【CVPR 2020】Dynamic Convolution：在卷积核上的注意力

轻量级卷积神经网络因计算预算限制深度和宽度，导致表示能力与性能不足。为此提出动态卷积，不增加网络深度或宽度，每层用多个并行卷积核，依输入注意力动态聚合。这既因核小高效，又因非线性聚合增强表示能力。将其用于MobileNetv3 - Small，ImageNet分类TOP - 1精度提2.9%，仅增4%Flops，COCO关键点检测提2.9AP。

Dynamic Convolution：在卷积核上的注意力

摘要

        轻量级卷积神经网络（CNNs）由于其较低的计算预算限制了CNNs的深度（卷积层数）和宽度（通道数），导致其表示能力有限，从而导致性能下降。 为了解决这个问题，我们提出了动态卷积，一种在不增加网络深度或宽度的情况下增加模型复杂性的新设计。 动态卷积不是每层使用一个卷积核，而是根据依赖于输入的注意力动态聚合多个并行的卷积核。 集合多个核不仅由于卷积核小而计算效率高，而且由于这些核通过注意力以非线性方式聚合而具有更强的表示能力。 通过对最先进的体系结构MobileNetv3-Small简单地使用动态卷积，ImageNet分类的TOP-1精度提高了2.9%，仅增加了4%的Flops，COCO关键点检测的增益达到了2.9AP。

免费影视、动漫、音乐、游戏、小说资源长期稳定更新！ 👉 点此立即查看 👈

1. Dynamic Convolution

        常规卷积对所有实例使用同样的卷积核，这会损害模型对实例的表示能力。为此，如图3所示，本文提出了Dynamic Convolution。与CondConv思想一样：首先创建一个可学习的卷积核库，然后使用路由函数预测每一卷积核的权重，从而得到针对该实例的专门卷积核。具体实现有两点不同：

CondConv仅使用一个简单的全连接层和Sigmoid函数生成权重（这会削弱表达能力），因此本文采用类似SE Layer的操作，激活函数使用Softmax函数（如图4所示，可以约束解空间）。在早期Dynamic Convolution使用几乎均匀的注意力以保证在早期，卷积核库中的卷积核可以有效地更新。这个通过设置Softmax函数的温度参数来实现，早期阶段使用较大的温度，然后进行线性衰减到1。

 

2. 代码复现

2.1 下载并导入所需要的包

In [1]

%matplotlib inlineimport paddleimport 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 figurefrom paddle import ParamAttrfrom paddle.nn.layer.norm import _BatchNormBaseimport math

登录后复制

2.2 创建数据集

In [2]

train_tfm = transforms.Compose([    transforms.Resize((130, 130)),    transforms.RandomResizedCrop(128),    transforms.RandomHorizontalFlip(0.5),    transforms.ToTensor(),    transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),])test_tfm = transforms.Compose([    transforms.Resize((128, 128)),    transforms.ToTensor(),    transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),])

登录后复制In [3]

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 [4]

batch_size=512

登录后复制In [5]

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 标签平滑

In [6]

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

2.4 AlexNet-DY

2.4.1 Dynamic Convolution

In [7]

class RoutingAttention(nn.Layer):    def __init__(self, inplanes, num_experts, ratio=4, temperature=30, end_epoches=10):        super().__init__()        self.avgpool = nn.AdaptiveAvgPool2D(1)        self.net = nn.Sequential(            nn.Conv2D(inplanes, int(inplanes//ratio), 1),            nn.ReLU(),            nn.Conv2D(int(inplanes//ratio), num_experts, 1)        )        self.temperature = temperature        self.step = self.temperature // end_epoches    def update_temperature(self):        if self.temperature > 1:            self.temperature -=self.step            if self.temperature < 1:                self.temperature = 1        return self.temperature    def set_temperature(self, temperature=1):        self.temperature = temperature        return self.temperature    def forward(self, x):        attn=self.avgpool(x)        attn=self.net(attn).reshape((attn.shape[0], -1))        return F.softmax(attn / self.temperature)

登录后复制In [8]

class DYConv2D(nn.Layer):    def __init__(self, inplanes, outplanes, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias_attr=True, num_experts=4):        super().__init__()        self.inplanes=inplanes        self.outplanes=outplanes        self.kernel_size=kernel_size        self.stride=stride        self.padding=padding        self.dilation=dilation        self.groups=groups        self.bias=bias_attr        self.num_experts=num_experts        self.routing=RoutingAttention(inplanes=inplanes, num_experts=num_experts)        self.weight=self.create_parameter((num_experts, outplanes, inplanes // groups, kernel_size, kernel_size),            default_initializer=nn.initializer.KaimingNormal()) # num_experts, out, in//g, k, k        if(bias_attr):            self.bias=self.create_parameter((num_experts, outplanes), default_initializer=nn.initializer.KaimingNormal())        else:            self.bias=None    def forward(self, x):        b, c, h, w = x.shape        attn = self.routing(x) # b, num_experts        x = x.reshape((1, -1, h, w))    #由于DY CNN对每一个样本都有不同的权重，因此为了使用F.conv2d，将batch维放入特征C中        weight = paddle.mm(attn, self.weight.reshape((self.num_experts, -1))).reshape(            (-1, self.inplanes//self.groups, self.kernel_size, self.kernel_size))  # b*out, in//g, k, k        if(self.bias is not None):            bias=paddle.mm(attn, self.bias.reshape((self.num_experts, -1))).reshape([-1])            output=F.conv2d(x, weight=weight, bias=bias, stride=self.stride, padding=self.padding, dilation=self.dilation, groups=self.groups * b)        else:            bias=None            output=F.conv2d(x, weight=weight, bias=bias, stride=self.stride, padding=self.padding, dilation=self.dilation, groups=self.groups * b)        output=output.reshape((b, self.outplanes, output.shape[-2], output.shape[-1]))        return output

登录后复制In [10]

model = DYConv2D(64, 128, 3, padding=1, stride=2, num_experts=4)paddle.summary(model, (4, 64, 224, 224))

登录后复制

W0131 21:58:42.897727   396 gpu_resources.cc:61] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.2, Runtime API Version: 11.2W0131 21:58:42.901930   396 gpu_resources.cc:91] device: 0, cuDNN Version: 8.2.

登录后复制

-------------------------------------------------------------------------------   Layer (type)         Input Shape          Output Shape         Param #    ===============================================================================AdaptiveAvgPool2D-1 [[4, 64, 224, 224]]     [4, 64, 1, 1]            0            Conv2D-1         [[4, 64, 1, 1]]       [4, 16, 1, 1]          1,040           ReLU-5          [[4, 16, 1, 1]]       [4, 16, 1, 1]            0            Conv2D-2         [[4, 16, 1, 1]]        [4, 4, 1, 1]           68       RoutingAttention-1  [[4, 64, 224, 224]]         [4, 4]               0       ===============================================================================Total params: 1,108Trainable params: 1,108Non-trainable params: 0-------------------------------------------------------------------------------Input size (MB): 49.00Forward/backward pass size (MB): 0.00Params size (MB): 0.00Estimated Total Size (MB): 49.01-------------------------------------------------------------------------------

登录后复制

{'total_params': 1108, 'trainable_params': 1108}

登录后复制

2.4.2 AlexNet-DY

In [9]

class AlexNet_DY(nn.Layer):    def __init__(self,num_classes=10):        super().__init__()        self.features=nn.Sequential(            nn.Conv2D(3, 48, kernel_size=11, stride=4, padding=11//2),            nn.BatchNorm(48),            nn.ReLU(),            nn.MaxPool2D(kernel_size=3, stride=2),            nn.Conv2D(48, 128, kernel_size=5, padding=2),            nn.BatchNorm(128),            nn.ReLU(),            nn.MaxPool2D(kernel_size=3, stride=2),            DYConv2D(128, 192, kernel_size=3, stride=1, padding=1, num_experts=2),            nn.BatchNorm(192),            nn.ReLU(),            DYConv2D(192, 192, kernel_size=3, stride=1, padding=1, num_experts=2),            nn.BatchNorm(192),            nn.ReLU(),            DYConv2D(192, 128, kernel_size=3, stride=1, padding=1, num_experts=2),            nn.BatchNorm(128),            nn.ReLU(),            nn.MaxPool2D(kernel_size=3, stride=2),        )        self.classifier=nn.Sequential(            nn.Linear(3 * 3 * 128, 2048),            nn.ReLU(),            nn.Dropout(0.5),            nn.Linear(2048, 2048),            nn.ReLU(),            nn.Dropout(0.5),            nn.Linear(2048, num_classes),        )    def forward(self,x):        x = self.features(x)        x = paddle.flatten(x, 1)        x=self.classifier(x)        return x

登录后复制In [ ]

model = AlexNet_DY(num_classes=10)paddle.summary(model, (4, 3, 128, 128))

登录后复制

2.5 训练

In [10]

learning_rate = 0.1n_epochs = 100paddle.seed(42)np.random.seed(42)

登录后复制登录后复制In [11]

def init_weight(m):        zeros = nn.initializer.Constant(0)        ones = nn.initializer.Constant(1)        if isinstance(m, (nn.Conv2D, nn.Linear)):            nn.initializer.KaimingNormal(m.weight)        if isinstance(m, nn.BatchNorm2D):            zeros(m.bias)            ones(m.weight)

登录后复制In [ ]

work_path = 'work/model'model = AlexNet_DY(num_classes=10)model.apply(init_weight)criterion = LabelSmoothingCrossEntropy()scheduler = paddle.optimizer.lr.MultiStepDecay(learning_rate=learning_rate, milestones=[30, 60, 90], verbose=False)optimizer = paddle.optimizer.SGD(parameters=model.parameters(), learning_rate=scheduler, weight_decay=1e-5)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 = accuracy_manager.compute(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()        optimizer.clear_grad()        train_loss += loss        train_num += len(y_data)    scheduler.step()    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 = val_accuracy_manager.compute(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'))    for i in model.features.children():        if isinstance(i, DYConv2D):            i.routing.update_temperature()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

2.6 实验结果

In [16]

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 [17]

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登录后复制登录后复制In [18]

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

登录后复制

登录后复制登录后复制登录后复制登录后复制In [19]

import timework_path = 'work/model'model = AlexNet_DY(num_classes=10)for i in model.features.children():        if isinstance(i, CondConv2D):            i.routing.set_temperature()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:1764

登录后复制In [20]

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 [21]

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)        ax.set_title("pt: " + str(pred[i]) + "\ngt: " + str(gt[i]))    return axes

登录后复制In [22]

work_path = 'work/model'X, y = next(iter(DataLoader(val_dataset, batch_size=18)))model = AlexNet_DY(num_classes=10)for i in model.features.children():        if isinstance(i, CondConv2D):            i.routing.set_temperature()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, 128, 128, 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

登录后复制登录后复制

3. AlexNet

3.1 AlexNet

In [23]

class AlexNet(nn.Layer):    def __init__(self,num_classes=10):        super().__init__()        self.features=nn.Sequential(            nn.Conv2D(3,48, kernel_size=11, stride=4, padding=11//2),            nn.BatchNorm2D(48),            nn.ReLU(),            nn.MaxPool2D(kernel_size=3,stride=2),            nn.Conv2D(48, 128, kernel_size=5, padding=2),            nn.BatchNorm2D(128),            nn.ReLU(),            nn.MaxPool2D(kernel_size=3,stride=2),            nn.Conv2D(128, 192, kernel_size=3, stride=1, padding=1),            nn.BatchNorm2D(192),            nn.ReLU(),            nn.Conv2D(192, 192, kernel_size=3, stride=1, padding=1),            nn.BatchNorm2D(192),            nn.ReLU(),            nn.Conv2D(192, 128,kernel_size=3, stride=1, padding=1),            nn.BatchNorm2D(128),            nn.ReLU(),            nn.MaxPool2D(kernel_size=3, stride=2),        )        self.classifier=nn.Sequential(            nn.Linear(3 * 3 * 128, 2048),            nn.ReLU(),            nn.Dropout(0.5),            nn.Linear(2048, 2048),            nn.ReLU(),            nn.Dropout(0.5),            nn.Linear(2048, num_classes),        )    def forward(self,x):        x = self.features(x)        x = paddle.flatten(x, 1)        x=self.classifier(x)        return x

登录后复制In [ ]

model = AlexNet(num_classes=10)paddle.summary(model, (1, 3, 128, 128))

登录后复制

3.2 训练

In [25]

learning_rate = 0.1n_epochs = 100paddle.seed(42)np.random.seed(42)

登录后复制登录后复制In [ ]

work_path = 'work/model1'model = AlexNet(num_classes=10)model.apply(init_weight)criterion = LabelSmoothingCrossEntropy()scheduler = paddle.optimizer.lr.MultiStepDecay(learning_rate=learning_rate, milestones=[30, 60, 90], verbose=False)optimizer = paddle.optimizer.SGD(parameters=model.parameters(), learning_rate=scheduler, weight_decay=1e-5)gate = 0.0threshold = 0.0best_acc = 0.0val_acc = 0.0loss_record1 = {'train': {'loss': [], 'iter': []}, 'val': {'loss': [], 'iter': []}}   # for recording lossacc_record1 = {'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 = accuracy_manager.compute(logits, labels)        accuracy_manager.update(acc)        if batch_id % 10 == 0:            loss_record1['train']['loss'].append(loss.numpy())            loss_record1['train']['iter'].append(loss_iter)            loss_iter += 1        loss.backward()        optimizer.step()        optimizer.clear_grad()        train_loss += loss        train_num += len(y_data)    scheduler.step()    total_train_loss = (train_loss / train_num) * batch_size    train_acc = accuracy_manager.accumulate()    acc_record1['train']['acc'].append(train_acc)    acc_record1['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 = val_accuracy_manager.compute(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_record1['val']['loss'].append(total_val_loss.numpy())    loss_record1['val']['iter'].append(loss_iter)    val_acc = val_accuracy_manager.accumulate()    acc_record1['val']['acc'].append(val_acc)    acc_record1['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

3.3 实验结果

In [27]

plot_learning_curve(loss_record1, title="loss", ylabel='CE Loss')

登录后复制

登录后复制登录后复制登录后复制登录后复制In [28]

plot_learning_curve(acc_record1, title="acc", ylabel='Accuracy')

登录后复制

登录后复制登录后复制登录后复制登录后复制In [29]

##### import timework_path = 'work/model1'model = AlexNet(num_classes=10)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:1822

登录后复制In [30]

work_path = 'work/model1'X, y = next(iter(DataLoader(val_dataset, batch_size=18)))model = AlexNet(num_classes=10)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, 128, 128, 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

登录后复制登录后复制

4. 对比实验结果