Add Metric.from_mask helper method (#3411)

[poojasethi]

Patch description

This change introduces a new from_mask helper function in the Metric class. It also refactors the compute_loss function in torch_generator_agent.py to call the from_mask helper when computing the loss, ppl, and token_acc.

Testing steps

Unit tests

pytest -v tests/test_metrics.py
Note that I added two new unit tests, test_average_metric_from_mask and test_ppl_metric_from_mask.

Manual logging

I manually verified loss, ppl, token_acc, and token_em in torch_generator_agent.py by logging both the original and new values like this and running the below command:

 parlai train_model --model examples/seq2seq \           
    --model-file /tmp/example_model \
    --task convai2 --batchsize 32 --num-epochs 2 --truncate 128 --seed 42 > metric_from_mask.txt

We can see that old_loss equals loss, old_ppl equals ppl, etc.

2022-11-22 17:35:17,429 INFO     | time:192s total_exs:6400 total_steps:200 epochs:0.05 time_left:7687s
    clen  clip  ctpb  ctps  ctrunc  ctrunclen  exps  exs  gnorm  llen  loss  lr  ltpb  \
   140.9     1  3457  3633   .5481      32.84 33.63 1600  15.06 12.85 8.829   1 411.2   
    ltps  ltrunc  ltrunclen  old_loss  old_ppl  old_token_acc  old_token_em  ppl  \
   432.1       0          0     8.829     6827          .1721             0 6827   
    token_acc  token_em  total_train_updates  tpb  tps   ups  
        .1721         0                  200 3868 4065 1.051

Other information

• I believe some of these files could potentially be refactored to use the new helper function as well. Happy to submit another PR, if helpful!
• I tried running autoformat.sh but it was running quite slowly... Tips on how to use this script would be helpful :)

[stephenroller]

This looks like it fixes that old issue! Actually @mojtaba-komeili looks like he may be able to use this.

Leaving one comment to see if we can go even one step cleaner

[stephenroller]

Can we use cls.many instead and get away without passing MyMetric as an extra parameter?

[poojasethi]

Ooh, yes! Good idea.

[stephenroller]

lgtm!

[poojasethi]

Sweet! @stephenroller and @klshuster should I wait for / help dig into the lint and CircleCI failures, or go ahead and land the PR?

[poojasethi]

Landed! (Discussed with @klshuster offline and it seems that the lint and CircleCI failures are not related to this PR)