Has ContextualWordEmbsAug gone slow?

[rajat-tech-002]

I had used ContextualWordEmbsAug for larger datasets earlier also. Has the code been significantly affected by some update. The augmentor speed is too slow now.

[050644zf]

I'm using this augmentor recently too and it takes 90 secs to process an sentence on my machine. and still take seconds on colab.

[rajat-tech-002]

@050644zf, so you also faced same issue?

[050644zf]

@050644zf, so you also faced same issue?

Yes, here is a testing notebook. You can see still takes 2.5sec to process an sentence.

[makcedward]

@rajat-tech-002
May I know which nlpaug version do you use? Supposed that speed should be improved for ContextualWordEmbsAug after applying tuning to fit multiple inputs to transformer's models.

[makcedward]

Tested from version 1.1.5 to 1.1.8

Performance is downgraded by 23%. (13.3s vs 16.4s) for 100 same inputs

[rajat-tech-002]

@rajat-tech-002 May I know which nlpaug version do you use? Supposed that speed should be improved for ContextualWordEmbsAug after applying tuning to fit multiple inputs to transformer's models.

I am using the latest version.

[rajat-tech-002]

Please close the issue.
I think on longer texts for eg. Yahoo dataset with average sentence length of 127, the augmenter would be very slow.
For eg. 300 seconds for 400 sentences on A100 GPU.
And it would be more if you do aug_max = None.
and aug_p = 30%

[makcedward]

Longer sentences affect performance as transformers need to handle more text rather than padding. Secondly, more percentage of augmentation affect performance more seriously.

More technical details:
For example, 10 tokens will be augmented, It needs to pass through transformers 10 times rather than 1 time. By BERT (or other models) design, masking language modeling predicts one token at a time.

However, I just notice there is major performance downgrade from 1.1.3 to 1.1.4. Time spending increased from 1s to 9s. The major change is adopting HuggingFace API rather than using my custom implementation. More tests need to conduct in order to identify the root cause.