fast-rcnn and faster-rcnn training

[b0unc3]

Hi,

is it possible to fine-tune fast/faster-rcnn models on digits?
Thanks.

[lukeyeager]

I haven't done it before. Do you have an implementation for Caffe or Torch already that you'd like to use?

[b0unc3]

I've been able to train a model using fast-rcnn (with the caffe-fast-rcnn implementation), but I would like to use digits to simplify and automate the training phase.

[lukeyeager]

DIGITS requires NVIDIA's fork of Caffe, which was last synced with BVLC/caffe on 09/29/2015 (https://github.com/NVIDIA/caffe/releases).

It looks like caffe-fast-rcnn is a different fork of caffe - so DIGITS will reject it.

If Caffe commits to a versioning scheme (BVLC/caffe#3311) which denotes a standardized API, DIGITS will be able to accept a much broader range of Caffe builds - possibly including the caffe-fast-rcnn fork.

Alternatively, if the code you need gets merged into BVLC's fork, then it will get picked up in NVIDIA's fork for our next update.

[marvision-ai]

Hi there, revisiting this. Is there a tutorial or at least some steps to get FasterRCNN working with digits? I have yet to see anything that covers beyond detectron.