前言
一、keras提供了三種定義模型的方式
1. 序列式(Sequential) API
序貫(sequential)API允許你為大多數問題逐層堆疊創建模型。雖然說對很多的應用來說，這樣的一個手法很簡單也解決了很多深度學習網絡結構的構建，但是它也有限制－它不允許你創建模型有共享層或有多個輸入或輸出的網絡。
2. 函數式(Functional) API
Keras函數式(functional)API為構建網絡模型提供了更為靈活的方式。
它允許你定義多個輸入或輸出模型以及共享圖層的模型。除此之外，它允許你定義動態(ad-hoc)的非週期性(acyclic)網絡圖。
模型是通過創建層的實例(layer instances)並將它們直接相互連接成對來定義的，然後定義一個模型(model)來指定那些層是要作為這個模型的輸入和輸出。
3.子類(Subclassing) API
補充知識：keras pytorch 構建模型對比
使用CIFAR10數據集，用三種框架構建Residual_Network作為例子，比較框架間的異同。
數據集格式
pytorch的數據集格式
import torch import torch.nn as nn import torchvision # Download and construct CIFAR-10 dataset. train_dataset = torchvision.datasets.CIFAR10(root='../../data/', train=True, download=True) # Fetch one data pair (read data from disk). image, label = train_dataset[0] print (image.size()) # torch.Size([3, 32, 32]) print (label) # 6 print (train_dataset.data.shape) # (50000, 32, 32, 3) # type(train_dataset.targets)==list print (len(train_dataset.targets)) # 50000 # Data loader (this provides queues and threads in a very simple way). train_loader = torch.utils.data.DataLoader(dataset=train_dataset, batch_size=64, shuffle=True) """ # 演示DataLoader返回的數據結構 # When iteration starts, queue and thread start to load data from files. data_iter = iter(train_loader) # Mini-batch images and labels. images, labels = data_iter.next() print(images.shape) # torch.Size([100, 3, 32, 32]) print(labels.shape) # torch.Size([100]) 可見經過DataLoader後，labels由list變成了pytorch內置的tensor格式 """ # 一般使用的話是下面這種 # Actual usage of the data loader is as below. for images, labels in train_loader: # Training code should be written here. pass
keras的數據格式
import keras from keras.datasets import cifar10 (train_x, train_y) , (test_x, test_y) = cifar10.load_data() print(train_x.shape) # ndarray 類型： (50000, 32, 32, 3) print(train_y.shape) # (50000, 1)
輸入網絡的數據格式不同
""" 1: pytorch 都是內置torch.xxTensor輸入網絡，而keras的則是原生ndarray類型 2: 對於multi-class的其中一種loss，即cross-entropy loss 而言， pytorch的api為 CorssEntropyLoss, 但y_true不能用one-hoe編碼！這與keras，tensorflow 都不同。tensorflow相應的api為softmax_cross_entropy 他們的api都僅限於multi-class classification 3*: 其實上面提到的api都屬於categorical cross-entropy loss, 又叫 softmax loss，是函數內部先進行了 softmax 激活，再經過cross-entropy loss。 這個loss是cross-entropy loss的變種， cross-entropy loss又叫logistic loss 或 multinomial logistic loss。 實現這種loss的函數不包括激活函數，需要自定義。 pytorch對應的api為BCEloss(僅限於 binary classification), tensorflow 對應的api為 log_loss。 cross-entropy loss的第二個變種是 binary cross-entropy loss 又叫 sigmoid cross- entropy loss。 函數內部先進行了sigmoid激活，再經過cross-entropy loss。 pytorch對應的api為BCEWithLogitsLoss, tensorflow對應的api為sigmoid_cross_entropy """ # pytorch criterion = nn.CrossEntropyLoss() ... for epoch in range(num_epochs): for i, (images, labels) in enumerate(train_loader): images = images.to(device) labels = labels.to(device) # Forward pass outputs = model(images) # 對於multi-class cross-entropy loss # 輸入y_true不需要one-hot編碼 loss = criterion(outputs, labels) ... # keras # 對於multi-class cross-entropy loss # 輸入y_true需要one-hot編碼 train_y = keras.utils.to_categorical(train_y,10) ... model.fit_generator(datagen.flow(train_x, train_y, batch_size=128), validation_data=[test_x,test_y], epochs=epochs,steps_per_epoch=steps_per_epoch, verbose=1) ...
整體流程
keras 流程
model = myModel() model.compile(optimizer=Adam(0.001),loss="categorical_crossentropy",metrics=["accuracy"]) model.fit_generator(datagen.flow(train_x, train_y, batch_size=128), validation_data=[test_x,test_y], epochs=epochs,steps_per_epoch=steps_per_epoch, verbose=1, workers=4) #Evaluate the accuracy of the test dataset accuracy = model.evaluate(x=test_x,y=test_y,batch_size=128) # 保存整個網絡 model.save("cifar10model.h5") """ # https://blog.csdn.net/jiandanjinxin/article/details/77152530 # 使用 # keras.models.load_model("cifar10model.h5") # 只保存architecture # json_string = model.to_json() # open('my_model_architecture.json','w').write(json_string) # 使用 # from keras.models import model_from_json #model = model_from_json(open('my_model_architecture.json').read()) # 只保存weights # model.save_weights('my_model_weights.h5') #需要在代碼中初始化一個完全相同的模型 # model.load_weights('my_model_weights.h5') #需要加載權重到不同的網絡結構（有些層一樣）中，例如fine-tune或transfer-learning，可以通過層名字來加載模型 # model.load_weights('my_model_weights.h5', by_name=True) """
pytorch 流程
model = myModel() # Loss and optimizer criterion = nn.CrossEntropyLoss() for epoch in range(num_epochs): for i, (images, labels) in enumerate(train_loader): images = images.to(device) labels = labels.to(device) # Forward pass outputs = model(images) loss = criterion(outputs, labels) # Backward and optimize # 將上次迭代計算的梯度值清0 optimizer.zero_grad() # 反向傳播，計算梯度值 loss.backward() # 更新權值參數 optimizer.step() # model.eval()，讓model變成測試模式，對dropout和batch normalization的操作在訓練和測試的時候是不一樣的 # eval（）時，pytorch會自動把BN和DropOut固定住，不會取平均，而是用訓練好的值。 # 不然的話，一旦test的batch_size過小，很容易就會被BN層導致生成圖片顏色失真極大。 model.eval() with torch.no_grad(): correct = 0 total = 0 for images, labels in test_loader: images = images.to(device) labels = labels.to(device) outputs = model(images) _, predicted = torch.max(outputs.data, 1) total += labels.size(0) correct += (predicted == labels).sum().item() print('Accuracy of the model on the test images: {} %'.format(100 * correct / total)) # Save the model checkpoint # 這是隻保存了weights torch.save(model.state_dict(), 'resnet.ckpt') """ # 使用 # myModel.load_state_dict(torch.load('params.ckpt')) # 若想保存整個網絡（architecture + weights) # torch.save(resnet, 'model.ckpt') # 使用 #model = torch.load('model.ckpt') """
對比流程
#https://blog.csdn.net/dss_dssssd/article/details/83892824 """ 1: 準備數據(注意數據格式不同） 2: 定義網絡結構model 3: 定義損失函數 4: 定義優化算法 optimizer 5: 訓練-keras 5.1:編譯模型（傳入loss function和optimizer等） 5.2:訓練模型（fit or fit_generator，傳入數據) 5: 訓練-pytorch 迭代訓練： 5.1:準備好tensor形式的輸入數據和標籤(可選) 5.2:前向傳播計算網絡輸出output和計算損失函數loss 5.3:反向傳播更新參數 以下三句話一句也不能少： 5.3.1:將上次迭代計算的梯度值清0 optimizer.zero_grad() 5.3.2:反向傳播，計算梯度值 loss.backward() 5.3.3:更新權值參數 optimizer.step() 6: 在測試集上測試-keras model.evaluate 6: 在測試集上測試-pytorch 遍歷測試集，自定義metric 7: 保存網絡（可選） 具體實現參考上面代碼 """
構建網絡
對比網絡
1、對於keras，不需要input_channels，函數內部會自動獲得，而pytorch則需要顯示聲明input_channels
2、對於pytorch Conv2d需要指定padding，而keras的則是same和valid兩種選項（valid即padding=0）
3、keras的Flatten操作可以視作pytorch中的view
4、keras的dimension一般順序是（H, W, C) (tensorflow 為backend的話），而pytorch的順序則是（ C, H, W)
5、具體的變換可以參照下方，但由於沒有學過pytorch，keras也剛入門，不能保證正確，日後學的更深入了之後再來看看。
pytorch 構建Residual-network
import torch import torch.nn as nn import torchvision import torchvision.transforms as transforms # Device configuration device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') # Hyper-parameters num_epochs = 80 learning_rate = 0.001 # Image preprocessing modules transform = transforms.Compose([ transforms.Pad(4), transforms.RandomHorizontalFlip(), transforms.RandomCrop(32), transforms.ToTensor()]) # CIFAR-10 dataset # train_dataset.data.shape #Out[31]: (50000, 32, 32, 3) # train_dataset.targets list # len(list)=5000 train_dataset = torchvision.datasets.CIFAR10(root='./data/', train=True, transform=transform, download=True) test_dataset = torchvision.datasets.CIFAR10(root='../../data/', train=False, transform=transforms.ToTensor()) # Data loader train_loader = torch.utils.data.DataLoader(dataset=train_dataset, batch_size=100, shuffle=True) test_loader = torch.utils.data.DataLoader(dataset=test_dataset, batch_size=100, shuffle=False) # 3x3 convolution def conv3x3(in_channels, out_channels, stride=1): return nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=stride, padding=1, bias=False) # Residual block class ResidualBlock(nn.Module): def __init__(self, in_channels, out_channels, stride=1, downsample=None): super(ResidualBlock, self).__init__() self.conv1 = conv3x3(in_channels, out_channels, stride) self.bn1 = nn.BatchNorm2d(out_channels) self.relu = nn.ReLU(inplace=True) self.conv2 = conv3x3(out_channels, out_channels) self.bn2 = nn.BatchNorm2d(out_channels) self.downsample = downsample def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) if self.downsample: residual = self.downsample(x) out += residual out = self.relu(out) return out # ResNet class ResNet(nn.Module): def __init__(self, block, layers, num_classes=10): super(ResNet, self).__init__() self.in_channels = 16 self.conv = conv3x3(3, 16) self.bn = nn.BatchNorm2d(16) self.relu = nn.ReLU(inplace=True) self.layer1 = self.make_layer(block, 16, layers[0]) self.layer2 = self.make_layer(block, 32, layers[1], 2) self.layer3 = self.make_layer(block, 64, layers[2], 2) self.avg_pool = nn.AvgPool2d(8) self.fc = nn.Linear(64, num_classes) def make_layer(self, block, out_channels, blocks, stride=1): downsample = None if (stride != 1) or (self.in_channels != out_channels): downsample = nn.Sequential( conv3x3(self.in_channels, out_channels, stride=stride), nn.BatchNorm2d(out_channels)) layers = [] layers.append(block(self.in_channels, out_channels, stride, downsample)) self.in_channels = out_channels for i in range(1, blocks): layers.append(block(out_channels, out_channels)) # [*[1,2,3]] # Out[96]: [1, 2, 3] return nn.Sequential(*layers) def forward(self, x): out = self.conv(x) # out.shape:torch.Size([100, 16, 32, 32]) out = self.bn(out) out = self.relu(out) out = self.layer1(out) out = self.layer2(out) out = self.layer3(out) out = self.avg_pool(out) out = out.view(out.size(0), -1) out = self.fc(out) return out model = ResNet(ResidualBlock, [2, 2, 2]).to(device) # pip install torchsummary or # git clone https://github.com/sksq96/pytorch-summary from torchsummary import summary # input_size=(C,H,W) summary(model, input_size=(3, 32, 32)) images,labels = iter(train_loader).next() outputs = model(images) # Loss and optimizer criterion = nn.CrossEntropyLoss() optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate) # For updating learning rate def update_lr(optimizer, lr): for param_group in optimizer.param_groups: param_group['lr'] = lr # Train the model total_step = len(train_loader) curr_lr = learning_rate for epoch in range(num_epochs): for i, (images, labels) in enumerate(train_loader): images = images.to(device) labels = labels.to(device) # Forward pass outputs = model(images) loss = criterion(outputs, labels) # Backward and optimize optimizer.zero_grad() loss.backward() optimizer.step() if (i+1) % 100 == 0: print ("Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}" .format(epoch+1, num_epochs, i+1, total_step, loss.item())) # Decay learning rate if (epoch+1) % 20 == 0: curr_lr /= 3 update_lr(optimizer, curr_lr) # Test the model model.eval() with torch.no_grad(): correct = 0 total = 0 for images, labels in test_loader: images = images.to(device) labels = labels.to(device) outputs = model(images) _, predicted = torch.max(outputs.data, 1) total += labels.size(0) correct += (predicted == labels).sum().item() print('Accuracy of the model on the test images: {} %'.format(100 * correct / total)) # Save the model checkpoint torch.save(model.state_dict(), 'resnet.ckpt')
keras 對應的網絡構建部分
""" #pytorch def conv3x3(in_channels, out_channels, stride=1): return nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=stride, padding=1, bias=False) """ def conv3x3(x,out_channels, stride=1): #out = spatial_2d_padding(x,padding=((1, 1), (1, 1)), data_format="channels_last") return Conv2D(filters=out_channels, kernel_size=[3,3], strides=(stride,stride),padding="same")(x) """ # pytorch # Residual block class ResidualBlock(nn.Module): def __init__(self, in_channels, out_channels, stride=1, downsample=None): super(ResidualBlock, self).__init__() self.conv1 = conv3x3(in_channels, out_channels, stride) self.bn1 = nn.BatchNorm2d(out_channels) self.relu = nn.ReLU(inplace=True) self.conv2 = conv3x3(out_channels, out_channels) self.bn2 = nn.BatchNorm2d(out_channels) self.downsample = downsample def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) if self.downsample: residual = self.downsample(x) out += residual out = self.relu(out) return out """ def ResidualBlock(x, out_channels, stride=1, downsample=False): residual = x out = conv3x3(x, out_channels,stride) out = BatchNormalization()(out) out = Activation("relu")(out) out = conv3x3(out, out_channels) out = BatchNormalization()(out) if downsample: residual = conv3x3(residual, out_channels, stride=stride) residual = BatchNormalization()(residual) out = keras.layers.add([residual,out]) out = Activation("relu")(out) return out """ #pytorch def make_layer(self, block, out_channels, blocks, stride=1): downsample = None if (stride != 1) or (self.in_channels != out_channels): downsample = nn.Sequential( conv3x3(self.in_channels, out_channels, stride=stride), nn.BatchNorm2d(out_channels)) layers = [] layers.append(block(self.in_channels, out_channels, stride, downsample)) self.in_channels = out_channels for i in range(1, blocks): layers.append(block(out_channels, out_channels)) # [*[1,2,3]] # Out[96]: [1, 2, 3] return nn.Sequential(*layers) """ def make_layer(x, out_channels, blocks, stride=1): # tf backend: x.output_shape[-1]==out_channels #print("x.shape[-1] ",x.shape[-1]) downsample = False if (stride != 1) or (out_channels != x.shape[-1]): downsample = True out = ResidualBlock(x, out_channels, stride, downsample) for i in range(1, blocks): out = ResidualBlock(out, out_channels) return out def KerasResidual(input_shape): images = Input(input_shape) out = conv3x3(images,16) # out.shape=(None, 32, 32, 16) out = BatchNormalization()(out) out = Activation("relu")(out) layer1_out = make_layer(out, 16, layers[0]) layer2_out = make_layer(layer1_out, 32, layers[1], 2) layer3_out = make_layer(layer2_out, 64, layers[2], 2) out = AveragePooling2D(pool_size=(8,8))(layer3_out) out = Flatten()(out) # pytorch 的nn.CrossEntropyLoss()會首先執行softmax計算 # 當換成keras時，沒有tf類似的softmax_cross_entropy # 自帶的categorical_crossentropy不會執行激活操作，因此得在Dense層加上activation out = Dense(units=10, activation="softmax")(out) model = Model(inputs=images,outputs=out) return model input_shape=(32, 32, 3) layers=[2, 2, 2] mymodel = KerasResidual(input_shape) mymodel.summary()
pytorch model summary
---------------------------------------------------------------- Layer (type) Output Shape Param # ================================================================ Conv2d-1 [-1, 16, 32, 32] 432 BatchNorm2d-2 [-1, 16, 32, 32] 32 ReLU-3 [-1, 16, 32, 32] 0 Conv2d-4 [-1, 16, 32, 32] 2,304 BatchNorm2d-5 [-1, 16, 32, 32] 32 ReLU-6 [-1, 16, 32, 32] 0 Conv2d-7 [-1, 16, 32, 32] 2,304 BatchNorm2d-8 [-1, 16, 32, 32] 32 ReLU-9 [-1, 16, 32, 32] 0 ResidualBlock-10 [-1, 16, 32, 32] 0 Conv2d-11 [-1, 16, 32, 32] 2,304 BatchNorm2d-12 [-1, 16, 32, 32] 32 ReLU-13 [-1, 16, 32, 32] 0 Conv2d-14 [-1, 16, 32, 32] 2,304 BatchNorm2d-15 [-1, 16, 32, 32] 32 ReLU-16 [-1, 16, 32, 32] 0 ResidualBlock-17 [-1, 16, 32, 32] 0 Conv2d-18 [-1, 32, 16, 16] 4,608 BatchNorm2d-19 [-1, 32, 16, 16] 64 ReLU-20 [-1, 32, 16, 16] 0 Conv2d-21 [-1, 32, 16, 16] 9,216 BatchNorm2d-22 [-1, 32, 16, 16] 64 Conv2d-23 [-1, 32, 16, 16] 4,608 BatchNorm2d-24 [-1, 32, 16, 16] 64 ReLU-25 [-1, 32, 16, 16] 0 ResidualBlock-26 [-1, 32, 16, 16] 0 Conv2d-27 [-1, 32, 16, 16] 9,216 BatchNorm2d-28 [-1, 32, 16, 16] 64 ReLU-29 [-1, 32, 16, 16] 0 Conv2d-30 [-1, 32, 16, 16] 9,216 BatchNorm2d-31 [-1, 32, 16, 16] 64 ReLU-32 [-1, 32, 16, 16] 0 ResidualBlock-33 [-1, 32, 16, 16] 0 Conv2d-34 [-1, 64, 8, 8] 18,432 BatchNorm2d-35 [-1, 64, 8, 8] 128 ReLU-36 [-1, 64, 8, 8] 0 Conv2d-37 [-1, 64, 8, 8] 36,864 BatchNorm2d-38 [-1, 64, 8, 8] 128 Conv2d-39 [-1, 64, 8, 8] 18,432 BatchNorm2d-40 [-1, 64, 8, 8] 128 ReLU-41 [-1, 64, 8, 8] 0 ResidualBlock-42 [-1, 64, 8, 8] 0 Conv2d-43 [-1, 64, 8, 8] 36,864 BatchNorm2d-44 [-1, 64, 8, 8] 128 ReLU-45 [-1, 64, 8, 8] 0 Conv2d-46 [-1, 64, 8, 8] 36,864 BatchNorm2d-47 [-1, 64, 8, 8] 128 ReLU-48 [-1, 64, 8, 8] 0 ResidualBlock-49 [-1, 64, 8, 8] 0 AvgPool2d-50 [-1, 64, 1, 1] 0 Linear-51 [-1, 10] 650 ================================================================ Total params: 195,738 Trainable params: 195,738 Non-trainable params: 0 ---------------------------------------------------------------- Input size (MB): 0.01 Forward/backward pass size (MB): 3.63 Params size (MB): 0.75 Estimated Total Size (MB): 4.38 ----------------------------------------------------------------
keras model summary
__________________________________________________________________________________________________ Layer (type) Output Shape Param # Connected to ================================================================================================== input_26 (InputLayer) (None, 32, 32, 3) 0 __________________________________________________________________________________________________ conv2d_103 (Conv2D) (None, 32, 32, 16) 448 input_26[0][0] __________________________________________________________________________________________________ batch_normalization_99 (BatchNo (None, 32, 32, 16) 64 conv2d_103[0][0] __________________________________________________________________________________________________ activation_87 (Activation) (None, 32, 32, 16) 0 batch_normalization_99[0][0] __________________________________________________________________________________________________ conv2d_104 (Conv2D) (None, 32, 32, 16) 2320 activation_87[0][0] __________________________________________________________________________________________________ batch_normalization_100 (BatchN (None, 32, 32, 16) 64 conv2d_104[0][0] __________________________________________________________________________________________________ activation_88 (Activation) (None, 32, 32, 16) 0 batch_normalization_100[0][0] __________________________________________________________________________________________________ conv2d_105 (Conv2D) (None, 32, 32, 16) 2320 activation_88[0][0] __________________________________________________________________________________________________ batch_normalization_101 (BatchN (None, 32, 32, 16) 64 conv2d_105[0][0] __________________________________________________________________________________________________ add_34 (Add) (None, 32, 32, 16) 0 activation_87[0][0] batch_normalization_101[0][0] __________________________________________________________________________________________________ activation_89 (Activation) (None, 32, 32, 16) 0 add_34[0][0] __________________________________________________________________________________________________ conv2d_106 (Conv2D) (None, 32, 32, 16) 2320 activation_89[0][0] __________________________________________________________________________________________________ batch_normalization_102 (BatchN (None, 32, 32, 16) 64 conv2d_106[0][0] __________________________________________________________________________________________________ activation_90 (Activation) (None, 32, 32, 16) 0 batch_normalization_102[0][0] __________________________________________________________________________________________________ conv2d_107 (Conv2D) (None, 32, 32, 16) 2320 activation_90[0][0] __________________________________________________________________________________________________ batch_normalization_103 (BatchN (None, 32, 32, 16) 64 conv2d_107[0][0] __________________________________________________________________________________________________ add_35 (Add) (None, 32, 32, 16) 0 activation_89[0][0] batch_normalization_103[0][0] __________________________________________________________________________________________________ activation_91 (Activation) (None, 32, 32, 16) 0 add_35[0][0] __________________________________________________________________________________________________ conv2d_108 (Conv2D) (None, 16, 16, 32) 4640 activation_91[0][0] __________________________________________________________________________________________________ batch_normalization_104 (BatchN (None, 16, 16, 32) 128 conv2d_108[0][0] __________________________________________________________________________________________________ activation_92 (Activation) (None, 16, 16, 32) 0 batch_normalization_104[0][0] __________________________________________________________________________________________________ conv2d_110 (Conv2D) (None, 16, 16, 32) 4640 activation_91[0][0] __________________________________________________________________________________________________ conv2d_109 (Conv2D) (None, 16, 16, 32) 9248 activation_92[0][0] __________________________________________________________________________________________________ batch_normalization_106 (BatchN (None, 16, 16, 32) 128 conv2d_110[0][0] __________________________________________________________________________________________________ batch_normalization_105 (BatchN (None, 16, 16, 32) 128 conv2d_109[0][0] __________________________________________________________________________________________________ add_36 (Add) (None, 16, 16, 32) 0 batch_normalization_106[0][0] batch_normalization_105[0][0] __________________________________________________________________________________________________ activation_93 (Activation) (None, 16, 16, 32) 0 add_36[0][0] __________________________________________________________________________________________________ conv2d_111 (Conv2D) (None, 16, 16, 32) 9248 activation_93[0][0] __________________________________________________________________________________________________ batch_normalization_107 (BatchN (None, 16, 16, 32) 128 conv2d_111[0][0] __________________________________________________________________________________________________ activation_94 (Activation) (None, 16, 16, 32) 0 batch_normalization_107[0][0] __________________________________________________________________________________________________ conv2d_112 (Conv2D) (None, 16, 16, 32) 9248 activation_94[0][0] __________________________________________________________________________________________________ batch_normalization_108 (BatchN (None, 16, 16, 32) 128 conv2d_112[0][0] __________________________________________________________________________________________________ add_37 (Add) (None, 16, 16, 32) 0 activation_93[0][0] batch_normalization_108[0][0] __________________________________________________________________________________________________ activation_95 (Activation) (None, 16, 16, 32) 0 add_37[0][0] __________________________________________________________________________________________________ conv2d_113 (Conv2D) (None, 8, 8, 64) 18496 activation_95[0][0] __________________________________________________________________________________________________ batch_normalization_109 (BatchN (None, 8, 8, 64) 256 conv2d_113[0][0] __________________________________________________________________________________________________ activation_96 (Activation) (None, 8, 8, 64) 0 batch_normalization_109[0][0] __________________________________________________________________________________________________ conv2d_115 (Conv2D) (None, 8, 8, 64) 18496 activation_95[0][0] __________________________________________________________________________________________________ conv2d_114 (Conv2D) (None, 8, 8, 64) 36928 activation_96[0][0] __________________________________________________________________________________________________ batch_normalization_111 (BatchN (None, 8, 8, 64) 256 conv2d_115[0][0] __________________________________________________________________________________________________ batch_normalization_110 (BatchN (None, 8, 8, 64) 256 conv2d_114[0][0] __________________________________________________________________________________________________ add_38 (Add) (None, 8, 8, 64) 0 batch_normalization_111[0][0] batch_normalization_110[0][0] __________________________________________________________________________________________________ activation_97 (Activation) (None, 8, 8, 64) 0 add_38[0][0] __________________________________________________________________________________________________ conv2d_116 (Conv2D) (None, 8, 8, 64) 36928 activation_97[0][0] __________________________________________________________________________________________________ batch_normalization_112 (BatchN (None, 8, 8, 64) 256 conv2d_116[0][0] __________________________________________________________________________________________________ activation_98 (Activation) (None, 8, 8, 64) 0 batch_normalization_112[0][0] __________________________________________________________________________________________________ conv2d_117 (Conv2D) (None, 8, 8, 64) 36928 activation_98[0][0] __________________________________________________________________________________________________ batch_normalization_113 (BatchN (None, 8, 8, 64) 256 conv2d_117[0][0] __________________________________________________________________________________________________ add_39 (Add) (None, 8, 8, 64) 0 activation_97[0][0] batch_normalization_113[0][0] __________________________________________________________________________________________________ activation_99 (Activation) (None, 8, 8, 64) 0 add_39[0][0] __________________________________________________________________________________________________ average_pooling2d_2 (AveragePoo (None, 1, 1, 64) 0 activation_99[0][0] __________________________________________________________________________________________________ flatten_2 (Flatten) (None, 64) 0 average_pooling2d_2[0][0] __________________________________________________________________________________________________ dense_2 (Dense) (None, 10) 650 flatten_2[0][0] ================================================================================================== Total params: 197,418 Trainable params: 196,298 Non-trainable params: 1,120 __________________________________________________________________________________________________
[lousu-xi ] keras的三種模型實現與區別說明已經有259次圍觀