git倉(cāng)庫(kù)地址：https://github.com/The-AI-Summer/self-attention-cv

用einsum和einops在PyTorch中實(shí)現(xiàn)計(jì)算機(jī)視覺的自我注意機(jī)制。專注于計(jì)算機(jī)視覺自注意模塊。

使用 pip 安裝

$ pip install self-attention-cv

如果你沒有GPU，最好是在環(huán)境中預(yù)裝好pytorch。

相關(guān)的文章

• How Attention works in Deep Learning
• How Transformers work in deep learning and NLP
• How the Vision Transformer (ViT) works in 10 minutes: an image is worth 16x16 words
• Understanding einsum for Deep learning: implement a transformer with multi-head self-attention from scratch
• How Positional Embeddings work in Self-Attention

示例代碼

Multi-head attention

import torch
from self_attention_cv import MultiHeadSelfAttention

model = MultiHeadSelfAttention(dim=64)
x = torch.rand(16, 10, 64)  # [batch, tokens, dim]
mask = torch.zeros(10, 10)  # tokens X tokens
mask[5:8, 5:8] = 1
y = model(x, mask)

Axial attention

import torch
from self_attention_cv import AxialAttentionBlock
model = AxialAttentionBlock(in_channels=256, dim=64, heads=8)
x = torch.rand(1, 256, 64, 64)  # [batch, tokens, dim, dim]
y = model(x)

Vanilla Transformer Encoder

import torch
from self_attention_cv import TransformerEncoder
model = TransformerEncoder(dim=64,blocks=6,heads=8)
x = torch.rand(16, 10, 64)  # [batch, tokens, dim]
mask = torch.zeros(10, 10)  # tokens X tokens
mask[5:8, 5:8] = 1
y = model(x,mask)

Vision Transformer使用ResNet50主干做圖像分類

import torch
from self_attention_cv import ViT, ResNet50ViT

model1 = ResNet50ViT(img_dim=128, pretrained_resnet=False, 
                        blocks=6, num_classes=10, 
                        dim_linear_block=256, dim=256)
# or
model2 = ViT(img_dim=256, in_channels=3, patch_dim=16, num_classes=10,dim=512)
x = torch.rand(2, 3, 256, 256)
y = model2(x) # [2,10]

使用Vision Transformer編碼器的Unet的復(fù)現(xiàn)

import torch
from self_attention_cv.transunet import TransUnet
a = torch.rand(2, 3, 128, 128)
model = TransUnet(in_channels=3, img_dim=128, vit_blocks=8,
vit_dim_linear_mhsa_block=512, classes=5)
y = model(a) # [2, 5, 128, 128]

Bottleneck Attention block

import torch
from self_attention_cv.bottleneck_transformer import BottleneckBlock
inp = torch.rand(1, 512, 32, 32)
bottleneck_block = BottleneckBlock(in_channels=512, fmap_size=(32, 32), heads=4, out_channels=1024, pooling=True)
y = bottleneck_block(inp)

位置嵌入可用

1D Positional Embeddings

import torch
from self_attention_cv.pos_embeddings import AbsPosEmb1D,RelPosEmb1D

model = AbsPosEmb1D(tokens=20, dim_head=64)
# batch heads tokens dim_head
q = torch.rand(2, 3, 20, 64)
y1 = model(q)

model = RelPosEmb1D(tokens=20, dim_head=64, heads=3)
q = torch.rand(2, 3, 20, 64)
y2 = model(q)

2D Positional Embeddings

import torch
from self_attention_cv.pos_embeddings import RelPosEmb2D
dim = 32  # spatial dim of the feat map
model = RelPosEmb2D(
    feat_map_size=(dim, dim),
    dim_head=128)

q = torch.rand(2, 4, dim*dim, 128)
y = model(q)

參考文獻(xiàn)

1. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... & Polosukhin, I. (2017). Attention is all you need. arXiv preprint arXiv:1706.03762.
2. Wang, H., Zhu, Y., Green, B., Adam, H., Yuille, A., & Chen, L. C. (2020, August). Axial-deeplab: Stand-alone axial-attention for panoptic segmentation. In European Conference on Computer Vision (pp. 108-126). Springer, Cham.
3. Srinivas, A., Lin, T. Y., Parmar, N., Shlens, J., Abbeel, P., & Vaswani, A. (2021). Bottleneck Transformers for Visual Recognition. arXiv preprint arXiv:2101.11605.
4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., ... & Houlsby, N. (2020). An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929.