Hi there, thanks for the awesome work, I use torchvision a lot :)
Was just wondering if there are any plans to make a new release and push it to PyPi soon?
It's been 8 months since the last one and I feel like there have been quite a few improvements on the master branch.
Cheers!
Yeah, and there are a lot of new models, any plan pretrained models?
@fmassa this is also a good time to consider merging in the layers branch and giving everyone access to the ROI* layers.
Hi,
Yes, we will be cutting a new release soon, it's been a while!
@varunagrawal I'm thinking about making the layers be part of 0.3, which would be the next release of torchvision.
@fmassa oh, it means that layers branch will be merged into master in nearest release?
It would be really great for any object detection task! Will wait :)
Yes, we plan on having other domains as well outside of image classification for 0.3.
It would be nice if the GPU version of NMS from maskrcnn-benchmark gets added as well.
@TheCodez nms is already in layers, isn't it?
@Oktai15 only the CPU version is, the GPU version is not implemented yet.
There is a CUDA implementation of nms in maskrcnn-benchmark that I plan to add to torchvision as well.
@fmassa I can add the GPU nms over the weekend. In the meantime, can we work out a plan for the new release? Right now, the timelines seem kind of vague and that seems to be rustling feathers.
@varunagrawal I'm writing a plan for v0.3 as we speak, and I'll share it here once it's ready.
About the GPU nms, we have it implemented in maskrcnn-benchmark, so I will just port it over from there I believe
@fmassa @TheCodez GPU version of NMS has been added to layers via #727
I rebased the latest master so you'll see merge conflicts with vision/layers but this would make merge to master a lot easier for you.
Another vote for a new release, currently getting:
/lib/python3.7/site-packages/torchvision/models/squeezenet.py:94: UserWarning: nn.init.kaiming_uniform is now deprecated in favor of nn.init.kaiming_uniform_.
init.kaiming_uniform(m.weight.data)
/lib/python3.7/site-packages/torchvision/models/squeezenet.py:92: UserWarning: nn.init.normal is now deprecated in favor of nn.init.normal_.
init.normal(m.weight.data, mean=0.0, std=0.01)
which is fixed in master. Please and thank you!
@stas00 sure! I'm writing the release notes for the new release, sorry for the delay!
@timonbimon this can be closed with the release of 0.2.3.
Most helpful comment
@fmassa this is also a good time to consider merging in the
layersbranch and giving everyone access to theROI*layers.