Open zimenglan-sysu-512 opened 5 years ago
Hi,
torchvision 0.3
models for object detection are not compatible with maskrcnn-benchmark.
I did a number of simplifications, like changing the anchors, using torchvision resnets and a few other changes.
What do you have in mind wrt supporting torchvision 0.3 models in maskrcnn-benchmark
?
You can override build_detection_model
to return one of the models from torchvision, right?
Hi @fmassa, What about detection specific ops, are they different than what we have here for roi_pooling and iou?
@botcs they are mostly the same, so those could be reused indeed.
@fmassa could it possibly mean that we wouldn't need to compile anything of this lib to train, or are there any more specific C/C++/CUDA codes here? If the first case is true, than maybe it could convince some to use this lib - and I would be happy to see what can I replace
There are new pieces of code that have been recently added, like deformable convs, which are not present in torchvision, so not everything can be replaced as of now
❓ Questions and Help
hi @fmassa since the torchvision 0.3 has been released, do u have a plan to support it in this repo?