students

students copied to clipboard

Users of this repo have to compile three different branches of Marian:

1. master: https://github.com/browsermt/students/blob/master/install.sh#L8-L9 (I think we can remove master)

2. Alham's quantization training branch: https://github.com/afaji/Marian/tree/fixed-quant

3. Nick's intgemm_reintegated_computestats branch

So a couple of questions:

1. How far is @afaji's fixed-quant branch from something that can be in master?
2. Can we merge fixed-quant and intgemm_reintegrated_computestats?

My intgemm_reintegrated_computestats is up to date with master. Uli took it and merged it with his mts branch, so @afaji should look to merge with data and write instructions. I have written instructions here https://github.com/browsermt/students/tree/master/train-student section 5.2 If I have done something wrong, please let me know.

Cheers,

Nick

will do it this week... so merge quantized training to master or to nick's branch?

@afaji, no with the branch /commit Id that this is pointing https://github.com/browsermt/mts/tree/master/3rd_party

Waiting for convergence on a branch, then I'll update cs<->en scripts.

This should be closed now? @kpu ?

I'm a bit confused on branches, I've been able to run the student training with marian-dev/master but this branch says it does not expect --optimize8 and --intgemm-shifted on the decoder but its marian-conv binary accepted ìntgemm8 as a gemm type. Should I use intgemm_reintegrated_computestats to use intgemm8 decoding?

@ZJaume --optimize8 and --intgemm-shifted were renamed. Where did you find them?

The current ones should be --int8shift or --int8shiftAlphaAll (as described here).

Found here.

From the list

• marian-dev/master
• marian-dev/intgemm_reintegrated_computestats
• browsermt/master
• afaifi/fixed-quant

Which of them are the preferred ones and for what? I know that you are currently developing a lot of stuff and it's difficult to have a single branch but it is a bit confusing to me :sweat_smile:

Sorry about that, I had forgotten to update this file with the new parameters

• marian-dev/intgemm_reintegrated_computestats
• browsermt/master Those have the same interface. You should always use browsermt/master, the rest are development branches. I try to keep the interface the same within all development branches, but no promises.

Thanks the help! So, to sum up, all of the features will go to browsermt/master at some point, right?

browsermt/master is up-to-date with intgemm_reintegrated_computestats. The goal is to work with upstream marian eventually, and all of those branches are temporary until we get all of our stuff upstream.