vit-pytorch实现 MobileViT注意力可视化

Posted 2023-03-02 白十月

项目链接 https://github.com/lucidrains/vit-pytorch

注意一下参数设置：

Parameters

1. image_size: int.
   Image size. If you have rectangular images, make sure your image size is the maximum of the width and height
2. patch_size: int.
   Number of patches. image_size must be divisible by patch_size.
   The number of patches is: n = (image_size // patch_size) ** 2 and n must be greater than 16.
3. num_classes: int.
   Number of classes to classify.
4. dim: int.
   Last dimension of output tensor after linear transformation nn.Linear(…, dim).
5. depth: int.
   Number of Transformer blocks.
6. heads: int.
   Number of heads in Multi-head Attention layer.
7. mlp_dim: int.
   Dimension of the MLP (FeedForward) layer.
8. channels: int, default 3.
   Number of image’s channels.
9. dropout: float between [0, 1], default 0…
   Dropout rate.
10. emb_dropout: float between [0, 1], default 0.
    Embedding dropout rate.
11. pool: string, either cls token pooling or mean pooling

image_size：表示图像大小的整数。图片应该是正方形的，并且image_size必须是宽度和高度中的最大值。
patch_size：表示补丁大小的整数。image_size必须能被 整除patch_size。补丁的数量计算为n =
(image_size // patch_size) ** 2并且n必须大于 16。 num_classes：一个整数，表示要分类的类数。
dim：一个整数，表示线性变换后输出张量的最后一维nn.Linear(…, dim)。 depth：一个整数，表示
Transformer 块的数量。 heads：一个整数，表示多头注意力层中的头数。 mlp_dim：一个整数，表示
MLP（前馈）层的维度。 channels：一个整数，表示图像中的通道数，默认值为3。 dropout：一个介于 0 和 1
之间的浮点数，代表辍学率。 emb_dropout：一个介于 0 和 1 之间的浮点数，表示嵌入丢失率。
pool：表示池化方法的字符串，可以是“cls token pooling”或“mean pooling”。

快速使用实例

import torch
from vit_pytorch import ViT

v = ViT(
    image_size = 256,
    patch_size = 32,
    num_classes = 1000,
    dim = 1024,
    depth = 6,
    heads = 16,
    mlp_dim = 2048,
    dropout = 0.1,
    emb_dropout = 0.1
)

img = torch.randn(1, 3, 256, 256)

preds = v(img) # (1, 1000)

SimpleViT

来自原论文的一些作者的更新建议对ViT进行简化，使其能够更快更好地训练。

这些简化包括2d正弦波位置嵌入、全局平均池（无CLS标记）、无辍学、批次大小为1024而不是4096，以及使用RandAugment和MixUp增强。他们还表明，最后的简单线性并不明显比原始MLP头差。

你可以通过导入SimpleViT来使用它，如下图所示

import torch
from vit_pytorch import SimpleViT

v = SimpleViT(
    image_size = 256,
    patch_size = 32,
    num_classes = 1000,
    dim = 1024,
    depth = 6,
    heads = 16,
    mlp_dim = 2048
)

img = torch.randn(1, 3, 256, 256)

preds = v(img) # (1, 1000)

可视化
Accessing Attention
If you would like to visualize the attention weights (post-softmax) for your research, just follow the procedure below

import torch
from vit_pytorch.vit import ViT

v = ViT(
    image_size = 256,
    patch_size = 32,
    num_classes = 1000,
    dim = 1024,
    depth = 6,
    heads = 16,
    mlp_dim = 2048,
    dropout = 0.1,
    emb_dropout = 0.1
)

# import Recorder and wrap the ViT

from vit_pytorch.recorder import Recorder
v = Recorder(v)

# forward pass now returns predictions and the attention maps

img = torch.randn(1, 3, 256, 256)
preds, attns = v(img)

# there is one extra patch due to the CLS token

attns # (1, 6, 16, 65, 65) - (batch x layers x heads x patch x patch)

本文介绍了 MobileViT，一种用于移动设备的轻量级通用视觉转换器。MobileViT 为全球信息处理与转换器提供了不同的视角。

您可以将其与以下代码一起使用（例如 mobilevit_xs）

import torch
from vit_pytorch.mobile_vit import MobileViT

mbvit_xs = MobileViT(
    image_size = (256, 256),
    dims = [96, 120, 144],
    channels = [16, 32, 48, 48, 64, 64, 80, 80, 96, 96, 384],
    num_classes = 1000
)

img = torch.randn(1, 3, 256, 256)

pred = mbvit_xs(img) # (1, 1000)