Is it possible to use Sagemaker clarify to generate shap explanations like in
this example or this
I can’t find any examples specific to huggingface transformers, but from the documentation it looks like it may just be a matter of passing text_config with granularity = “token”?
Let me check with the AWS Team. I haven’t used SageMaker Clarify yet.
@MaximusDecimusMeridi the AWS team shared with me this example:
This file has been truncated.
"# Explaining text sentiment analysis using SageMaker Clarify"
" 1. [Overview](#Overview)\n",
" 1. [Prerequisites and Data](#Prerequisites-and-Data)\n",
" 1. [Initialize SageMaker](#Initialize-SageMaker)\n",
" 1. [Loading the data: Women's Ecommerce clothing reviews Dataset](#Loading-the-data:-Women's-ecommerce-clothing-reviews-dataset) \n",
" 1. [Data preparation for model training](#Data-preparation-for-model-training) \n",
" 1. [Train and Deploy Hugging Face Model](#Train-and-Deploy-Hugging-Face-Model)\n",
" 1. [Train model with Hugging Face estimator](#Train-model-with-Hugging-Face-estimator)\n",
i haven’t looked at it. So would be nice if you could report back on this and the experience
@philschmid It looks great, haven’t gone through in detail yet but this is the output from running it
Are there any native HF tools for explainability? I know there’s
exBERT but it looks more like a UI tool for now. Is there anything for getting token level explanation programatically? I’ve found a few independently made tools:
But is there anything directly in HF to do this?
And no we currently have nothing from HF directly.