Attention + Convolution transformer

This is an experimental architecture leveraging convolution blocks with attention blocks to model both the short and long range dynamics of the input tokens. The flow is the following: x -> convolution block -> attn -> FFN

Install

pip3 install --upgrade atc-torch

Usage

import torch
from attnconv.main import ATCTransformer

model = ATCTransformer(
    dim=512,
    depth=6,
    num_tokens=20000,
    dim_head=64,
    heads=8,
    ff_mult=4,
)

x = torch.randint(0, 20000, (1, 512))
logits = model(x)  # (1, 1024, 20000)
print(logits)

Training

python3 train.py

License

MIT