复杂神经网络模型的实现离不开"融合"操作。常见融合操作如下:
(1)求和,求差
# 求和 layers.Add(inputs) # 求差 layers.Subtract(inputs)inputs: 一个输入张量的列表(列表大小至少为 2),列表的shape必须一样才能进行求和(求差)操作。
例子:
input1 = keras.layers.Input(shape=(16,)) x1 = keras.layers.Dense(8, activation='relu')(input1) input2 = keras.layers.Input(shape=(32,)) x2 = keras.layers.Dense(8, activation='relu')(input2) added = keras.layers.add([x1, x2]) out = keras.layers.Dense(4)(added) model = keras.models.Model(inputs=[input1, input2], outputs=out)(2)乘法
# 输入张量的逐元素乘积(对应位置元素相乘,输入维度必须相同) layers.multiply(inputs) # 输入张量样本之间的点积 layers.dot(inputs, axes, normalize=False)dot即矩阵乘法,例子1:
x = np.arange(10).reshape(1, 5, 2) y = np.arange(10, 20).reshape(1, 2, 5) # 三维的输入做dot通常像这样指定axes,表示矩阵的第一维度和第二维度参与矩阵乘法,第0维度是batchsize tf.keras.layers.Dot(axes=(1, 2))([x, y]) # 输出如下: <tf.Tensor: shape=(1, 2, 2), dtype=int64, numpy= array([[[260, 360], [320, 445]]])>例子2:
x1 = tf.keras.layers.Dense(8)(np.arange(10).reshape(5, 2)) x2 = tf.keras.layers.Dense(8)(np.arange(10, 20).reshape(5, 2)) dotted = tf.keras.layers.Dot(axes=1)([x1, x2]) dotted.shape TensorShape([5, 1])(3)联合:
# 所有输入张量通过 axis 轴串联起来的输出张量。 layers.add(inputs,axis=-1) inputs: 一个列表的输入张量(列表大小至少为 2)。axis: 串联的轴。例子:
x1 = tf.keras.layers.Dense(8)(np.arange(10).reshape(5, 2)) x2 = tf.keras.layers.Dense(8)(np.arange(10, 20).reshape(5, 2)) concatted = tf.keras.layers.Concatenate()([x1, x2]) concatted.shape TensorShape([5, 16])(4)统计操作
求均值layers.Average()
input1 = tf.keras.layers.Input(shape=(16,)) x1 = tf.keras.layers.Dense(8, activation='relu')(input1) input2 = tf.keras.layers.Input(shape=(32,)) x2 = tf.keras.layers.Dense(8, activation='relu')(input2) avg = tf.keras.layers.Average()([x1, x2]) # x_1 x_2 的均值作为输出 print(avg) # <tf.Tensor 'average/Identity:0' shape=(None, 8) dtype=float32> out = tf.keras.layers.Dense(4)(avg) model = tf.keras.models.Model(inputs=[input1, input2], outputs=out)layers.Maximum()用法相同。
假设要构造这样一个模型:
(1)模型具有以下三个输入
工单标题(文本输入),工单的文本正文(文本输入),以及用户添加的任何标签(分类输入)(2)模型将具有两个输出:
介于 0 和 1 之间的优先级分数(标量 Sigmoid 输出)应该处理工单的部门(部门范围内的 Softmax 输出)。模型大概长这样:
接下来开始创建这个模型。
(1)模型的输入
num_tags = 12 num_words = 10000 num_departments = 4 title_input = keras.Input(shape=(None,), name="title") # Variable-length sequence of ints body_input = keras.Input(shape=(None,), name="body") # Variable-length sequence of ints tags_input = keras.Input(shape=(num_tags,), name="tags") # Binary vectors of size `num_tags`(2)将输入的每一个词进行嵌入成64-dimensional vector
title_features = layers.Embedding(num_words,64)(title_input) body_features = layers.Embedding(num_words,64)(body_input)(3)处理结果输入LSTM模型,得到 128-dimensional vector
title_features = layers.LSTM(128)(title_features) body_features = layers.LSTM(32)(body_features)(4)concatenate融合所有的特征
x = layers.concatenate([title_features, body_features, tags_input])(5)模型的输出
# 输出1,回归问题 priority_pred = layers.Dense(1,name="priority")(x) # 输出2,分类问题 department_pred = layers.Dense(num_departments,name="department")(x)(6)定义模型
model = keras.Model( inputs=[title_input, body_input, tags_input], outputs=[priority_pred, department_pred], )(7)模型编译
编译此模型时,可以为每个输出分配不同的损失。甚至可以为每个损失分配不同的权重,以调整其对总训练损失的贡献。
model.compile( optimizer=keras.optimizers.RMSprop(1e-3), loss={ "priority": keras.losses.BinaryCrossentropy(from_logits=True), "department": keras.losses.CategoricalCrossentropy(from_logits=True), }, loss_weights=[1.0, 0.2], )(8)模型的训练
# Dummy input data title_data = np.random.randint(num_words, size=(1280, 10)) body_data = np.random.randint(num_words, size=(1280, 100)) tags_data = np.random.randint(2, size=(1280, num_tags)).astype("float32") # Dummy target data priority_targets = np.random.random(size=(1280, 1)) dept_targets = np.random.randint(2, size=(1280, num_departments)) # 通过字典的形式将数据fit到模型 model.fit( {"title": title_data, "body": body_data, "tags": tags_data}, {"priority": priority_targets, "department": dept_targets}, epochs=2, batch_size=32, )通过add来实现融合操作,模型的基本结构如下:
# 实现第一个块 _input = keras.Input(shape=(32,32,3)) x = layers.Conv2D(32,3,activation='relu')(_input) x = layers.Conv2D(64,3,activation='relu')(x) block1_output = layers.MaxPooling2D(3)(x) # 实现第二个块 x = layers.Conv2D(64,3,padding='same',activation='relu')(block1_output) x = layers.Conv2D(64,3,padding='same',activation='relu')(x) block2_output = layers.add([x,block1_output]) # 实现第三个块 x = layers.Conv2D(64, 3, activation="relu", padding="same")(block2_output) x = layers.Conv2D(64, 3, activation="relu", padding="same")(x) block_3_output = layers.add([x, block2_output]) # 进入全连接层 x = layers.Conv2D(64,3,activation='relu')(block_3_output) x = layers.GlobalAveragePooling2D()(x) x = layers.Dense(256, activation="relu")(x) x = layers.Dropout(0.5)(x) outputs = layers.Dense(10)(x)模型的定义与编译:
model = keras.Model(_input,outputs,name='resnet') model.compile( optimizer=keras.optimizers.RMSprop(1e-3), loss='sparse_categorical_crossentropy', metrics=["acc"], )模型的训练
(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data() # 归一化 x_train = x_train.astype("float32") / 255 x_test = x_test.astype("float32") / 255 model.fit(tf.expand_dims(x_train,-1), y_train, batch_size=64, epochs=1, validation_split=0.2)注:当loss = =keras.losses.CategoricalCrossentropy(from_logits=True)时,需对标签进行one-hot:
y_train = keras.utils.to_categorical(y_train, 10)