Uformer - Pytorch
Implementation of Uformer, Attention-based Unet, in Pytorch. It will only offer the concat-cross-skip connection.
This repository will be geared towards use in a project for learning protein structures. Specifically, it will include the ability to condition on time steps (needed for DDPM), as well as 2d relative positional encoding using rotary embeddings (instead of the bias on the attention matrix in the paper).
Install
$ pip install uformer-pytorch
Usage
import torch
from uformer_pytorch import Uformer
model = Uformer(
dim = 64, # initial dimensions after input projection, which increases by 2x each stage
stages = 4, # number of stages
num_blocks = 2, # number of transformer blocks per stage
window_size = 16, # set window size (along one side) for which to do the attention within
dim_head = 64,
heads = 8,
ff_mult = 4
)
x = torch.randn(1, 3, 256, 256)
pred = model(x) # (1, 3, 256, 256)
To condition on time for DDPM training
import torch
from uformer_pytorch import Uformer
model = Uformer(
dim = 64,
stages = 4,
num_blocks = 2,
window_size = 16,
dim_head = 64,
heads = 8,
ff_mult = 4,
time_emb = True # set this to true
)
x = torch.randn(1, 3, 256, 256)
time = torch.arange(1)
pred = model(x, time = time) # (1, 3, 256, 256)
Citations
@misc{wang2021uformer,
title = {Uformer: A General U-Shaped Transformer for Image Restoration},
author = {Zhendong Wang and Xiaodong Cun and Jianmin Bao and Jianzhuang Liu},
year = {2021},
eprint = {2106.03106},
archivePrefix = {arXiv},
primaryClass = {cs.CV}
}
84 Nov 22, 2022
1 Oct 21, 2021
4 Dec 21, 2021
1.8k Jan 7, 2023
3 Jun 29, 2022
556 Jan 4, 2023
71 Dec 30, 2022
5 Nov 21, 2022
0 Sep 2, 2021
1 Nov 22, 2021
129 Dec 22, 2022
164 Dec 05, 2022
175 Jan 07, 2023
3 Oct 29, 2022
31 Nov 25, 2022
1 Nov 30, 2021
38 Dec 12, 2022
282 Jan 09, 2023
2 Jan 05, 2022
21 Jul 27, 2022
14 Nov 22, 2022
87 Nov 30, 2022
2 Dec 06, 2022
8 Jun 30, 2022
1 Nov 13, 2021
32 Jan 02, 2023
523 Jan 05, 2023
118 Dec 24, 2022
19 Dec 06, 2022