https://github.com/kyegomez/ct

Implementation of the attention and transformer from "Building Blocks for a Complex-Valued Transformer Architecture"
https://github.com/kyegomez/ct

artificial-intelligence attention attention-is-all-you-need attention-mechanism complex-numbers complex-signals dotproduct fourier-transform

Last synced: 6 days ago
JSON representation

Implementation of the attention and transformer from "Building Blocks for a Complex-Valued Transformer Architecture"

• Host: GitHub
• URL: https://github.com/kyegomez/ct
• Owner: kyegomez
• License: mit
• Created: 2023-10-07T15:05:33.000Z (about 1 year ago)
• Default Branch: main
• Last Pushed: 2024-03-11T02:31:36.000Z (8 months ago)
• Last Synced: 2024-09-18T06:15:44.476Z (about 2 months ago)
• Topics: artificial-intelligence, attention, attention-is-all-you-need, attention-mechanism, complex-numbers, complex-signals, dotproduct, fourier-transform
• Language: Python
• Homepage: https://discord.gg/qUtxnK2NMf
• Size: 2.16 MB
• Stars: 8
• Watchers: 2
• Forks: 0
• Open Issues: 1
• Metadata Files:
  • Readme: README.md
  • Funding: .github/FUNDING.yml
  • License: LICENSE

Awesome Lists containing this project

README

        
[![Multi-Modality](agorabanner.png)](https://discord.gg/qUtxnK2NMf)

# Complex Transformer (WIP)

The open source implementation of the attention and transformer from "Building Blocks for a Complex-Valued Transformer Architecture" where they propose an an attention mechanism for complex valued signals or images such as MRI and remote sensing.

They present:

- complex valued scaled dot product attention

- complex valued layer normalization

- results show improved robustness to overfitting while maintaing performance wbhen compared to real valued transformer

## Install

`pip install complex-attn`

## Usage

```python

import torch

from ct.attention import ComplexAttention

# # Usage example

dim = 512

heads = 8

seq_len = 512

batch_size = 32

q = torch.randn(batch_size, seq_len, dim) + 1j * torch.randn(

    batch_size, seq_len, dim

)

k = torch.randn(batch_size, seq_len, dim) + 1j * torch.randn(

    batch_size, seq_len, dim

)

v = torch.randn(batch_size, seq_len, dim) + 1j * torch.randn(

    batch_size, seq_len, dim

)

attention_layer = ComplexAttention(

    dim, 

    heads, 

    qk_norm=True,

    dropout=0.1,

)

attn_output = attention_layer(q, k, v)

print("Attention Output Shape:", attn_output.shape)

```

# Architecture

- I use regular norm instead of complex norm for simplicity

# License

MIT

# Citations

```

@article{2306.09827,

Author = {Florian Eilers and Xiaoyi Jiang},

Title = {Building Blocks for a Complex-Valued Transformer Architecture},

Year = {2023},

Eprint = {arXiv:2306.09827},

Doi = {10.1109/ICASSP49357.2023.10095349},

}

```