★★★ 本文源自AlStudio社区精品项目，【点击此处】查看更多精品内容 >>>

1.摘要

近年来，有大量的卷积模型通过堆叠不同感受野的卷积以及采用金字塔结构的网络模型提取特征，但这些模型往往忽视了全局信息的提取。直到vision transformer的提出，首次将transformer引入视觉领域，并在全局信息建模展现了更好的性能，但不可忽视的是transformer在处理高分辨率图片时会产生大量的计算开销。最近，ConvNeXt,在传统残差结构的基础上，使用了更为先进的训练技巧，使传统卷积的性能可以和ViT不相上下，这让我们重新思考能否设计一种全新的结构可以大幅减低计算开销的同时，有着transformer一样的全局特征提取的能力，Conv2Former使用了transformer一样的QKV结构，但采用卷积生成权重加权，为我们进一步设计卷积模型提供了一种思路。

![image.png]

!mkdir /home/aistudio/Conv2Former-libraries
!pip install paddlex -t /home/aistudio/Conv2Former-libraries

import paddle
import numpy as np
from paddle.vision.datasets import Cifar10
from paddle.vision.transforms import Transpose
from paddle.io import Dataset, DataLoader
from paddle import nn
import paddle.nn.functional as F
import paddle.vision.transforms as transforms
import os
#import matplotlib.pyplot as plt
#from matplotlib.pyplot import figure
import sys 
sys.path.append('/home/aistudio/Conv2Former-libraries')
import paddlex

一些训练tricks,labelsoomthing and droppath.

class LabelSmoothingCrossEntropy(nn.Layer):
    def __init__(self, smoothing=0.1):
        super().__init__()
        self.smoothing = smoothing

    def forward(self, pred, target):

        confidence = 1. - self.smoothing
        log_probs = F.log_softmax(pred, axis=-1)
        idx = paddle.stack([paddle.arange(log_probs.shape[0]), target], axis=1)
        nll_loss = paddle.gather_nd(-log_probs, index=idx)
        smooth_loss = paddle.mean(-log_probs, axis=-1)
        loss = confidence * nll_loss + self.smoothing * smooth_loss

        return loss.mean()

def drop_path(x, drop_prob=0.0, training=False):
        """
        Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
        the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
        See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ...
        """
        if drop_prob == 0.0 or not training:
            return x
        keep_prob = paddle.to_tensor(1 - drop_prob)
        shape = (paddle.shape(x)[0],) + (1,) * (x.ndim - 1)
        random_tensor = keep_prob + paddle.rand(shape, dtype=x.dtype)
        random_tensor = paddle.floor(random_tensor)  # binarize
        output = x.divide(keep_prob) * random_tensor
        return output

class DropPath(nn.Layer):
        def __init__(self, drop_prob=None):
            super(DropPath, self).__init__()
            self.drop_prob = drop_prob

        def forward(self, x):
            return drop_path(x, self.drop_prob, self.training)

2.数据载入及增强

（数据集：cifar-10）
作者采用了一些常见的数据增强方式（未完全复现）：MixUp、CutMix、Stochastic Depth、 Random Erasing 、Label Smoothing、RandAug 、Layer Scale

train_tfm = transforms.Compose([
    transforms.Resize((32,32)),
    transforms.ColorJitter(brightness=0.2,contrast=0.2, saturation=0.2),
    paddlex.transforms.MixupImage(),
    #transforms.Cutmix(),
    transforms.RandomResizedCrop(32, scale=(0.6, 1.0)),
    transforms.RandomErasing(),
    transforms.RandomHorizontalFlip(0.5),
    transforms.RandomRotation(20),
    transforms.ToTensor(),
    transforms.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
])

test_tfm = transforms.Compose([
    transforms.Resize((32,32)),
    transforms.ToTensor(),
    transforms.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
])

batch_size=256

paddle.vision.set_image_backend('cv2')
# 使用Cifar10数据集
train_dataset = Cifar10(data_file='./data/cifar-10-python.tar.gz', mode='train', transform = train_tfm,)
val_dataset = Cifar10(data_file='./data/cifar-10-python.tar.gz', mode='test',transform = test_tfm)
print("train_dataset: %d" % len(train_dataset))
print("val_dataset: %d" % len(val_dataset))


train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, drop_last=True, num_workers=2)
val_loader = DataLoader(val_dataset, batch_size=batch_size, shuffle=False, drop_last=False, num_workers=2)

train_dataset: 50000
val_dataset: 10000

0

3.模型创建

3.1 Conv2Former模块创建

由于显存以及训练条件限制，我们将原文设计的224乘224的输入改为32乘32输入，并采用Conv2Former-N的模型参数进行堆叠，即{C1, C2, C3, C4}={64, 128, 256, 512}；{L1, L2, L3, L4}={2, 2, 8, 2}

class MLP(nn.Layer):
    def __init__(self, dim, mlp_ratio=4,  drop=0.,):
        super().__init__()

        self.norm = nn.LayerNorm(dim, epsilon=1e-6,)

        self.fc1 = nn.Conv2D(dim, dim * mlp_ratio, 1)
        self.pos = nn.Conv2D(dim * mlp_ratio, dim * mlp_ratio, 3, padding=1, groups=dim * mlp_ratio)
        self.fc2 = nn.Conv2D(dim * mlp_ratio, dim, 1)
        self.act = nn.GELU()
        self.drop = nn.Dropout(drop)

    def forward(self, x):
        B, C, H, W = x.shape

        x = self.norm(x.transpose([0, 2, 3, 1])).transpose([0, 3, 1, 2])
        x = self.fc1(x)
        x = self.act(x)
        x = x + self.act(self.pos(x))
        x = self.fc2(x)

        return x

class ConvMod(nn.Layer):
    def __init__(self, dim):
        super().__init__()

        self.norm = nn.LayerNorm(dim, epsilon=1e-6,)
        self.a = nn.Sequential(
            nn.Conv2D(dim, dim, 1),
            nn.GELU(),
            nn.Conv2D(dim, dim, 11, padding=5, groups=dim)
        )

        self.v = nn.Conv2D(dim, dim, 1)
        self.proj = nn.Conv2D(dim, dim, 1)

    def forward(self, x):
        B, C, H, W = x.shape

        x = self.norm(x.transpose([0, 2, 3, 1])).transpose([0, 3, 1, 2])
        a = self.a(x)
        x = a * self.v(x)
        x = self.proj(x)

        return x

3.2Convolutional modulation
作者在此处采用了11乘11的大卷积核，作者通过实验，发现Conv2Former在卷积核大小进一步增大时，性能可以进一步加强，故最终将卷积核大小设置为11乘11。也许是因为这么大的感受野最后赋予了模型更强的全局信息获取能力。

class Block(nn.Layer):
    def __init__(self, dim, mlp_ratio=4, drop=0., drop_path=0.,):
        super().__init__()

        self.attn = ConvMod(dim)
        self.mlp = MLP(dim, mlp_ratio, drop=drop)
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()

    def forward(self, x):
        x = x + self.drop_path(self.attn(x))
        x = x + self.drop_path(self.mlp(x))
        return x

class BasicLayer(nn.Layer):

    def __init__(self, dim, depth, mlp_ratio=4., drop=0., drop_path=0.,downsample=True):

        super(BasicLayer, self).__init__()
        self.dim = dim
        self.drop_path = drop_path

        # build blocks
        self.blocks = nn.LayerList([
            Block(dim=dim, mlp_ratio=mlp_ratio, drop=drop, drop_path=drop_path[i],)
            for i in range(depth)
        ])

        # patch merging layer
        if downsample:
            self.downsample = nn.Sequential(
                nn.GroupNorm(num_groups=1, num_channels=dim),
                nn.Conv2D(dim, dim * 2, kernel_size=2, stride=2,bias_attr=False)
            )
        else:
            self.downsample = None

    def forward(self, x):
        for blk in self.blocks:
            x = blk(x)
        if self.downsample is not None:
            x = self.downsample(x)
        return x

class Conv2Former(nn.Layer):

    def __init__(self, num_classes=10, depths=(2,2,8,2), dim=(64,128,256,512), mlp_ratio=2.,drop_rate=0.,
                 drop_path_rate=0.15, **kwargs):
        super().__init__()

        norm_layer = nn.LayerNorm
        self.num_classes = num_classes
        self.num_layers = len(depths)
        self.dim = dim
        self.mlp_ratio = mlp_ratio
        self.pos_drop = nn.Dropout(p=drop_rate)

        # stochastic depth decay rule
        dpr = [x.item()
               for x in paddle.linspace(0, drop_path_rate, sum(depths))]

        # build layers
        self.layers = nn.LayerList()
        for i_layer in range(self.num_layers):
            layer = BasicLayer(dim[i_layer],
                               depth=depths[i_layer],
                               mlp_ratio=self.mlp_ratio,
                               drop=drop_rate,
                               drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
                               downsample=(i_layer < self.num_layers - 1),
                               )
            self.layers.append(layer)
        self.fc1 = nn.Conv2D(3, 64, 1)
        self.norm = norm_layer(512, epsilon=1e-6,)
        self.avgpool = nn.AdaptiveAvgPool2D(1)
        self.head = nn.Linear(512, num_classes) \
            if num_classes > 0 else nn.Identity()

        self.apply(self._init_weights)
        
    def _init_weights(self, m):
        tn = nn.initializer.TruncatedNormal(std=.02)
        zeros = nn.initializer.Constant(0.)
        ones = nn.initializer.Constant(1.)
        if isinstance(m, nn.Linear):
            tn(m.weight)
            if isinstance(m, nn.Linear) and m.bias is not None:
                zeros(m.bias)
        elif isinstance(m, (nn.Conv1D, nn.Conv2D)):
            tn(m.weight)
            if m.bias is not None:
                zeros(m.bias)
        elif isinstance(m, (nn.LayerNorm, nn.GroupNorm)):
            zeros(m.bias)
            ones(m.weight)

    def forward_features(self, x):

        x = self.fc1(x)
        x = self.pos_drop(x)

        for layer in self.layers:
            x = layer(x)

        x = self.norm(x.transpose([0, 2, 3, 1]))
        x = x.transpose([0, 3, 1, 2])
        x = self.avgpool(x)
        x = paddle.flatten(x, 1)
        return x

    def forward(self, x):
        x = self.forward_features(x)
        x = self.head(x)
        return x

#参数设置
learning_rate = 0.001
n_epochs = 50
paddle.seed(42)
np.random.seed(42)
batch_size = 256

work_path = './work/model'

# conv2Former模型打印
model = Conv2Former(num_classes=10, depths=(2,2,8,2),dim=(64,128,256,512), mlp_ratio=2,drop_path_rate=0.1)
params_info=paddle.summary(model,input_size=(1, 3, 32, 32))
print(params_info)

criterion = LabelSmoothingCrossEntropy()
scheduler = paddle.optimizer.lr.CosineAnnealingDecay(learning_rate=learning_rate, T_max=50000 // batch_size * n_epochs,
                                                     verbose=False)
optimizer = paddle.optimizer.Adam(parameters=model.parameters(), learning_rate=scheduler, weight_decay=1e-5)

gate = 0.0
threshold = 0.0
best_acc = 0.0
val_acc = 0.0
loss_record = {'train': {'loss': [], 'iter': []}, 'val': {'loss': [], 'iter': []}}  # for recording loss
acc_record = {'train': {'acc': [], 'iter': []}, 'val': {'acc': [], 'iter': []}}  # for recording accuracy

loss_iter = 0
acc_iter = 0

for epoch in range(n_epochs):
    # ---------- Training set----------
    model.train()
    train_num = 0.0
    train_loss = 0.0

    val_num = 0.0
    val_loss = 0.0
    accuracy_manager = paddle.metric.Accuracy()
    val_accuracy_manager = paddle.metric.Accuracy()
    print("#===epoch: {}, lr={:.10f}===#".format(epoch, optimizer.get_lr()))
    for batch_id, data in enumerate(train_loader):
        x_data, y_data = data
        labels = paddle.unsqueeze(y_data, axis=1)

        logits = model(x_data)

        loss = criterion(logits, y_data)

        acc = paddle.metric.accuracy(logits, labels)
        accuracy_manager.update(acc)
        if batch_id % 10 == 0:
            loss_record['train']['loss'].append(loss.numpy())
            loss_record['train']['iter'].append(loss_iter)
            loss_iter += 1

        loss.backward()

        optimizer.step()
        scheduler.step()
        optimizer.clear_grad()

        train_loss += loss
        train_num += len(y_data)

    total_train_loss = (train_loss / train_num) * batch_size
    train_acc = accuracy_manager.accumulate()
    acc_record['train']['acc'].append(train_acc)
    acc_record['train']['iter'].append(acc_iter)
    acc_iter += 1
    # Print the information.
    print("#===epoch: {}, train loss is: {}, train acc is: {:2.2f}%===#".format(epoch, total_train_loss.numpy(),
                                                                                train_acc * 100))

 # ---------- Validation ----------
    model.eval()

    for batch_id, data in enumerate(val_loader):
        x_data, y_data = data
        labels = paddle.unsqueeze(y_data, axis=1)
        with paddle.no_grad():
            logits = model(x_data)

        loss = criterion(logits, y_data)

        acc = paddle.metric.accuracy(logits, labels)
        val_accuracy_manager.update(acc)

        val_loss += loss
        val_num += len(y_data)

    total_val_loss = (val_loss / val_num) * batch_size
    loss_record['val']['loss'].append(total_val_loss.numpy())
    loss_record['val']['iter'].append(loss_iter)
    val_acc = val_accuracy_manager.accumulate()
    acc_record['val']['acc'].append(val_acc)
    acc_record['val']['iter'].append(acc_iter)

    print(
        "#===epoch: {}, val loss is: {}, val acc is: {:2.2f}%===#".format(epoch, total_val_loss.numpy(), val_acc * 100))

    # ===================save====================
    if val_acc > best_acc:
        best_acc = val_acc
        paddle.save(model.state_dict(), os.path.join(work_path, 'best_model.pdparams'))
        paddle.save(optimizer.state_dict(), os.path.join(work_path, 'best_optimizer.pdopt'))

print(best_acc)
paddle.save(model.state_dict(), os.path.join(work_path, 'final_model.pdparams'))
paddle.save(optimizer.state_dict(), os.path.join(work_path, 'final_optimizer.pdopt'))

## 4.结论与讨论

4.1结论

本项目通过展现Conv2Former论文中的网络结构，对Conv2Former-N在飞桨框架下完成复现并进行初步训练，在没有预训练的基础上，对在50个epoch训练以后，模型在验证集上的准确率显著提升，在Cifar-10数据集上产生了有一定竞争力的结果，这证明了Conv2Former的模块设计具有一定的优越性，能够在大幅减少计算负担的同时，提升模型性能，同时，也为transformer的可解释性以及卷积模块的重新设计提供了新的思路。

Model	Parameter	Val Acc
Conv2Former-N	8,847,978	0.82
Swin-T	27,527,044	0.75

注：Swin-T实验结果来自浅析 Swin Transformer，模型为swin_tiny。

4.2不足与展望

1. 可以复现Conv2Former-B、Conv2Former-L等更深层次，更大参数量的模型，并且在大参数量模型下，可以通过通过在image-21k上进行预训练，后在目标数据集上进行微调，获得更好的模型性能。
2. 可以将模型应用在语义分割、目标检测等下游任务中测试性能。

此文章为搬运
原项目链接