ju-resplande / viz_embeddings

Embeddings projections using TensorFlow Projector and HuggingFace Transformers

viz_embeddings

Embeddings projections for TensorFlow Projector using HuggingFace Transformers.

Example 1, Example 2

Requirements

The only requirements are: HuggingFace and NTLK.

Using pip :

 pip install -r requirements.txt

Usage

1. Run extract_embbedings to produce embeddings.tsv and vocab.tsv

   from extract import extract_embeddings
   
   extract_embeddings(text, model_name, level_name, aggr_func=torch.mean, filter_func=lambda x: len(x) >= 3, unique = True, doc_stride=1, do_lower=False)

   Args:

   text (str): input text

   model_name (str): Hugging Face model name

   level_name (str): text segmentation level name

   aggr_func (func): function to aggregate tensor embbedings

   filter_func (func): filter vocabulary level function

   unique (bool): if terms on level_name should be unique

   do_lower (bool): when unique = True, whether considere cased

2. View or Publish

   • Upload embeddings.tsv and vocab.tsv to Embedding Projector

     OR

   • Generate a config.json file as in examples/config.son

   • Access link http://projector.tensorflow.org/?config={your config_url}