nikolausmayer
commented
6 years ago You can use smaller input images, or a smaller model: FlowNet2-c needs less memory than FlowNet2-C, and FlowNet2-CS needs less memory than FlowNet2-CSS.
You could also rewrite the runtime script to execute the network layer-by-layer instead of loading the whole thing into memory at once. However, that'll be much slower, and (I imagine) also quite a bit of work.
Other than that, you'll need retraining.
When I run the demo,an error occur is
F0404 14:23:42.628688 24692 syncedmem.cpp:56] Check failed: error == cudaSuccess (2 vs. 0) out of memory Check failure stack trace: I know this is due to the small memory of my graphics card.Is there some parameter like yaml file to solve this problem?Thanks a lot.