deprecated API in class roi_proposal()?

[chakpongchung]

roi_proposal_net = roi_proposal(rpn_net, gt_boxes, im_dims, eval_mode, flags)

roi_proposal_net = roi_proposal(rpn_net, gt_boxes, im_dims, eval_mode)

It seems like you have changed these functions. Would you be willing to share the update? from Networks.faster_rcnn_networks import rpn, roi_proposal, fast_rcnn

[kevinjliang]

Hi! I believe the changes to faster_rcnn_networks (including roi_proposal) have already been pushed. We're trying to move away from using the 'flags' dict as much as possible.

[chakpongchung]

well, I tried to git clone recursively again:

rpn_net = rpn(featureMaps,gt_boxes,im_dims,_feat_stride,eval_mode,flags)

TypeError Traceback (most recent call last)

in () ----> 1 rpn_net = rpn(featureMaps,gt_boxes,im_dims,_feat_stride,eval_mode,flags) TypeError: __init__() takes 6 positional arguments but 7 were given

[kevinjliang]

Ah, okay. The Development/ folder mostly contains experimental code and testing for debugging. We don't actively maintain any of that code.

I can update that notebook when I get the chance though, if that's of interest.

[chakpongchung]

Thank you for your quick response.

I tried to follow the demo page too. Where is the faster_rcnn_clutter.py now? python faster_rcnn_clutter.py -n [Model num, ex 1] -e [Num of epochs, ex 3]

[kevinjliang]

It's been removed. I actually just pushed a rather hefty number of updates yesterday; I was planning on updating the README once I take care of a few other things and hoping no one tried to run the demo until then, but it looks like you beat me to it.

Try running (from Models/): python faster_rcnn_conv5.py -n 1 -e 5 -i 1

Might take a few hours, but it'll save a bunch of PNGs of the results at Data/clutteredMNIST/Outputs/

[kevinjliang]

Still have some instructions I need to add, but for the purposes of the demo, the README has been updated. I've also fixed the iPython notebook you were asking about.

[chakpongchung]

Hi Kevin,

Thank you for your update. Just as feedback, now the notebook crashed at

fast_rcnn_net.get_bbox_refinement().get_shape() python faster_rcnn_conv5.py -n 1 -e 5 -i 1

is working. I have two 1080TI, can I use two gpu to train it?

[kevinjliang]

Notebook appears to be working on my end. Are you talking about line 25?

I believe if you include the flag -g 2 when you run the model, it should use both your GPUs. @dancsalo, confirm?

[chakpongchung]

yes, this line:

fast_rcnn_net.get_bbox_refinement().get_shape()