Supervision optimization

[chenzhehuai]

Supervision optimization to make the algorithm linear-time in the minibatch size, per minibatch, instead of quadratic-time.

• by: change to call Concat(const Fst &fst1, MutableFst *fst2)

• details:

previously we call "Concat(MutableFst<Arc> *fst1, const Fst<Arc> &fst2)" from AppendSupervision() by:

  for (int32 i = 0; i < input.size(); i++) {
    const Supervision &src = *(input[i]);
    fst::Concat(&output_supervision->back().fst, src.fst);
  }

now change to call "Concat(const Fst<Arc> &fst2, MutableFst<Arc> *fst1)" by

  for (int32 i =  input.size()-1; i >= 0; i--) {
    const Supervision &src = *(input[i]);
    fst::Concat(src.fst, &output_supervision->back().fst);
  }

---

• run chain-supervision-test:

/export/a12/zchen/works/concat/chain-supervision-test.log

• test whether the output of nnet3-chain-merge-egs is the same before and after modification:

zchen@g02:/export/a12/zchen/works/concat$  nnet3-chain-merge-egs --minibatch-size=256,128,64 ark:exp/chain/tdnn_lstm_1e_sp/egs/storage/2/cegs.1.ark ark,t:tmp/test3.txt
zchen@g02:/export/a12/zchen/works/concat$  nnet3-chain-merge-egs --minibatch-size=256,128,64 ark:exp/chain/tdnn_lstm_1e_sp/egs/storage/2/cegs.1.ark ark,t:tmp/test.txt
diff /export/a12/zchen/works/concat/tmp/test.txt /export/a12/zchen/works/concat/tmp/test3.txt

[chenzhehuai]

add things below

1. simplify the interface AppendSupervision by: i) delete compactify 2) use single output_supervision instead of vector
2. modify discriminative-supervision.cc correspondingly, along with test files and other calling functions
3. test correctness:

zchen@g02:/export/a12/zchen/works/concat$ nnet3-chain-merge-egs --minibatch-size=256,128,64 ark:exp/chain/tdnn_lstm_1e_sp/egs/storage/2/cegs.1.ark ark,t:tmp/test4.txt
zchen@g02:/export/a12/zchen/works/concat$ diff tmp/test4.txt tmp/test3.txt     

4. test speedup:

/export/a12/zchen/works/concat/test_time.sh

zchen@login:/export/a12/zchen/works/concat$ grep real watanabe*
watanabe.log:real       3m28.864s
watanabe.log.new:real   0m23.637s

[danpovey]

A few cosmetic comments.
Please check the changes carefully, as we may not test all of them (e.g. the discriminative-training ones).

[danpovey]

here and elsewhere, please put space around the operator -, i.e. 'num_inputs - 2'. This is how we normally do it. And I prefer 'i >= 0' to 'i > -1'.

[danpovey]

no need for parentheses.

[danpovey]

please make this a bit longer, and start with a capital.

[chenzhehuai]

done.
I also do a test on discriminative-supervision.cc, now it is the same

/export/a12/zchen/works/concat/test_disc.sh

[danpovey]

thanks! Will merge when tests finish.