I design a NN using for instance segmentation which has some deconv layers. When I deploy NN using SDK, there is no error and the application can run on A311D successfully. But the NPU outputs are incorrect. I take time to find where the results are inconsistent with the GPU. I find that intermediate results before deconv layer are correct. But the intermediate results after deconv layer are incorrect.
I use SDK6.4.3 and Caffe. And the environment is updated using ‘apt full-grade’.But when I depolyed the same netwok using SDK 6.4.0.10 some time ago, NPU results are correct!