Hello
Firstly I really appreciate your work.
I wonder if you have in mind to support image size 512 soon.
Thanks
I hope to start working on that in the next few days, as I should have a little bit of time. I will probably be able to give you a more accurate time frame after I start working on it though, and since this is all done in my free-time I can't guarantee anything, but hopefully in the next couple weeks! I want to knock out all of the additional features on this network so I can get going on DSSD. If you want to help though or have any code that you've worked on, as always feel free to submit a PR.
Hi, has anyone tried to implement SSD 512? Is it a matter of specifying the correct config, or some other changes are required?
@lopuhin Any chance you'd be willing to put out a PR including the 512 implementation? Thanks for implementing it!
@amdegroot sure, it would be awesome! I also moved code around a bit to make it pip-installable, I hope this is ok? Right now it needs a couple of fixes to make it ready for PR - bring back pytorch 0.2.0 support and fix VOC training example, I hope I'll be able to fix this issues in the next few weeks.
@lopuhin
Hi, I copied your code from your ssd.pytorch and has tried to reproduce the result. However, after setting every parameters, the out=net(images) in train.py can not move forward any more. This leads to training difficulties, could you help me out? any advice? thank you.
https://github.com/qijiezhao/pytorch-ssd
Anyone that is urgently requiring a 512 version of ssd is welcomed to visit my repo.
(I have reproduced the proposed results of the raw paper.)
I support args.dim for the convenience for adjusting the input size.
Then, Dataloader of KITTI is also supported.
However, some details may cause errors are still under fixing, please feel free to let me know.
@qijiezhao nicely done! Could you plz share your SSD512 weights? It seems not available in your repo.
Hopefully I implemented SSD 512 in lopuhin@600e711 lopuhin@cfa226c and lopuhin@3be5e7f
I can鈥檛 even believe that the 512 input size can be rebuilt so easily.
Most helpful comment
Hopefully I implemented SSD 512 in https://github.com/lopuhin/ssd.pytorch/commit/600e7112461247ed3a5b3a01b8792eeddd558f6b https://github.com/lopuhin/ssd.pytorch/commit/cfa226c797f4c80a3bc18a76f8213d9948fe3708 and https://github.com/lopuhin/ssd.pytorch/commit/3be5e7fb825bd28157ab360b24ac5663fa4bd0c6