CPU Training error: smooth_L1_loss_layer Not Implemented Yet

[sunshineatnoon]

I modified train_net.py to use cpu for training. I also disabled MATLAB and use the pre-computed selective search bounding box for VOC2007. However, I got this error during training:

Solving...
F0906 14:48:34.166416 13116 smooth_L1_loss_layer.cpp:39] Not Implemented Yet
*** Check failure stack trace: ***

I looked into fast-rcnn/caffe-fast-rcnn/src/caffe/layers/smooth_L1_loss_layer.cpp and found following code:

template <typename Dtype>
void SmoothL1LossLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
    const vector<Blob<Dtype>*>& top) {
  NOT_IMPLEMENTED;
}

template <typename Dtype>
void SmoothL1LossLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,
    const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
  NOT_IMPLEMENTED;
}

So I guess this layer is only implemented for GPU, not for CPU. Does this mean I can't use CPU to train fast-rcnn?

[kaifeichen]

+1

[ghost]

@sunshineatnoon @kaifeichen @rbgirshick I got the same problem. Did you solve this problem ???

[sunshineatnoon]

@haimansx No, I have not solved the problem, I think the only way to solve the problem is by implementing the two functions above in c++

[kaifeichen]

Hi @haimansx, I think @sunshineatnoon is right. I just used the GPU version and it worked.

[tanjoreg]

Hi @sunshineatnoon I have followed your blog to train Fast-RCNN my own dataset and i have stumbled upon this error. I am using CPU Mode. Were you able to solve this issue?

@kaifeichen Does that mean Fast-RCNN is not implemented for CPU mode? Any updates on this would help me.

[sunshineatnoon]

@tanjoreg No, I haven't solved this issue. Fast Rcnn can only be trained on GPU as far as I know. But it can be tested on CPU.

[tanjoreg]

@sunshineatnoon Thank you very much,for the swift reply. Your blog helped me a lot. I have tested the imagenet demo on CPU and it works. I was trying to train my own dataset, which resulted in the error. Does that mean I have hit the wall with CPU mode?

[sunshineatnoon]

As I mentioned above, I think the way to solve this issue is to implement the two functions above in c++.

[hnflee]

+1

[athulase]

@sunshineatnoon: Were you able to train R-CNN in cpu? or is it still in the status of not running training on cpu mode?

[HamdiHamed1992]

I am facing the same problem!
Any help?
If someone implemented those 2 functions in cpp, can he post the code?

[PeterJiangwy]

@HamdiHamed1992 , Do you solve this problem. There is someone write these 2 functions, but it still not work.

[HamdiHamed1992]

@PeterJiangwy sorry no. I am using a CPU to train CNN and interested to segmentation not object localisation. Therefore i did not focus alot with RCNN, Fast RCNN and Faster RCNN. I can suggest you to use other network architectures and change just the final classification layers to your wanted task. Choose some already existing models that can run on CPU... (Segnet 4 example, based on VGG16 Model arch). If you need anything, feel free to ask.

[PeterJiangwy]

@ HamdiHamed1992, Thank you. I’m newer to this scope. Now I’m trying to do something about object detection. Do you have any suggestions? Thank you very much for your help.

在 2018年1月16日，18:22，HamdiHamed1992 [email protected] 写道：

@PeterJiangwy https://github.com/peterjiangwy sorry no. I am using a CPU to train CNN and interested to segmentation not object localisation. Therefore i did not focus alot with RCNN, Fast RCNN and Faster RCNN. I can suggest you to use other network architectures and change just the final classification layers to your wanted task. Choose some already existing models that can run on CPU... (Segnet 4 example, based on VGG16 Model arch). If you need anything, feel free to ask.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/rbgirshick/fast-rcnn/issues/53#issuecomment-357916252, or mute the thread https://github.com/notifications/unsubscribe-auth/ARw4KVKdCASfaD5tX2Q8yVBqxn2rzqZzks5tLHhLgaJpZM4F4ju-.