Learning in CPU

[marifnst]

dear all

i try to implement faster r cnn in CPU mode. actually, faster r cnn officially doesnt provide training in CPU mode.

so, i made changes:

• roi_pooling_layer.cpp -->conversion based on roi_pooling_layer.cu
• smooth_L1_loss_layer.cpp --> conversion based on smooth_L1_loss_layer.cu

it's compiled successfully and training can be run smoothly using ZF network (my training process still in progress right now). anybody can help to give suggestion about my changes on attached code ? is my code conversion correct ?

regards muhammad arifn nasution

changes code.zip

[Austriker]

@marifnst Have a look at this PR. Some people are also trying to implement de CPU mode.

[marifnst]

@Austriker thank you for your information. i will visit the thread. my code running well and i intend to compare learning result in GPU mode. because i intend to train faster r cnn in very low resource (ex mobile GPU) will be shared soon.

[saiprabhakar]

@Austriker @marifnst I implemented the code based on GPU code of smooth l1 loss, but the gpu code doesnt take weights into consideration in its back propagation. Is this a mistake or am I missing something?

[indsak]

@marifnst Can you please tell me what all to be done after changing the files for training FRCNN in CPU mode? Do i need to change the files or is it possible by giving any argument in the train option

[abhigoku10]

@marifnst hi tq for sharing the .cpp files is it possible to share the .cu files also since when i replace the existing files with urs i am ot able to get any changes

[sdews]

@abhigoku10 After replaceing the existing files with the above two .cpp files , you should cd $FRCN_ROOT/caffe-fast-rcnn sudo make clean sudo make -j $(($(nproc) + 1)) && make pycaffe

[Kshitij-Parashar]

Hi @marifnst, I followed all the instructions and replaced both the files but could not get train in cpu mode. Please help me, I would be highly greatful!