Pre-trained file cannot be loaded with pytorch 1.0.1 about mlcpp HOT 5 CLOSED

Could you please provide more details, normally files exported from pytorch can't be directly loaded to with C++ API. Please look at the StateLoader class, there are details how to export from python pytorch.

from mlcpp.

Thanks for replying.
The details as follows,

1. I compiled mask_rcnn_pytorch in VS 2017 x64 on Pytorch 1.0.1 for windows downloaded from pytorch.org

2. I download 3 pre-trained files from the addresses listed in readme(https://drive.google.com/file/d/1Y9Qq4u-DxiM_eYJ7w4-VXWp4XmUeoU3t/view?usp=sharing...)

3. I failed to load the pre-trained files, because they were created by pytorch 1.0 preiview and pytorch 1.01 does not take them, Actually if there is "TORCH1" in the head of file, it will be rejected

4. I tried to convert mask_rcnn_cocol.pth to C++ compatible in project https://github.com/multimodallearning/pytorch-mask-rcnn, with pytorch 1.01, unfortunately there is conflicts in model. The code as below,
   model.load_state_dict(torch.load(COCO_MODEL_PATH))
   example = torch.rand(1, 3, 224, 224)
   traced_script_module = torch.jit.trace(model, example)
   traced_script_module.save("D:/model.pt")

Any suggestions? thanks!

from mlcpp.

Please use this one https://github.com/mjstevens777/pytorch-mask-rcnn python implementation because original from multimodallearning is outdated.

from mlcpp.

Please check this guy's job at your convenience. https://github.com/darolt/mask_rcnn

from mlcpp.

I've update parameters file to be compatible with PyTorch v1.0.1.

from mlcpp.