mx-maskrcnn: MXNetError: [11:03:30] /mx-maskrcnn/incubator-mxnet/mshadow/mshadow/././././cuda/tensor_gpu-inl.cuh:110: Check failed: err == cudaSuccess (7 vs. 0) Name: MapPlanKernel ErrStr:too many resources requested for launch

First of all, many thanks to Tusimple and @Zehaos , it helps the open source community a lot.

I have met the same problem with the latest version. But I have solved this problem by using this commit commit b3e4a91885ae0a468a6162e129c25a381e6cc1a3 I suggest @jinfagang you try this commit

Environment information:

Ubuntu 16.04
Nvidia Titan X, driver 384.90
CUDA Version 8.0.61(use 'cat /usr/local/cuda/version.txt')
cudnn 6.0.21 (use 'cat /usr/local/cuda/include/cudnn.h | grep CUDNN_MAJOR -A 2' to check)

Some details of my case and solution: I have met this problem with driver version 375.26, cudnn 5.1 and maskrcnn commit b68216d. I have tried several different solutions, but only this environment config could work.

If someone who met the similar problem, please have a try with the following steps:

1. Update your driver to 384.90
2. Make sure your cuda is 8.0.61
3. Use cudnn  6.0.21
4. git clone --recursive https://github.com/TuSimple/mx-maskrcnn.git 
5. cd mx-maskrcnn
6. git checkout b3e4a91
7. cp rcnn/CXX_OP/* incubator-mxnet/src/operator/
8. build MXNet from source, please refer to the [tutorial](https://mxnet.incubator.apache.org/get_started/build_from_source.html)

Hope it works.

I’m not sure what causes this problem, maybe there exist conflicts between the operators and cudnn/ driver.

There is another solution [Fast-RCNN] Proposal operator failed: too many resources requested for launch #6204** , but I didn’t try it.

Problems solved by everyone’s effort, please be open and be patient to problem and issues. I believe the author spent a lot of effort on this project. It is meaningless to abuse