Lux new

[jeremiedb]

Migration from v0.3 to v0.4 by moving from Flux/Zygote to Lux/Reactant

[jeremiedb]

@AdityaPandeyCN I've just merged the jdb/grant-review along your branch into lux-new to have the bulk of the Lux adaptation in lux-new.
However, issues with deval and infers still seem present:

• If deval is specified, training on gpu crashes with Scalar indexing. And on CPU, it having using dtrain as deval doubles the training time, which is abnormal. Evaluation shouldn't take more than a single epoch training time.
• Inference should returns return an inference for each of the requested datapoint. By using partial=false for DataIter, this condition can't be met. The idea of having partial=false was just to simplify debugging process. It should be possible to get a complete inference efficiently with Lux tools (might be a comination of loop on complete batch using the cmpiled batch inference (ex: apply(x, ps, st)) along a single @jit apply step for the last batch)

[AdityaPandeyCN]

@jeremiedb Let me have a look at it again, I was able to run this locally on gpu, but may have missed committing some part of the code.

[AdityaPandeyCN]

Hello @jeremiedb I have raised a PR(#27) with the fix for the above problems:-

1. Inference is working now.
2. For the deval case when running on GPU, the timings are good(not much difference) but in the CPU case the timings are still a big problem. I managed to bring them bit lower but still not very good. I have tried taking more things under compile to fast up but still not good for the cpu case.