search

MegEngine / FINet

This is the official MegEngine implementation of FINet: Dual Branches Feature Interaction for Partial-to-Partial Point Cloud Registration, AAAI 2022
MIT License
24 stars 1 forks source link

I've met some trouble in using megengine #3

Closed fffire-oss closed 1 year ago

fffire-oss commented 1 year ago

Traceback (most recent call last): File "train.py", line 148, in train_proc(params) File "train.py", line 136, in main train_and_evaluate(model, manager) File "train.py", line 92, in train_and_evaluate train(model, manager, gm) File "train.py", line 56, in train gm.backward(loss["total"]) File "/home/jzf/anaconda3/envs/meg/lib/python3.7/site-packages/megengine/autodiff/grad_manager.py", line 289, in backward self._grad(ys, dys) File "/home/jzf/anaconda3/envs/meg/lib/python3.7/site-packages/megengine/core/autodiff/grad.py", line 67, in call self._impl.backward(ys, dys) File "/home/jzf/anaconda3/envs/meg/lib/python3.7/site-packages/megengine/core/autodiff/grad.py", line 20, in backward return core2.backward(self, ys, dys) RuntimeError: assertion `val.shape().is_scalar()' failed at ../../../../../../src/opr/impl/tensor_manip.cpp:1011: bool mgb::opr::Split::infer_shape(size_t, megdnn::TensorShape&, const mgb::cg::static_infer::InpVal&) extra message: shapes for Split must be scalars

fffire-oss commented 1 year ago

I use MegEngine of the latest version==1.11.0

hxwork commented 1 year ago

Hi,

Thanks for your interest! I am sorry that I only tested my code using MegEngine==1.7.0. Maybe downgrading the MegEngine version can solve your problem.

Best regards, Hao

xaioqi16 commented 1 year ago

I encountered the same problem, how did you solve it

hxwork commented 1 year ago

Hi,

Have you tried to downgrade the MegEngine? Another workaround is using the Pytorch version of our work, please refer to this repo.

xaioqi16 commented 1 year ago

Thanks for your suggestion, I'm trying the pytorch version.