https://github.com/valaydave/attention_to_gif
Visualize transition of attention weights across layers in a Transformer as a GIF
https://github.com/valaydave/attention_to_gif
pytorch self-attention transformer visualisation
Last synced: 4 months ago
JSON representation
Visualize transition of attention weights across layers in a Transformer as a GIF
- Host: GitHub
- URL: https://github.com/valaydave/attention_to_gif
- Owner: valayDave
- License: mit
- Created: 2021-03-07T21:21:19.000Z (over 4 years ago)
- Default Branch: master
- Last Pushed: 2021-03-08T01:48:11.000Z (over 4 years ago)
- Last Synced: 2025-03-20T04:39:38.642Z (7 months ago)
- Topics: pytorch, self-attention, transformer, visualisation
- Language: Python
- Homepage:
- Size: 2.57 MB
- Stars: 3
- Watchers: 2
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: Readme.md
- License: LICENSE.txt
Awesome Lists containing this project
README
# Attention To Gif
A small library to make gif from Attention Weights collected from various layers of a [Transformer](https://en.wikipedia.org/wiki/Transformer_(machine_learning_model)).
## Setup
```sh
pip install git+https://github.com/valayDave/attention_to_gif
```
## Examples
[[Example Colab Link](https://colab.research.google.com/drive/1gGiBztUnbuPBW4c_UKle-slx1LojyBYd?usp=sharing)]
### View Video in Jupyter Notebook
```python
from attention_to_gif to import AttentionVisualizer
from moviepy.editor import VideoClip
# attention_tensor : ( num_layers ,batch_size ,num_heads ,seq_len_x ,seq_len_y )
visulizer = AttentionVisualizer(attention_tensor,x_label_toks=tokens,y_label_toks=tokens,fig_size=(10,10),chosen_head=2)
animation = VideoClip(make_frame=visulizer,duration=visulizer.num_layers)
animation.ipython_display(fps = 3, loop = False, autoplay = False)
```
### Save Gif :
```python
from attention_to_gif import AttentionVisualizer
# attention_tensor : ( num_layers ,batch_size ,num_heads ,seq_len_x ,seq_len_y )
visulizer = AttentionVisualizer(attention_tensor,x_label_toks=tokens,y_label_toks=tokens,fig_size=(10,10),chosen_head=2)
visulizer.save_visualisation(viz_name='Attention-Viz.gif',fps = 3)
```
### Create Single Plot in Matplot Lib
```python
from attention_to_gif import AttentionVisualizer
import matplotlib.pyplot as plt
# attention_tensor : ( num_layers ,batch_size ,num_heads ,seq_len_x ,seq_len_y )
visulizer = AttentionVisualizer(torch.randn(10,1,8,15,15),x_label_toks=tokens,y_label_toks=tokens,fig_size=(10,10),chosen_head=2)
fig = visulizer.create_single_plot(fig_size=(200,200))
plt.show()
fig.savefig('attention.png')
```
