win10, cuda 9.0, python 3.5环境下复现 Flow-Guided Feature Aggregation for Video Object Detection 问题总结

demo

1. error:Microsoft Visual C++ is required
   参考链接，在vs2017中配置vc++ 14

2. LINK : fatal error LNK1158: cannot run ‘rc.exe’
   参考链接 将rc.exe、rcdll.dll从C:\Program Files (x86)\Windows Kits\8.1\bin\x64复制到C:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\bin

3. ‘dict’ object has no attribute ‘iteritems’
   对应文件的dict.iteritems改为dict.items （python 2 与 python 3的区别）
   源代码使用python 2编写的，因此许多代码需要改动，主要为print、dict、pickle等函数。

4. LINK : fatal error LNK1181: 无法打开输入文件“ID=2.obj”

   C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.0\bin\nvcc.exe --compiler-bindir “C:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\bin” --shared “-LD:\Program Files\Anaconda3\libs” “-LD:\Program Files\Anaconda3\PCbuild\amd64” “-LC:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\LIB\amd64” “-LC:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\ATLMFC\LIB\amd64” “-LC:\Program Files (x86)\Windows Kits\10\lib\10.0.17134.0\ucrt\x64” “-LC:\Program Files (x86)\Windows Kits\NETFXSDK\4.6.1\lib\um\x64” “-LC:\Program Files (x86)\Windows Kits\10\lib\10.0.17134.0\um\x64” -lcublas build\temp.win-amd64-3.5\Release\gpu_nms.obj -o build\lib.win-amd64-3.5\gpu_nms.cp35-win_amd64.pyd --linker-options=/nologo,/INCREMENTAL:NO,/LTCG,/MANIFEST:EMBED,ID=2,/MANIFESTUAC:NO,/IMPLIB:build\temp.win-amd64-3.5\Release\gpu_nms.cp35-win_amd64.lib,/NODEFAULTLIB:libcmt.lib
   gpu_nms.obj
   LINK : fatal error LNK1181: 无法打开输入文件“ID=2.obj”
   error: command ‘C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.0\bin\nvcc.exe’ failed with exit status 2

   有一个曲线救国的方法：点此。出现改行错误后，将"ID=2,"去掉，手动执行该命令。有一个曲线救国的方法：点此。出现改行错误后，将"ID=2,"去掉，手动执行该命令。

5. mxnet.base.NotImplementedForSymbol: Function __iadd__ (namely operator "+=") with arguments (<class 'mxnet.symbol.symbol.Symbol'>, <class 'int'>) is not implemented for Symbol and only available in NDArray.
   szha的回答
   '.\fgfa_rfcn\symbols\resnet_v1_101_flownet_rfcn.py’中的 aggregated_conv_feat += tiled_weight * warp_list[i]改为
   aggregated_conv_feat = aggregated_conv_feat + tiled_weight * warp_list[i]

6. 由于编译mxnet太麻烦，所以直接pip安装0.12.1版本；测试demo.py无问题。

训练

1. fail to open fgfa_rfcn_vid-0002.params

   mxnet.base.MXNetError: [15:13:33] D:\Program Files (x86)\Jenkins\workspace\mxnet\mxnet\dmlc-core\src\io\local_filesys.cc:166: Check failed: allow_null LocalFileSystem: fail to open "./output/fgfa_rfcn/imagenet_vid\resnet_v1_101_flownet_imagenet_vid_rfcn_end2end_ohem\VID_val_videos\DET_train_30classes_VID_train_15frames\fgfa_rfcn_vid-0002.params"

   注释掉experiments/fgfa_rfcn/fgfa_rfcn_end2end_train_test.py 的第20行。painterdrown的回答

2. TypeError: a bytes-like object is required, not ‘str’

   Traceback (most recent call last):
   File “_ctypes/callbacks.c”, line 234, in ‘calling callback function’
   File “D:\Program Files\Anaconda3\lib\site-packages\mxnet\operator.py”, line 621, in creator
   op_prop = prop_cls(**kwargs)
   File “Flow-Guided-Feature-Aggregation/experiments/fgfa_rfcn…\fgfa_rfcn\operator_py\proposal_target.py”, line 94, in init
   self._cfg = pickle.loads(cfg)
   TypeError: a bytes-like object is required, not ‘str’

   先将cfg用pickle.dumps序列化，再用pickle.loads反序列化，思路是没有问题的。但是不知为何，pickle.loads时cfg的格式变成了字符串而不是bytes——将值复制出来却又是bytes。索性将cfg序列化至磁盘，然后从磁盘中读出来。方法如下

   fgfa_rfcn/symbols/resnet_v1_101_flownet_rfcn.py, line 965, 增加

   with open('cfg.pck', 'wb') as fp:
       pickle.dump(cfg, fp)
   

   且原本的970行改为 cfg=None

   fgfa_rfcn\operator_py\proposal_target.py, line 36、line 94, 改为

   with open('cfg.pck', 'rb') as fp:
       self._cfg = pickle.load(fp)
   

   终于，成功进行训练。

   Epoch[0] Batch [100] Speed: 0.44 samples/sec Train-RPNAcc=0.907294, RPNLogLoss=0.296552, RPNL1Loss=0.189087,
   RCNNAcc=0.807704, RCNNLogLoss=1.681007, RCNNL1Loss=0.130364,
   Epoch[0] Batch [200] Speed: 0.44 samples/sec Train-RPNAcc=0.923818, RPNLogLoss=0.242203, RPNL1Loss=0.156525,
   RCNNAcc=0.861396, RCNNLogLoss=1.166992, RCNNL1Loss=0.163834,

测试

尚未进行