facebookresearch/ParlAI

New Metric.from_mask helper method

Open

#3,411 opened on Jan 26, 2021

View on GitHub
 (2 comments) (0 reactions) (1 assignee)Python (10,371 stars) (2,089 forks)batch import
EnhancementHelp WantedMediumnever-stale

Description

We have quite a few instances where we have some per-token losses/metrics along with a corresponding mask

metric_per_token # torch.Tensor of shape (batchsize, num_tokens)
mask # torch.BoolTensor of shape (batchsize, num_tokens)

And we want a per-batch-example average:

tokens_per_ex = mask.long().sum(dim=-1)
metric_per_ex = (metric_per_token * mask).sum(dim=-1)
metrics: List[MyMetric] = MyMetric.many(metric_per_ex, tokens_per_ex)
self.record_local_metric('metric_name', metrics)

I'd like us to have a helper classmethod in Metric called from_mask:

class Metric:
    @classmethod
    def from_mask(cls, metric_per_token, token_mask):
        # returns the equivalent of the "metrics" object above

Once this is done, add unit tests for this (test AverageMetric and PPLMetric directly). Checkpoint there.

After you've implemented this, upgrade TorchGeneratorAgent to use your new helper, upgrading the code for loss, ppl, and token_acc.

Example: https://github.com/facebookresearch/ParlAI/blob/67433e376fc361dee5aa045cb6bb2b68d3faa478/parlai/core/torch_generator_agent.py#L711-L727

See if you can find at least one other place who can benefit from upgrading this pattern.

Contributor guide