【CVPR2023】FasterNet：追逐更高FLOPS、更快的神经网络

FasterNet聚焦于提升神经网络速度，指出仅减少FLOPs未必降低延迟，关键在于提高每秒浮点运算（FLOPS）。其提出部分卷积（PConv），减少冗余计算与内存访问。基于此构建的FasterNet在多设备上速度更快，且精度不俗，如微型版比MobileVit - XXS快数倍且精度更高，大型版准确率高且吞吐量提升。

FasterNet：追逐更高FLOPS、更快的神经网络

摘要

        为了设计快速神经网络，许多工作都集中在减少浮点运算的数量（FLOPs）上。 然而，我们观察到FLOPs的减少并不一定会导致延迟的类似程度的减少。 这主要源于低效率的每秒浮点运算(FLOPS)。 为了实现更快的网络，我们回顾了流行的操作，并证明如此低的FLOPS主要是由于操作频繁的内存访问，特别是深度卷积。 因此，我们提出了一种新的部分卷积(PConv)，通过同时减少冗余计算和内存访问，可以更有效地提取空间特征。 在Ponv的基础上，我们进一步提出了FasterNet，这是一个新的神经网络家族，它在各种设备上获得了比其他网络更高的运行速度，而不影响各种视觉任务的准确性。 例如，在ImageNet1K上，我们的微型FasterNet-T0在GPU、CPU和ARM处理器上分别比MobileVit-XXS块3.1×、3.1×和2.5×，同时精度提高2.9%。 我们的大型FasterNet-L实现了令人印象深刻的83.5%的Top-1准确率，与新兴的Swin-B不相上下，同时在GPU上提高了49%的推断吞吐量，并在CPU上节省了42%的计算时间。

1. FasterNet

        本文思考了一个问题：怎样才能更快？之前的工作大多使用FLOPs来表示神经网络的快慢，但是某些操作（如DWConv）实际运行并不快，这主要是因为频繁的内存访问。本文提出了新的见解：设计一个低FLOPs高FLOPS的操作，这样可以加快网络运行速度。由此，本文作者提出了一个“T型”的卷积——PConv，主要思想是DWConv虽然FLOPs小，但是由于频繁的内存访问导致FLOPS也小。由于网络存在冗余通道，那我是不是可以设计一个网络只用一部分去做空间计算，作者就尝试了这一想法，发现效果非常好，速度快，精度高。具体的操作如图5所示：

       

        基于PConv和传统的分层Transformer，本文提出了一个新的网络架构——FasterNet，结构图如图4所示：

       

2. 代码复现

2.1 下载并导入所需的库

In [ ]

!pip install paddlex

登录后复制    In [ ]

%matplotlib inlineimport paddleimport paddle.fluid as fluidimport numpy as npimport matplotlib.pyplot as pltfrom paddle.vision.datasets import Cifar10from paddle.vision.transforms import Transposefrom paddle.io import Dataset, DataLoaderfrom paddle import nnimport paddle.nn.functional as Fimport paddle.vision.transforms as transformsimport osimport matplotlib.pyplot as pltfrom matplotlib.pyplot import figureimport paddleximport mathimport itertools

登录后复制    

2.2 创建数据集

In [3]

train_tfm = transforms.Compose([    transforms.RandomResizedCrop(224),    transforms.ColorJitter(brightness=0.2,contrast=0.2, saturation=0.2),    transforms.RandomHorizontalFlip(0.5),    transforms.RandomRotation(20),    paddlex.transforms.MixupImage(),    transforms.ToTensor(),    transforms.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),])test_tfm = transforms.Compose([    transforms.Resize((224, 224)),    transforms.ToTensor(),    transforms.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),])

登录后复制    In [4]

paddle.vision.set_image_backend('cv2')# 使用Cifar10数据集train_dataset = Cifar10(data_file='data/data152754/cifar-10-python.tar.gz', mode='train', transform = train_tfm)val_dataset = Cifar10(data_file='data/data152754/cifar-10-python.tar.gz', mode='test',transform = test_tfm)print("train_dataset: %d" % len(train_dataset))print("val_dataset: %d" % len(val_dataset))

登录后复制        

train_dataset: 50000val_dataset: 10000

登录后复制        In [5]

batch_size=256

登录后复制    In [6]

train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, drop_last=True, num_workers=4)val_loader = DataLoader(val_dataset, batch_size=batch_size, shuffle=False, drop_last=False, num_workers=4)

登录后复制    

2.3 模型的创建

2.3.1 标签平滑

In [7]

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()

登录后复制    

2.3.2 DropPath

In [8]

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 outputclass 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)

登录后复制    

2.3.3 FasterNet模型的创建

In [9]

class PConv(nn.Layer):    def __init__(self, dim, kernel_size=3, n_div=4):        super().__init__()                self.dim_conv = dim // n_div        self.dim_untouched = dim - self.dim_conv        self.conv = nn.Conv2D(self.dim_conv, self.dim_conv, kernel_size, padding=(kernel_size - 1) // 2, bias_attr=False)    def forward(self, x):                x1, x2 = paddle.split(x, [self.dim_conv, self.dim_untouched], axis=1)        x1 = self.conv(x1)        x = paddle.concat([x1, x2], axis=1)        return x

登录后复制    In [10]

class FasterNetBlock(nn.Layer):    def __init__(self, dim, expand_ratio=2, act_layer=nn.ReLU, drop_path_rate=0.0):        super().__init__()        self.pconv = PConv(dim)        self.conv1 = nn.Conv2D(dim, dim * expand_ratio, 1, bias_attr=False)        self.bn = nn.BatchNorm2D(dim * expand_ratio)        self.act_layer = act_layer()        self.conv2 = nn.Conv2D(dim * expand_ratio, dim, 1, bias_attr=False)        self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0.0 else nn.Identity()    def forward(self, x):        residual = x        x = self.pconv(x)        x = self.conv1(x)        x = self.bn(x)        x = self.act_layer(x)        x = self.conv2(x)        x = residual + self.drop_path(x)        return x

登录后复制    In [11]

class FasterNet(nn.Layer):    def __init__(self, in_channel=3, embed_dim=40, act_layer=nn.ReLU, num_classes=1000, depths=[1, 2, 8, 2], drop_path=0.0):        super().__init__()        self.stem = nn.Sequential(            nn.Conv2D(in_channel, embed_dim, 4, stride=4, bias_attr=False),            nn.BatchNorm2D(embed_dim),            act_layer()        )        drop_path_list = [x.item() for x in paddle.linspace(0, drop_path, sum(depths))]        self.feature = []        embed_dim = embed_dim        for idx, depth in enumerate(depths):                        self.feature.append(nn.Sequential(                *[FasterNetBlock(embed_dim, act_layer=act_layer, drop_path_rate=drop_path_list[sum(depths[:idx]) + i]) for i in range(depth)]            ))            if idx < len(depths) - 1:                self.feature.append(nn.Sequential(                    nn.Conv2D(embed_dim, embed_dim * 2, 2, stride=2, bias_attr=False),                    nn.BatchNorm2D(embed_dim * 2),                    act_layer()                ))                embed_dim = embed_dim * 2                self.feature = nn.Sequential(*self.feature)        self.avg_pool = nn.AdaptiveAvgPool2D(1)                self.conv1 = nn.Conv2D(embed_dim, 1280, 1, bias_attr=False)        self.act_layer = act_layer()        self.fc = nn.Linear(1280, num_classes)    def forward(self, x):        x = self.stem(x)        x = self.feature(x)                x = self.avg_pool(x)        x = self.conv1(x)        x = self.act_layer(x)        x = self.fc(x.flatten(1))        return x

登录后复制    In [12]

def fasternet_t0():    num_classes=10    embed_dim = 40    depths = [1, 2, 8, 2]    drop_path_rate = 0.0    act_layer = nn.GELU    return FasterNet(embed_dim=embed_dim, act_layer=act_layer, num_classes=num_classes, depths=depths, drop_path=drop_path_rate)def fasternet_t1():    num_classes=10    embed_dim = 64    depths = [1, 2, 8, 2]    drop_path_rate = 0.02    act_layer = nn.GELU    return FasterNet(embed_dim=embed_dim, act_layer=act_layer, num_classes=num_classes, depths=depths, drop_path=drop_path_rate)def fasternet_t2():    num_classes=10    embed_dim = 96    depths = [1, 2, 8, 2]    drop_path_rate = 0.05    act_layer = nn.ReLU    return FasterNet(embed_dim=embed_dim, act_layer=act_layer, num_classes=num_classes, depths=depths, drop_path=drop_path_rate)def fasternet_s():    num_classes=10    embed_dim = 128    depths = [1, 2, 13, 2]    drop_path_rate = 0.03    act_layer = nn.ReLU    return FasterNet(embed_dim=embed_dim, act_layer=act_layer, num_classes=num_classes, depths=depths, drop_path=drop_path_rate)def fasternet_m():    num_classes=10    embed_dim = 144    depths = [3, 4, 18, 3]    drop_path_rate = 0.05    act_layer = nn.ReLU    return FasterNet(embed_dim=embed_dim, act_layer=act_layer, num_classes=num_classes, depths=depths, drop_path=drop_path_rate)def fasternet_l():    num_classes=10    embed_dim = 192    depths = [3, 4, 18, 3]    drop_path_rate = 0.05    act_layer = nn.ReLU    return FasterNet(embed_dim=embed_dim, act_layer=act_layer, num_classes=num_classes, depths=depths, drop_path=drop_path_rate)

登录后复制    

2.3.4 模型的参数

In [ ]

model = fasternet_t0()paddle.summary(model, (1, 3, 224, 224))

登录后复制    

       

In [ ]

model = fasternet_t1()paddle.summary(model, (1, 3, 224, 224))

登录后复制    

       

In [ ]

model = fasternet_t2()paddle.summary(model, (1, 3, 224, 224))

登录后复制    

       

In [ ]

model = fasternet_s()paddle.summary(model, (1, 3, 224, 224))

登录后复制    

       

In [ ]

model = fasternet_m()paddle.summary(model, (1, 3, 224, 224))

登录后复制    

       

In [ ]

model = fasternet_l()paddle.summary(model, (1, 3, 224, 224))

登录后复制    

       

2.4 训练

In [19]

learning_rate = 0.001n_epochs = 100paddle.seed(42)np.random.seed(42)

登录后复制    In [ ]

work_path = 'work/model'# FasterNet-T0model = fasternet_t0()criterion = LabelSmoothingCrossEntropy()scheduler = paddle.optimizer.lr.CosineAnnealingDecay(learning_rate=learning_rate, T_max=50000 // batch_size * n_epochs, verbose=False)optimizer = paddle.optimizer.AdamW(parameters=model.parameters(), learning_rate=scheduler, weight_decay=0.005)gate = 0.0threshold = 0.0best_acc = 0.0val_acc = 0.0loss_record = {'train': {'loss': [], 'iter': []}, 'val': {'loss': [], 'iter': []}}   # for recording lossacc_record = {'train': {'acc': [], 'iter': []}, 'val': {'acc': [], 'iter': []}}      # for recording accuracyloss_iter = 0acc_iter = 0for epoch in range(n_epochs):    # ---------- Training ----------    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'))

登录后复制    

       

2.5 结果分析

In [21]

def plot_learning_curve(record, title="loss", ylabel='CE Loss'):    ''' Plot learning curve of your CNN '''    maxtrain = max(map(float, record['train'][title]))    maxval = max(map(float, record['val'][title]))    ymax = max(maxtrain, maxval) * 1.1    mintrain = min(map(float, record['train'][title]))    minval = min(map(float, record['val'][title]))    ymin = min(mintrain, minval) * 0.9    total_steps = len(record['train'][title])    x_1 = list(map(int, record['train']['iter']))    x_2 = list(map(int, record['val']['iter']))    figure(figsize=(10, 6))    plt.plot(x_1, record['train'][title], c='tab:red', label='train')    plt.plot(x_2, record['val'][title], c='tab:cyan', label='val')    plt.ylim(ymin, ymax)    plt.xlabel('Training steps')    plt.ylabel(ylabel)    plt.title('Learning curve of {}'.format(title))    plt.legend()    plt.show()

登录后复制    In [22]

plot_learning_curve(loss_record, title="loss", ylabel='CE Loss')

登录后复制        

登录后复制登录后复制                In [23]

plot_learning_curve(acc_record, title="acc", ylabel='Accuracy')

登录后复制        

登录后复制登录后复制                In [24]

import timework_path = 'work/model'model = fasternet_t0()model_state_dict = paddle.load(os.path.join(work_path, 'best_model.pdparams'))model.set_state_dict(model_state_dict)model.eval()aa = time.time()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)bb = time.time()print("Throughout:{}".format(int(len(val_dataset)//(bb - aa))))

登录后复制        

Throughout:982

登录后复制        In [25]

def get_cifar10_labels(labels):      """返回CIFAR10数据集的文本标签。"""    text_labels = [        'airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog',        'horse', 'ship', 'truck']    return [text_labels[int(i)] for i in labels]

登录后复制    In [26]

def show_images(imgs, num_rows, num_cols, pred=None, gt=None, scale=1.5):      """Plot a list of images."""    figsize = (num_cols * scale, num_rows * scale)    _, axes = plt.subplots(num_rows, num_cols, figsize=figsize)    axes = axes.flatten()    for i, (ax, img) in enumerate(zip(axes, imgs)):        if paddle.is_tensor(img):            ax.imshow(img.numpy())        else:            ax.imshow(img)        ax.axes.get_xaxis().set_visible(False)        ax.axes.get_yaxis().set_visible(False)        if pred or gt:            ax.set_title("pt: " + pred[i] + "\ngt: " + gt[i])    return axes

登录后复制    In [27]

work_path = 'work/model'X, y = next(iter(DataLoader(val_dataset, batch_size=18)))model = fasternet_t0()model_state_dict = paddle.load(os.path.join(work_path, 'best_model.pdparams'))model.set_state_dict(model_state_dict)model.eval()logits = model(X)y_pred = paddle.argmax(logits, -1)X = paddle.transpose(X, [0, 2, 3, 1])axes = show_images(X.reshape((18, 224, 224, 3)), 1, 18, pred=get_cifar10_labels(y_pred), gt=get_cifar10_labels(y))plt.show()

登录后复制        

Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).

登录后复制        

登录后复制                In [ ]

!pip install interpretdl

登录后复制    In [29]

import interpretdl as it

登录后复制    In [30]

work_path = 'work/model'model = fasternet_t0()model_state_dict = paddle.load(os.path.join(work_path, 'best_model.pdparams'))model.set_state_dict(model_state_dict)

登录后复制    In [31]

X, y = next(iter(DataLoader(val_dataset, batch_size=18)))lime = it.LIMECVInterpreter(model)

登录后复制    In [32]

lime_weights = lime.interpret(X.numpy()[3], interpret_class=y.numpy()[3], batch_size=100, num_samples=10000, visual=True)

登录后复制        

100%|██████████| 10000/10000 [00:46<00:00, 212.97it/s]

登录后复制        

登录后复制                

3. 对比实验

对比实验见DWConv.ipynb

总结

        FasterNet从FLOPs和FLOPS两个角度重新审视卷积操作对于神经网络的影响，提出了新的神经网络家族——FasterNet。FasterNet不仅速度快，准确率也高。