2 tests from HardSpatialTransformerLayerTest/1, where TypeParam = caffe::CPUDevice<double>时发生错误 about rmpe

[Fang-Haoshu]

Hi @siyilingting

Yes, as the info said: "Check failed: false Don't use the CPU implementation! If you really want to, delete the CHECK in st_layer.cpp file. Line number: 242-243. De_transform is not supported."

The STN layer doesn't support CPU at present, so you can try to make it with GPU.

BTW, usually I don't run 'make runtest', so very THANK YOU for telling me the issue about fish-bike,jpg

[siyilingting]

Thank you for your advice and I will try it again, ^_^

[siyilingting]

Hello, @Fang-Haoshu
There are two problems that puzzles me.

1. When I run the command ./data/MPII_COCO14/create_data.sh in the Train/Eval, I find the code searches datas in "/RMPE/data/MPII_COCO14/JPEGImages" instead of "/data/MPII_COCO14/images" . After changing the name, the code runs successfully.
2. But next when I run python examples/rmpe/ssd_pascal_MPII_COCO14VGG.py, it showed me these:
   F0112 19:03:33.095942 10139 io.cpp:63] Check failed: fd != -1 (-1 vs. -1) File not found: models/VGG_SSD/VGG_ILSVRC_16_layers_fc_reduced.caffemodel
   *** Check failure stack trace: ***
   @ 0x7f1b98426e6d (unknown)
   @ 0x7f1b98428ced (unknown)
   @ 0x7f1b98426a5c (unknown)
   @ 0x7f1b9842963e (unknown)
   @ 0x7f1b9de1e2d3 caffe::ReadProtoFromBinaryFile()
   @ 0x7f1b9de49134 caffe::ReadNetParamsFromBinaryFileOrDie()
   @ 0x7f1b9ddc30d7 caffe::Net<>::CopyTrainedLayersFromBinaryProto()
   @ 0x7f1b9ddc3146 caffe::Net<>::CopyTrainedLayersFrom()
   @ 0x408d45 CopyLayers()
   @ 0x40964c train()
   @ 0x406b9c main
   @ 0x7f1b901b9b15 __libc_start_main
   @ 0x4074a5 (unknown)

And I didn't find the file " VGG_ILSVRC_16_layers_fc_reduced.caffemodel " on my computer.
Could you give me some advice? Thank you very much!

[Fang-Haoshu]

Hi, thanks for your report!

For training human detector, you need to download the models & get more details at SSD's official repo:https://github.com/weiliu89/caffe/tree/ssd

Best

[siyilingting]

Hi, thank you for your advice!
But there is another problem. When I run the command ./data/MPII_COCO14/create_data.sh , it tells me that data_transformer.cpp:444] Check failed: channels == img_channels (3 vs. 1). So I have found the images with one channel and delete them in /home/wbr/RMPE/data/MPII_COCO14/train.txt . Now it seems running correctly and it is training human detector. Is it right to delete data? I really don't understand.
Thank you and looking forward to your responds.

Best wishes !

[Fang-Haoshu]

It's not a big problem.

I will try to figure out why. Before that, you can train with your current setting

Best

[Fang-Haoshu]

should be fixed now

Hi! I am facing the same problem
2 tests from HardSpatialTransformerLayerTest/1, where TypeParam = caffe::CPUDevice
But it seems that delete line 242-243 in st_layer.cpp doesn't work.
I am wondering how you finally solve this. @siyilingting
Thank you!