JVET-AC0057 AhG14 SADL v4 changes [F. Galpin, T. Dumas, P. Bordes, E. François (InterDigital)]

This contribution presents updates in the Small AdHoc Deep Learning (SADL) library. The update contains both some performance improvement and new features.

New features include support for sparse matrix-vector multiplication for network using dense layers, slicing layer, grouped convolutions and separable convolutions.

New features in SADL version 4:

• Sparse matrix-vector multiplication, no need to change model, useful for intra prediction
• Separable convolutions: useful for model complexity reduction
• Grouped convolution: useful for model complexity reduction. This changes the model format, and it is recommended for existing models to be updated to this new model format.
• Unit testing: useful for regression testing on models, and layers within a model
• New layer for slicing tensor: this was not listed in the contribution but included in the SADL update

It was reported that it would be desirable to automate the regression testing of merge requests. This is currently not done for NNVC. It was commented by software coordinator that some additional testing may be accommodated, but if excessive testing is performed, there could be computing resource issues on the git server. This is delegated to the AHG3 and AHG14 software coordinators to find a reasonable solution to meet the request.

The proposed actions are discussed as follows:

• It was proposed to update the SADL version in NNVC to the version 4 for the next software version. Currently SADL version 3 is used, and it was commented that it would be beneficial to use the latest SADL version in NNVC. The AHG14 software coordinator commented that there are no known issues with this proposed SADL version update (other than the need to update models to new format as noted above)
• It was proposed to make the SADL repository public to be aligned with the NNVC repository. Currently, SADL is not public whereas NNVC is public. This means that to clone SADL repo, MPEG credentials would be needed, whereas cloning the NNVC repo does not require credentials. Due to the fact that NNVC has dependency on SADL, this discrepancy creates some access difficulty for participants. It was commented that there did not seem to be specific reasons for keeping SADL private. It was commented that making resource publicly available seems to be a good practice in general. It was also commented that other than SADL, all other software repos on the git server are public (NNVC was made public in the last meeting). This was agreed. K. Sühring will take the action to make SADL public.

Decision (SW): Use SADL version 4 in the next NNVC software release.