ConvMixer是基于卷积层进行Mixer操作的模型，结构简单却精度不错。它与MLP Mixer类似，通过交替混合channel和token维度信息提取图像特征，但用卷积替代MLP。其用逐通道卷积提取token信息，1x1卷积提取channel信息，官方提供三个预训练模型，在ImageNet 1k验证集上表现良好，还可从头或微调训练。
☞☞☞AI 智能聊天, 问答助手, AI 智能搜索, 免费无限量使用 DeepSeek R1 模型☜☜☜
ConvMixer 与 MLP Mixer 模型一样模型的结构都十分简单
同样是通过 channel 和 token 两个维度的信息进行交替混合，实现图像特征的有效提取
只不过 ConvMixer 使用的基础网络层为卷积，而 MLP Mixer 使用的是 MLP（多层感知机）
在 ConvMixer 模型中：
使用 Depthwise Convolution（逐通道卷积） 来提取 token 间的相关信息，类似 MLP Mixer 中的 token-mixing MLP
使用 Pointwise Convolution（1x1 卷积） 来提取 channel 间的相关信息，类似 MLP Mixer 中的 channel-mixing MLP
然后将两种卷积交替执行，混合两个维度的信息
模型的大致架构如下图所示：
这是一款可爱的 happy birthday to you 字母矢量素材，提供了 AI 和 EPS 和免扣 PNG 等格式，含 JPG 预览图。
8
import paddle.nn as nnclass Residual(nn.Layer):
# Residual Block（残差层）
# y = f(x) + x
def __init__(self, fn):
super().__init__()
self.fn = fn def forward(self, x):
return self.fn(x) + xdef ConvMixer(dim, depth, kernel_size=9, patch_size=7, act=nn.GELU, n_classes=1000):
# ConvMixer Model
# dim: hidden channal dim（ConvMixer 网络的隐藏层通道数）
# depth: num of ConvMixer Block（网络层数也是其中 ConvMixer 层的数量）
# kernel_size: kernel_size of Convolution in ConvMixer Block（ConvMixer 层中的卷积层的卷积核大小）
# patch_size: patch_size in Patch Embedding （Patch Embedding 时 Patch 的大小）
# act: activate function（激活函数）
# n_classes: num of classes（输出的类别数量）
return nn.Sequential( # Patch Embedding
# Conv(kernel_size = stride = patch_size) + GELU + BN
# 使用一个卷积核大小和步长都等于 Patch 大小的卷积层进行输入图像 Embedding 的操作
# 并连接一个 GELU 激活函数和 BN 批归一化层
nn.Conv2D(3, dim, kernel_size=patch_size, stride=patch_size),
act(),
nn.BatchNorm2D(dim), # ConvMixer Block x N(depth)
# N(depth) 个 ConvMixer 层
*[nn.Sequential( # Residual Block + Depthwise Convolution + GELU + BN
# 逐通道卷积提取 Token 之间的信息
# 并连接一个 GELU 激活函数和 BN 批归一化层
# 最后与原输入进行一个残差连接
Residual(nn.Sequential(
nn.Conv2D(dim, dim, kernel_size, groups=dim, padding="same"),
act(),
nn.BatchNorm2D(dim)
)), # Pointwise Convolution + GELU + BN
# 1x1 卷积提取 Channel 之间的信息
# 并连接一个 GELU 激活函数和 BN 批归一化层
nn.Conv2D(dim, dim, kernel_size=1),
act(),
nn.BatchNorm2D(dim)
) for i in range(depth)], # Output Layers
nn.AdaptiveAvgPool2D((1,1)),
nn.Flatten(),
nn.Linear(dim, n_classes)
)import paddledef convmixer_1536_20(pretrained=False, **kwargs):
model = ConvMixer(1536, 20, kernel_size=9, patch_size=7, **kwargs) if pretrained:
params = paddle.load('/home/aistudio/data/data111600/convmixer_1536_20_ks9_p7.pdparams')
model.set_dict(params) return modeldef convmixer_1024_20(pretrained=False, **kwargs):
model = ConvMixer(1024, 20, kernel_size=9, patch_size=14, **kwargs) if pretrained:
params = paddle.load('/home/aistudio/data/data111600/convmixer_1024_20_ks9_p14.pdparams')
model.set_dict(params) return modeldef convmixer_768_32(pretrained=False, **kwargs):
model = ConvMixer(768, 32, kernel_size=7, patch_size=7, act=nn.ReLU, **kwargs) if pretrained:
params = paddle.load('/home/aistudio/data/data111600/convmixer_768_32_ks7_p7_relu.pdparams')
model.set_dict(params) return modelmodel = convmixer_768_32(pretrained=True) x = paddle.randn((1, 3, 224, 224)) out = model(x)print(out.shape) model.eval() out = model(x)print(out.shape)
ConvMixer 与其他一些先进模型的精度对比：
具体的精度表现如下表：
!mkdir data/ILSVRC2012
!tar -xf ~/data/data68594/ILSVRC2012_img_val.tar -C ~/data/ILSVRC2012
import osimport cv2import numpy as npimport paddleimport paddle.vision.transforms as Tfrom PIL import Image# 构建数据集class ILSVRC2012(paddle.io.Dataset):
def __init__(self, root, label_list, transform, backend='pil'):
self.transform = transform
self.root = root
self.label_list = label_list
self.backend = backend
self.load_datas() def load_datas(self):
self.imgs = []
self.labels = [] with open(self.label_list, 'r') as f: for line in f:
img, label = line[:-1].split(' ')
self.imgs.append(os.path.join(self.root, img))
self.labels.append(int(label)) def __getitem__(self, idx):
label = self.labels[idx]
image = self.imgs[idx] if self.backend=='cv2':
image = cv2.imread(image) else:
image = Image.open(image).convert('RGB')
image = self.transform(image) return image.astype('float32'), np.array(label).astype('int64') def __len__(self):
return len(self.imgs)
val_transforms = T.Compose([
T.Resize(int(224 / 0.96), interpolation='bicubic'),
T.CenterCrop(224),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])# 配置模型model = convmixer_1536_20(pretrained=True)
model = paddle.Model(model)
model.prepare(metrics=paddle.metric.Accuracy(topk=(1, 5)))# 配置数据集val_dataset = ILSVRC2012('data/ILSVRC2012', transform=val_transforms, label_list='data/data68594/val_list.txt', backend='pil')# 模型验证acc = model.evaluate(val_dataset, batch_size=128, num_workers=0, verbose=1)print(acc)Eval begin...
step 391/391 [==============================] - acc_top1: 0.8137 - acc_top5: 0.9562 - 3s/step
Eval samples: 50000
{'acc_top1': 0.81366, 'acc_top5': 0.95616}根据论文的模型配置训练一下 CIFAR-10 数据集的 BaseLine：
由于没有严格对齐各项训练参数，所以训练结果可能应该会有差异
import osimport cv2import numpy as npimport paddleimport paddle.nn as nnimport paddle.vision.transforms as Tfrom paddle.vision.datasets import Cifar10from PIL import Imagefrom paddle.callbacks import EarlyStopping, VisualDL, ModelCheckpoint
train_transforms = T.Compose([
T.Resize(int(224 / 0.96), interpolation='bicubic'),
T.RandomCrop(224),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
val_transforms = T.Compose([
T.Resize(int(224 / 0.96), interpolation='bicubic'),
T.CenterCrop(224),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
model = ConvMixer(256, 8)
opt = paddle.optimizer.Adam(learning_rate=1e-5, parameters=model.parameters())
model = paddle.Model(model)
model.prepare(optimizer=opt, loss=nn.CrossEntropyLoss(), metrics=paddle.metric.Accuracy(topk=(1, 5)))
train_dataset = Cifar10(transform=train_transforms, backend='pil', mode='train')
val_dataset = Cifar10(transform=val_transforms, backend='pil', mode='test')
checkpoint = ModelCheckpoint(save_dir='save')
earlystopping = EarlyStopping(monitor='acc_top1',
mode='max',
patience=3,
verbose=1,
min_delta=0,
baseline=None,
save_best_model=True)
vdl = VisualDL('log')
model.fit(train_dataset, val_dataset, batch_size=32, num_workers=0, epochs=10, save_dir='save', callbacks=[checkpoint, earlystopping, vdl], verbose=1)import osimport cv2import numpy as npimport paddleimport paddle.nn as nnimport paddle.vision.transforms as Tfrom paddle.vision.datasets import Cifar10from PIL import Imagefrom paddle.callbacks import EarlyStopping, VisualDL, ModelCheckpoint
train_transforms = T.Compose([
T.Resize(int(224 / 0.96), interpolation='bicubic'),
T.RandomCrop(224),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
val_transforms = T.Compose([
T.Resize(int(224 / 0.96), interpolation='bicubic'),
T.CenterCrop(224),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
model = convmixer_768_32(n_classes=10, pretrained=True)
opt = paddle.optimizer.Adam(learning_rate=1e-5, parameters=model.parameters())
model = paddle.Model(model)
model.prepare(optimizer=opt, loss=nn.CrossEntropyLoss(), metrics=paddle.metric.Accuracy(topk=(1, 5)))
train_dataset = Cifar10(transform=train_transforms, backend='pil', mode='train')
val_dataset = Cifar10(transform=val_transforms, backend='pil', mode='test')
checkpoint = ModelCheckpoint(save_dir='save')
earlystopping = EarlyStopping(monitor='acc_top1',
mode='max',
patience=3,
verbose=1,
min_delta=0,
baseline=None,
save_best_model=True)
vdl = VisualDL('log')
model.fit(train_dataset, val_dataset, batch_size=32, num_workers=0, epochs=1, save_dir='save', callbacks=[checkpoint, earlystopping, vdl], verbose=1)以上就是ConvMixer：Patches are all you need?的详细内容，更多请关注php中文网其它相关文章！
每个人都需要一台速度更快、更稳定的 PC。随着时间的推移，垃圾文件、旧注册表数据和不必要的后台进程会占用资源并降低性能。幸运的是，许多工具可以让 Windows 保持平稳运行。
Copyright 2014-2025 https://www.php.cn/ All Rights Reserved | php.cn | 湘ICP备2023035733号
195
收藏
