RFCN profiling and training on own network

[NM]

Hi,

I am trying to implement the RFCN pipeline using the python fork with Microsoft Caffe (with GPU enabled) with my own small pre-trained network, with three convolutional layers and 2 fully connected layers ( like LeNet)

Question1: 

To baseline and profile i ran the demo_rfcn under RFCN_ROOT/tools (with Resnet-50) and got timings as:

"detection took 0.102s for 300 object proposals" for a test image of 1280x720p . 

I then ran just the Res-Net 50 part in the protoxt using ./build/tools/caffe -time option to see how long Res-net 50 takes . That ouput was "190.91ms for a forward pass"

How is that possible? Is 0.102s not the end-to-end timing (not including reading the i/p and displaying detections+nms). If so what does it represent 

Question2: 

If i want to use my with my own small pre-trained network, with three convolutional layers and 2 fully connected layers ( like LeNet)

Does anyone have any pointers on how i should proceed?

A. If i change the train- prototxt (RFCN with Resnet-50) in the place where the Res-net 50 architecture is described and include my model specs (with elimination of the 2 fullyconnected layers), and start re-train RFCN with VOC 2012+2007 train&val-- is that on the right track? Or will i have to make changes in the latter layers of ROI and/or RPN and/or additional convolutional layers (with dilation) to make it compatible

B. If the above works ---  i am willing to modify y network specs (make it deeper) however i am not sure how to access which layers would have to be added (i.e. Inception modules or/and more on towards Resnet architecture etc) Any pointers on that?

Any help is appreciated!! 

Thank you

[ustbw...@163.com]

hello. how to use the small network instead of Restnet50?

在 2016年11月10日星期四 UTC+8上午10:48:00，NM写道：