0 for win10 + cudnn 5.0 for win10，重新编译caffe，配置后测试，问题解决。

错误10：TypeError: ‘numpy.float64’ object cannot be interpreted as an index

Traceback (most recent call last):
  File "./tools/train_net.py", line 140, in 
    max_iters=args.max_iters)
  File "G:\DL\py-faster-rcnn-sp\FRCNN\tools\..\lib\fast_rcnn\train.py", line 160, in train_net
    model_paths = sw.train_model(max_iters)
  File "G:\DL\py-faster-rcnn-sp\FRCNN\tools\..\lib\fast_rcnn\train.py", line 101, in train_model
    self.solver.step(1)
  File "G:\DL\py-faster-rcnn-sp\FRCNN\tools\..\lib\rpn\proposal_target_layer.py", line 67, in forward
    rois_per_image, self._num_classes)
  File "G:\DL\py-faster-rcnn-sp\FRCNN\tools\..\lib\rpn\proposal_target_layer.py", line 168, in _sample_rois
    fg_inds = npr.choice(fg_inds, size=fg_rois_per_this_image, replace=False)
  File "mtrand.pyx", line 1197, in mtrand.RandomState.choice
TypeError: 'numpy.float64' object cannot be interpreted as an index

解决：numpy1.12及以后版本删除了使用浮点数索引的功能，一种解决方案是采用低版本numpy(numpy 1.11.0)，另一种解决方案是通过添加astype(np.int)将产生的float型数据强转为int。第一种方案笔者没测过，这里采用第二种解决方案：共有4处需修改

lib/roi_data_layer/minibatch.py line 26
更改前：fg_rois_per_image = np.round(cfg.TRAIN.FG_FRACTION * rois_per_image)
更改后：fg_rois_per_image = np.round(cfg.TRAIN.FG_FRACTION * rois_per_image).astype(np.int)

lib/datasets/ds_utils.py line 12 
更改后：hashes = np.round(boxes * scale).dot(v).astype(np.int)

lib/fast_rcnn/test.py line 129
更改后：hashes = np.round(blobs['rois'] * cfg.DEDUP_BOXES).dot(v).astype(np.int)

lib/rpn/proposal_target_layer.py line 60
更改后：fg_rois_per_image = np.round(cfg.TRAIN.FG_FRACTION * rois_per_image).astype(np.int)

---

参考资料：

1. MS Caffe
2. 官方Caffe-windows 配置与示例运行
3. Caffe: Faster-RCNN Python版本配置 (Windows)
4. TypeError: ‘numpy.float64’ object cannot be interpreted as an index