BertViz

Visualize Attention in Transformer Models (BERT, GPT2, T5, etc.)

Quick Tour • Getting Started • Colab tutorial • Blog • Paper • Citation

BertViz is an interactive tool for visualizing attention in Transformer language models such as BERT, GPT2, or T5. It can be run inside a Jupyter or Colab notebook through a simple Python API that supports most Huggingface models. BertViz extends the Tensor2Tensor visualization tool by Llion Jones, adding multiple views that each offer a unique lens into the attention mechanism.

This is the official code repository for A Multiscale Visualization of Attention in the Transformer Model by Jesse Vig.

🚀 Quick Tour

Head View

The head view visualizes attention for one or more attention heads in the same layer. It is based on the excellent Tensor2Tensor visualization tool by Llion Jones.

🕹 Try out the head view in an interactive Colab tutorial (all visualizations pre-loaded).

Model View

The model view shows a bird's-eye view of attention across all layers and heads.

🕹 Try out the model view in an interactive Colab tutorial (all visualizations pre-loaded).

Neuron View

The neuron view visualizes individual neurons in the query and key vectors and shows how they are used to compute attention.

🕹 Try out the neuron view in an interactive Colab tutorial (all visualizations pre-loaded).

⚡️ Getting Started

Colab

Add the following cell at the beginning of a Colab notebook to use BertViz:

!pip install bertviz

Sample code

Run the following code to load the distbert-base-uncased model and display it in the model view:

from transformers import AutoTokenizer, AutoModel, utils
from bertviz import model_view
utils.logging.set_verbosity_error()  # Suppress standard warnings

tokenizer = AutoTokenizer.from_pretrained("distilbert-base-uncased")
model = AutoModel.from_pretrained("distilbert-base-uncased", output_attentions=True)
inputs = tokenizer.encode("The cat sat on the mat", return_tensors='pt')
outputs = model(inputs)
attention = outputs[-1]  # Output includes attention weights when output_attentions=True
tokens = tokenizer.convert_ids_to_tokens(inputs[0]) 
model_view(attention, tokens)

The visualization may take a few seconds to load. Feel free to experiment with different input texts and models. See Documentation for additional use cases and examples.

Jupyter Notebook

You may also run BertViz locally inside a Jupyter Notebook.

Installation

pip install bertviz

You must also have Jupyter Notebook and ipywidgets installed:

pip install jupyterlab
pip install ipywidgets

If you have any issues installing Jupyter or ipywidgets, consult the documentation (jupyter, ipwidgets).

Sample code

You may use the same sample code from above.

To create a new notebook, first start Jupyter Notebook:

jupyter notebook

Then click New to create a new notebook, and select Python 3 (ipykernel) if prompted.

Running included notebooks

You may also run any of the sample notebooks included with BertViz:

git clone --depth 1 [email protected]:jessevig/bertviz.git
cd bertviz/notebooks
jupyter notebook

🕹 Interactive Tutorial

Check out this Colab notebook for an interactive tutorial on BertViz. Note: all visualizations are pre-loaded, so there is no need to execute any cells.

📖 Documentation

Table of contents

• Self-attention models (BERT, GPT-2, etc.)
  • Head and Model Views
  • Neuron View
• Encoder-decoder models (BART, MarianMT, etc.)
• Installing from source
• Additional options
  • Dark / light mode
  • Filtering layers
  • Setting default layer/head(s)
  • Visualizing sentence pairs
  • Non-Huggingface models

Self-attention models (BERT, GPT-2, etc.)

Head and Model Views

First load a Huggingface model, either a pre-trained model as shown below, or your own fine-tuned model. Be sure to set output_attention=True.

from transformers import AutoTokenizer, AutoModel, utils
utils.logging.set_verbosity_error()  # Suppress standard warnings
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
model = AutoModel.from_pretrained("bert-base-uncased", output_attentions=True)

Then prepare inputs and compute attention:

inputs = tokenizer.encode("The cat sat on the mat", return_tensors='pt')
outputs = model(inputs)
attention = outputs[-1]  # Output includes attention weights when output_attentions=True
tokens = tokenizer.convert_ids_to_tokens(inputs[0]) 

Finally, display the attention weights using the head_view or model_view functions:

from bertviz import head_view
head_view(attention, tokens)

Examples: DistilBERT (Model View Notebook, Head View Notebook)

For full API, please refer to the source code for the head view or model view.

Neuron View

The neuron view is invoked differently than the head view or model view, due to requiring access to the model's query/key vectors, which are not returned through the Huggingface API. It is currently limited to custom versions of BERT, GPT-2, and RoBERTa included with BertViz.

# Import specialized versions of models (that return query/key vectors)
from bertviz.transformers_neuron_view import BertModel, BertTokenizer
from bertviz.neuron_view import show

model_type = 'bert'
model_version = 'bert-base-uncased'
do_lower_case = True
sentence_a = "The cat sat on the mat"
sentence_b = "The cat lay on the rug"
model = BertModel.from_pretrained(model_version, output_attentions=True)
tokenizer = BertTokenizer.from_pretrained(model_version, do_lower_case=do_lower_case)
show(model, model_type, tokenizer, sentence_a, sentence_b, layer=2, head=0)

Examples: BERT (Notebook, Colab) • GPT-2 (Notebook, Colab) • RoBERTa (Notebook)

For full API, please refer to the source.

Encoder-decoder models (BART, MarianMT, etc.)

The head view and model view both support encoder-decoder models.

First, load an encoder-decoder model:

from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("Helsinki-NLP/opus-mt-en-de")
model = AutoModel.from_pretrained("Helsinki-NLP/opus-mt-en-de", output_attentions=True)

Then prepare the inputs and compute attention:

encoder_input_ids = tokenizer("She sees the small elephant.", return_tensors="pt", add_special_tokens=True).input_ids
decoder_input_ids = tokenizer("Sie sieht den kleinen Elefanten.", return_tensors="pt", add_special_tokens=True).input_ids

outputs = model(input_ids=encoder_input_ids, decoder_input_ids=decoder_input_ids)

encoder_text = tokenizer.convert_ids_to_tokens(encoder_input_ids[0])
decoder_text = tokenizer.convert_ids_to_tokens(decoder_input_ids[0])

Finally, display the visualization using either head_view or model_view.

from bertviz import model_view
model_view(
    encoder_attention=outputs.encoder_attentions,
    decoder_attention=outputs.decoder_attentions,
    cross_attention=outputs.cross_attentions,
    encoder_tokens= encoder_text,
    decoder_tokens = decoder_text
)

You may select Encoder, Decoder, or Cross attention from the drop-down in the upper left corner of the visualization.

Examples: MarianMT (Notebook) • BART (Notebook)

For full API, please refer to the source code for the head view or model view.

Installing from source

git clone https://github.com/jessevig/bertviz.git
cd bertviz
python setup.py develop

Additional options

Dark / light mode

The model view and neuron view support dark (default) and light modes. You may set the mode using the display_mode parameter:

model_view(attention, tokens, display_mode="light")

Filtering layers

To improve the responsiveness of the tool when visualizing larger models or inputs, you may set the include_layers parameter to restrict the visualization to a subset of layers (zero-indexed). This option is available in the head view and model view.

Example: Render model view with only layers 5 and 6 displayed

model_view(attention, tokens, include_layers=[5, 6])

For the model view, you may also restrict the visualization to a subset of attention heads (zero-indexed) by setting the include_heads parameter.

Setting default layer/head(s)

In the head view, you may choose a specific layer and collection of heads as the default selection when the visualization first renders. Note: this is different from the include_heads/include_layers parameter (above), which removes layers and heads from the visualization completely.

Example: Render head view with layer 2 and heads 3 and 5 pre-selected

head_view(attention, tokens, layer=2, heads=[3,5])

You may also pre-select a specific layer and single head for the neuron view.

Visualizing sentence pairs

Some models, e.g. BERT, accept a pair of sentences as input. BertViz optionally supports a drop-down menu that allows user to filter attention based on which sentence the tokens are in, e.g. only show attention between tokens in first sentence and tokens in second sentence.

Head and model views

To enable this feature when invoking the head_view or model_view functions, set the sentence_b_start parameter to the start index of the second sentence. Note that the method for computing this index will depend on the model.

Example (BERT):

from bertviz import head_view
from transformers import AutoTokenizer, AutoModel, utils
utils.logging.set_verbosity_error()  # Suppress standard warnings

# NOTE: This code is model-specific
model_version = 'bert-base-uncased'
model = AutoModel.from_pretrained(model_version, output_attentions=True)
tokenizer = AutoTokenizer.from_pretrained(model_version)
sentence_a = "the rabbit quickly hopped"
sentence_b = "The turtle slowly crawled"
inputs = tokenizer.encode_plus(sentence_a, sentence_b, return_tensors='pt')
input_ids = inputs['input_ids']
token_type_ids = inputs['token_type_ids'] # token type id is 0 for Sentence A and 1 for Sentence B
attention = model(input_ids, token_type_ids=token_type_ids)[-1]
sentence_b_start = token_type_ids[0].tolist().index(1) # Sentence B starts at first index of token type id 1
token_ids = input_ids[0].tolist() # Batch index 0
tokens = tokenizer.convert_ids_to_tokens(token_ids)    
head_view(attention, tokens, sentence_b_start)

Neuron view

To enable this option in the neuron view, simply set the sentence_a and sentence_b parameters in neuron_view.show().

Non-Huggingface models

The head view and model view may be used to visualize self-attention for any standard Transformer model, as long as the attention weights are available and follow the format specified in head_view and model_view (which is the format returned from Huggingface models). In some case, Tensorflow checkpoints may be loaded as Huggingface models as described in the Huggingface docs.

⚠️ Limitations

Tool

• This tool is designed for shorter inputs and may run slowly if the input text is very long and/or the model is very large. To mitigate this, you may wish to filter the layers displayed by setting the include_layers parameter, as described above.
• When running on Colab, some of the visualizations will fail (runtime disconnection) when the input text is long. To mitigate this, you may wish to filter the layers displayed by setting the include_layers parameter, as described above.
• The neuron view only supports the custom BERT, GPT-2, and RoBERTa models included with the tool. This view needs access to the query and key vectors, which required modifying the model code (see transformers_neuron_view directory), which has only been done for these three models. Also, only one neuron view may be included per notebook.

Attention as "explanation"

Visualizing attention weights illuminates a particular mechanism within the model architecture but does not necessarily provide a direct explanation for model predictions. See [1, 2, 3].

🔬 Paper

A Multiscale Visualization of Attention in the Transformer Model (ACL System Demonstrations 2019).

Citation

@inproceedings{vig-2019-multiscale,
    title = "A Multiscale Visualization of Attention in the Transformer Model",
    author = "Vig, Jesse",
    booktitle = "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: System Demonstrations",
    month = jul,
    year = "2019",
    address = "Florence, Italy",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/P19-3007",
    doi = "10.18653/v1/P19-3007",
    pages = "37--42",
}

Authors

Jesse Vig (homepage)

🙏 Acknowledgments

We are grateful to the authors of the following projects, which are incorporated into this repo:

• https://github.com/tensorflow/tensor2tensor
• https://github.com/huggingface/pytorch-pretrained-BERT

License

This project is licensed under the Apache 2.0 License - see the LICENSE file for details