search

weiliu89 / caffe

Caffe: a fast open framework for deep learning.
http://caffe.berkeleyvision.org/
Other
4.77k stars 1.67k forks source link

Porting to Caffe2 #561

Open mjohnst opened 7 years ago

mjohnst commented 7 years ago

I'm interested in porting the SSD project to Caffe2. I realize there's a lot of code that needs to be re-written for Caffe2, but if it's not too much trouble, I'd appreciate a high-level list of what would need to be added to support training (or possibly also converting) a SSD model in Caffe2.

From the commit history, it looks like you've added these custom layers (and added/changed a variety of utils and tools):

annotated_data_layer
detection_evaluate_layer
detection_output_layer
multibox_loss_layer
normalize_layer
permute_layer
prior_box_layer
smooth_L1_loss_layer
video_data_layer

I can likely re-implement the layers in Caffe2, are there any other important components needed to support SSD models?

weiliu89 commented 7 years ago

These seem to be complete. Of course, there are many utility functions (e.g. bbox_utils.cpp etc.)

mjohnst commented 7 years ago

Thanks Wei. I left out the tools/utils and stuff because there're are a lot :)

wjxiz1992 commented 7 years ago

@mjohnst Hi, What's your progress now? I would like to port it to Caffe2 too. Which step have you reached?

mjohnst commented 7 years ago

Hi @wjxiz1992 I'm no longer working on the project that required this conversion, so I haven't made any progress. Sorry!