I've just finished the first draft of tensorflow i...

# mathematicsa

altavir

11/10/2021, 5:56 PMI've just finished the first draft of tensorflow integration for KMath (using graph lazy computation of course). https://github.com/mipt-npm/kmath/tree/feature/tensorflow. Comments are welcome (especially from **@zaleslaw**). I plan to add more operations as extensions, direct operation builder (use tensorflow graph directly with some improvements from a special scope in algebra). Compilation of MST to graph (expressions and preliminary optimizations).

🔝 2

z

zaleslaw

11/10/2021, 6:31 PMHi, great news, what is the goal of this integration, where can I find some examples?

a

altavir

11/10/2021, 8:10 PMThe goal is to provide low-level tensor and StructureND API implementation over TF. And to allow people create TF lazy graph without the usual boilerplate. Right now I have only one test that shows it is actually working, but we can implement some examples by **@Ролан**'s team since they rely only on tenor API. Several operations are still not implemented, but I will fix that.

z

zaleslaw

11/12/2021, 12:10 PMSo, the idea is pretty clear, I will wait for full implementation or more tests/examples, to understand all the benefits

a

altavir

11/12/2021, 12:28 PMThe primary benefit is that you can do computations with "slow" core and then switch to "fast" tensorflow. As for tensorflow itself, we use lazy graph API even for intermediate computations, so in theory it should be very fast. Need to test all that though.

8 Views