Martin Simonovsky
Martin Simonovsky
Hi, thanks a lot for releasing a 3rd party implementation of the paper. Neverthless, I'm afraid there is some problem with your code or at least the hyperparameters are not...
The requirements for contiguous tensors seem to be too strict with R3. Actually, cudnn supports operations on non-contiguous data well; there are just some constraints, typically that input and gradInput...
I think the behavior on L141 in `gsp_graph_multiresolution.m` is wrong in the case when `largest_eigenvector(1)==0`, which now leads to filling `largest_eigenvector` with 0s as `sign(0)=0`. I suggest L141 being executed...
The current code throws "GSP_GRAPH_SPARSIFY: Epsilon out of required range" if `gsp_graph_multiresolution()` comes to the point that the current graph has three nodes. I think L175 in gsp_graph_multiresolution.m should be...
I have encountered weird behavior when turning on optimization. The following code computes a simple MLP with cross-entropy loss. The loss can be computed in two ways (`CrossEntropyCriterion` or `LogSoftMax`&`ClassNLLCriterion`)....
## Motivation There are still many Python packages around which stick with the old 3.20 protobuf version. However, this is typically the very last release of major version 3: 3.20.3....
Currently, nD cells are currently always loaded as linearized 1D tables. They should be loaded as nD Lua tables instead.