Anyone succeed in converting pth to onnx format?

[liamsun2019]

Hi Pals,

I tried to convert the trained pth model to onnx format, but encounter several issues. The code segment are as follows:

model = Trainer.build_model(cfg)
state = torch.load(cfg.MODEL.WEIGHTS, map_location=lambda storage, loc: storage)
model.load_state_dict(state['model'])
model.eval()
model.cuda()

must be 32-divisable

dummy_input = torch.randn(1, 3, 448, 448).to("cuda")
torch.onnx.export(model, dummy_input, "model.onnx", verbose=True, input_names=['image'], output_names=['pred'])

The used config file is centermask_lite_V_19_slim_dw_eSE_FPN_ms_4x.yaml.

a. The input to the network whose backbone is a rcnn is a list which contains dictionaries composed of input image data and some other attirbutes. Looks like there's no way to pass similar input for onnx exporting. I added a member function in rcnn.py (detectron2/modeling/meta_arch/rcnn.py) which accepts such input to solve this issue.

b. Based on above, following error rises up during conversion:
RuntimeError: Failed to export an ONNX attribute 'onnx::Sub', since it's not constant, please try to make things (e.g., kernel size) static if possible

I cannot google any clues to the error. Anyone can kindly help? Thanks.

[liamsun2019]

reduplicated issue