Converting e2e_faster_rcnn_X_101_32x8d_FPN_1x_caffe2.pkl to protobuf

[karanchawla]

❓ Questions and Help

Do we use the Detectron's script to convert the pkl config files to protobuf for C++ deployment?

I'm getting the following error when I try to convert the aforementioned file using that python script:

(result, consumed) = self._buffer_decode(data, self.errors, final)
UnicodeDecodeError: 'utf-8' codec can't decode byte 0x80 in position 0: invalid start byte

[fmassa]

Hi,

Could you expand on what you are trying to do?
It's not very clear to me.

Thanks!

[karanchawla]

I'm trying to run this model as part of a C++ project and was attempting to convert the pkl file to a pytorch model file, however, being new to pytorch I wasn't sure how to proceed with this. So, I was trying to use Detectron's script for converting the model to protobuf. Do you have any pointers for that?

[fmassa]

Hi,

If you want to use Detectron for C++ deployment, you should ask your questions in https://github.com/facebookresearch/Detectron, as they will be in a better position to help you.

That being said, I will at some point add support for exporting PyTorch models into C++ using new functionality from PyTorch 1.0. But it might still take a few weeks for that.

Given that this doesn't look like a question / issue with maskrcnn-benchmark, I'm closing this issue.

[karanchawla]

Hi, thanks! Sorry, if I miscommunicated, I meant I have been trying to deploy maskrcnn-benchmark on C++ but since I didn't find any tools to convert the models to C++ compatible format, I tried using the tools provided with Detectron.

I see, could you point me to how I should go about deploying this on C++? Maybe I can implement it and open a PR?

Thanks!

[fmassa]

Hi,

I indeed misunderstood your question.

I'll be looking into supporting jit tracing in the future. This will probably involve modifying the way I use C++ extensions.
This means that it will be possible to have maskrcnn-benchmark running in standalone C++ apps.

I can send you some pointers on how to approach that, but I don't have all the answers ahead of time, and there might be some subtle issues that might require some digging.

I'm currently tracking this issue in #27, so to avoid duplicates I'll be closing this issue. Let's continue the discussion there.