cbecker

Is this still working with the current master at https://github.com/pytorch/pytorch ? I tried different options and I'm still hitting the 65535 symbol limit.

@pjh5 I am building Caffe2 from the pytorch repo https://github.com/pytorch/pytorch I am using the options: ``` USE_OPENCV=OFF USE_OPENMP=OFF BUILD_ONNX_PYTHON=OFF BUILD_CUSTOM_PROTOBUF=ON BUILD_TEST=OFF USE_METAL=OFF USE_GFLAGS=OFF USE_GLOG=OFF USE_NNPACK=OFF USE_NCCL=OFF CUDA_USE_STATIC_RUNTIME=OFF CAFFE2_STATIC_LINK_CUDA=OFF USE_OBSERVERS=OFF ```...

I'll try that tomorrow (I don't have a Windows machine where I am now) with BUILD_ATEN and BUILD_BINARY off. I was using /Ob2 and Release, but not sure about these...

@pjh5 I tried with BUILD_BINARY and BUILD_ATEN OFF with master from today, and the problem is still there. Is this an issue that will be fixed soon? I am wondering...

I am glad to contribute to this, as I believe it's a big blocker for anyone wanting to use caffe2 on Windows. I am thinking what would be the best...

Thanks @peterjc123 , that will help. Nonetheless, wouldn't it then make sense to limit symbol exporting in all platforms, to be consistent and to easy development?

I've been looking for an alternative to scp, and @pdcastro's solution works great. Maybe eventually we'll have that within balena-cli, I hope :)

On second thoughts, maybe the solution would be to limit the amount of samples per tree according to the class that has the least amount of samples, so that the...