Google SSD on Caffe Windows

[michael ben ami]

Hi everyone.. 

i'm trying to run and use google SSD(https://github.com/weiliu89/caffe/tree/ssd ) code on caffe but in windows 

has someone already done something like that ? 

is all the python API / dependence are the same between the two versions of caffe (Linux/Win) ? 

Thanks 

Michael

[Daniel Moodie]

Are you trying to train or do inference?

This branch I believe is just ssd + windows

https://github.com/dtmoodie/caffe/tree/merge_ssd

This branch is ssd + nvcaffe + windows + optional boost logging backend

https://github.com/dtmoodie/caffe/tree/boostlog

There is a bug with training ssd that exists even in the author's original code that causes a deadlock while training on many gpus.

Furthermore, training on windows is not recommended because you'll run into many bugs.  For example, lmdb database size cannot exceed 4GB on windows.

[Daniel Moodie]

I must correct myself.

There is a bug in my branch that causes a deadlock while training on many GPUs.  This may be due to the merging of nvcaffe and ssd.  Not sure as to the cause of the bug.  Training is still not recommended on windows because of limitation of lmdb size.

[michael ben ami]

Thanks Daniel i'll check this.. 

did you try to train in the original repo and use the Net in yours ? 

[Daniel Moodie]

A bit of an update on these branches:

dtmoodie/caffe/tree/merge 

- Training works on Linux, very slow training compared to the next branch.

- Inference works on Windows / Linux

Training should work on Windows with this repo, however on Windows you cannot create an LMDB file greater than 4GB, so I think I only trained one model on Windows then I started just training on Linux.  I have used this repo for both Windows and Linux training and I can confirm it creates a good model.

dtmoodie/caffe/tree/boostlog 

- Training works on Linux, as long as you don't do any testing during training.  Much faster due to merging with nvidia/caffe.  Can achieve about 60% utilization of 8 titan X pascals.

- I've verified that training successfully converges, however I cannot perform testing while solving the model, there is a deadlock somewhere that causes training to freeze after testing completes.

- Inference does not work on windows, works on Linux. (https://github.com/NVIDIA/caffe/issues/284)

Training should also work from this repo but since I cannot comment on the issues with inference, I cannot be certain of if it will be successful.

There might be an intermediate commit / branch with just nvcaffe + ssd but boostlog should effectively be the same thing if you leave settings at default.  Boost log branch adds a cmake flag that builds caffe using as the logging backend instead of glog.  This makes it possible for me to integrate into a logging application that I use that is based off of boost log.

[michael ben ami]

so for now i;ll check first the dtmoodie/caffe/tree/merge 

thanks!!! 

[johns...@gmail.com]

Hello Daniel, I don`t know how to use cmake to build ssd on window with this repo(https://github.com/dtmoodie/caffe/tree/merge_ssd) , error is Cmake can not find something Like glog, boost ,BTW, I have NugetPackages from MicroSoft-caffe-windows

在 2016年12月6日星期二 UTC+8上午5:15:52，Daniel Moodie写道：