Some fixes for the Explaining a Question Answering Transformers Model Example Notebook #3685
Conversation
There was a problem hiding this comment.
Thanks for the great PR and the in-depth analysis.
I have three suggestions:
- Could we somehow get rid of the "TF-TRT Warning: Could not find TensorRT" warning? Either just delete the warning in the .ipynb file or install the according stuff. This is optional, if you struggle too much with it, please just go ping me here.
- I would suggest to swap the order of things so that
def f_start(questions):
return f(questions, tokenized_qs, True)
def f_end(questions):
return f(questions, tokenized_qs, False)tokenized_qs is defined before it is first used in the functions. My IDE gives me some warnings about this, so please just define it first.
3. Could you please run the cells in order, so that the appropriate numbers are displayed in the notebook?
|
@CloseChoice , thanks for the tips. I have updated the notebook with the following:
|
|
Thanks for your PR. I looked deeper into this and I have to say, our current implementation (not your notebook) does seem incompatible with this notebook, so your solution is a clever workaround. Thanks. |
This pull request concerns the example "Explaining a Question Answering Transformers Model"
Problems
A: Re-Tokenization Issue
The analysis does not work if a word is divided into several tokens and one of the subwords is also an independent word in the tokenizer's vocabulary.
In the analysis that this example code puts forward, the input is tokenized and masked. But then, the input is detokenized, and the part that is not masked, is retokenized. This can result in a different retokenization than the original one. This, then results in a different sequence length, causing shape mismatch errors.
For example, consider the tokenization:
['what', 'did', 'i', 'eat', 'for', 'now', '?', '[SEP]', 'i', 'picked', 'up', 'a', 'bag', 'of', 'peanuts', 'and', 'rai', '##sin', '##s', 'for', 'a', 'snack', '.', 'i', 'wanted', 'a', 'sweet', '##er', 'snack', 'out', 'so', 'i', 'ate', 'them', 'for', 'now', '.']The problem arises when a part of the text including "rai" gets masked. In this case, "sins" will not be retokenized as "##sin" and "##s", but as "sins", which is recognized as a full word in the tokenizer's vocabulary. Thanks to Yuka Wolter for pointing out this bug to me.
B: Inflexible Input Formatting
The current example only works for input formatted like this:
However, it fails if the input is formatted without whitespaces before and after [SEP]:
Proposed changes
To address these issues, I propose rewriting f(questions, start) to avoid retokenizing after masking. The modified code below delivers the same results as the original but should be more robust because it avoids retokenization.
The idea is to use the same text masker as the original code, but force it to output ids instead of strings:
explainer_start = shap.Explainer(f_start, shap.maskers.Text(tokenizer=pmodel.tokenizer, output_type='ids'))This way, f(questions, start) works directly with ids and does not need to retokenize anything.
Remaining problems
This example still has some remaining problems:
Checklist