Last updated: 2022-06-27, 11:26
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
序列（Sequential）模型适合于简单的层堆栈，每层只有一个输入张量和一个输出张量。
从语法上来说，下面的序列模型：
# 定义 3 层 Sequential 模型
model = keras.Sequential(
[
layers.Dense(2, activation='relu', name='layer1'),
layers.Dense(3, activation='relu', name='layer2'),
layers.Dense(4, name='layer3')
]
)
# 对示例样本调用模型
x = tf.ones((3, 3))
y = model(x)
与下面的函数定义等价：
# 创建 3 个 layer
layer1 = layers.Dense(2, activation="relu", name="layer1")
layer2 = layers.Dense(3, activation="relu", name="layer2")
layer3 = layers.Dense(4, name="layer3")
# 在测试输入上调用 layers
x = tf.ones((3, 3))
y = layer3(layer2(layer1(x)))
以下情形不适合用 Sequential 模型：
将 layer 列表传递给 Sequential 的构造函数来创建序列模型：
model = keras.Sequential(
[
layers.Dense(2, activation="relu"),
layers.Dense(3, activation="relu"),
layers.Dense(4),
]
)
这些 layers 可以通过 layers 属性访问：
>>> model.layers
[<keras.layers.core.dense.Dense at 0x1e0d4f2a580>,
<keras.layers.core.dense.Dense at 0x1e0d4f2ad90>,
<keras.layers.core.dense.Dense at 0x1e0d4f2aa30>]
也可以通过 add() 逐步添加 layer：
model = keras.Sequential()
model.add(layers.Dense(2, activation="relu"))
model.add(layers.Dense(3, activation="relu"))
model.add(layers.Dense(4))
还可以使用 pop() 方法移除 layer：
>>> model.pop()
>>> print(len(model.layers))
2
可以使用 name 参数为创建的模型命名：
model = keras.Sequential(name="my_sequential")
model.add(layers.Dense(2, activation="relu", name="layer1"))
model.add(layers.Dense(3, activation="relu", name="layer2"))
model.add(layers.Dense(4, name="layer3"))
通常来说，Keras 中的所有 layers 需要知道其输入 shape，才能创建对应的 weights。
>>> layer = layers.Dense(3)
>>> layer.weights # 空 list
[]
>>> # 在测试输入上调用 layer
>>> x = tf.ones((1, 4))
>>> y = layer(x)
>>> layer.weights # Now it has weights, of shape (4, 3) and (3,)
[<tf.Variable 'dense_6/kernel:0' shape=(4, 3) dtype=float32, numpy=
array([[-0.57910645, 0.01889718, 0.81052744],
[-0.6837379 , -0.6663338 , -0.09140235],
[ 0.91084313, -0.40995848, 0.41564345],
[-0.3284163 , -0.52211523, 0.5777253 ]], dtype=float32)>,
<tf.Variable 'dense_6/bias:0' shape=(3,) dtype=float32, numpy=array([0., 0., 0.], dtype=float32)>]
上述行为也适用于 Sequential 模型，在实例化没有输入的 Sequential 模型时，它还没有构建，因此也没有 weights，此时调用 model.weights 会出错。在接受输入数据后，模型才创建 weights：
>>> model = keras.Sequential([
layers.Dense(2, activation='relu'),
layers.Dense(3, activation='relu'),
layers.Dense(4)
]) # 此时没有 weights
>>> model.weights # 会抛出错误
ValueError: Weights for model sequential_4 have not yet been created. Weights are created when the Model is first called on inputs or `build()` is called with an `input_shape`.
>>> model.summary() # 也会抛出错误
ValueError: This model has not yet been built. Build the model first by calling `build()` or by calling the model on a batch of data.
>>> x = tf.ones((1, 4))
>>> y = model(x)
>>> len(model.weights)
6
>>> model.summary()
Model: "sequential_4"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense_7 (Dense) (1, 2) 10
dense_8 (Dense) (1, 3) 9
dense_9 (Dense) (1, 4) 16
=================================================================
Total params: 35
Trainable params: 35
Non-trainable params: 0
以递增的方式构建 Sequential 模型时，并随时使用 summary 显示模型对验证模型的准确性很有效。此时可以通过向模型传递 Input 对象来启动模型，这样它从一开始就知道输入 shape：
model = keras.Sequential()
model.add(keras.Input(shape=(4,)))
model.add(layers.Dense(2, activation='relu'))
model.summary()
Model: "sequential_3"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense_7 (Dense) (None, 2) 10
=================================================================
Total params: 10
Trainable params: 10
Non-trainable params: 0
_________________________________________________________________
需要注意的是，Input 对象没有作为 model.layers 的一部分进行显示，因为它不是一个 layer:
>>> model.layers
[<keras.layers.core.dense.Dense at 0x12ad7175730>]
还有一种更简单的指定输入 shape 的方法，对第一层传入 input_shape 参数即可：
>>> model = keras.Sequential()
>>> model.add(layers.Dense(2, activation='relu', input_shape=(4,)))
>>> model.summary()
Model: "sequential_7"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense_12 (Dense) (None, 2) 10
=================================================================
Total params: 10
Trainable params: 10
Non-trainable params: 0
这样预定义输入 shape 的模型总有 weights 和输出 shape。
一般来说，如果知道输入 shape，对 Sequential 模型建议提前指定 shape。
在构建新的 Sequential 模型时，使用 add() 逐步添加 layers，并使用 summary() 查看模型摘要很有用。例如，可以查看 Conv2D 和 MaxPooling2D 层如何向下采样图像特征：
>>> model = keras.Sequential()
>>> model.add(keras.Input(shape=(250, 250, 3))) # 250x250 RGB 图片
>>> model.add(layers.Conv2D(32, 5, strides=2, activation='relu'))
>>> model.add(layers.Conv2D(32, 3, activation='relu'))
>>> model.add(layers.MaxPooling2D(3))
# 此时输出的 shape 是多少？不好计算，但是可以直接输出
>>> model.summary()
Model: "sequential_8"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d (Conv2D) (None, 123, 123, 32) 2432
conv2d_1 (Conv2D) (None, 121, 121, 32) 9248
max_pooling2d (MaxPooling2D (None, 40, 40, 32) 0
)
=================================================================
Total params: 11,680
Trainable params: 11,680
Non-trainable params: 0
# 此时输出 shape 为 (40, 40, 32)，继续向下采样
>>> model.add(layers.Conv2D(32, 3, activation="relu"))
>>> model.add(layers.Conv2D(32, 3, activation="relu"))
>>> model.add(layers.MaxPooling2D(3))
>>> model.add(layers.Conv2D(32, 3, activation="relu"))
>>> model.add(layers.Conv2D(32, 3, activation="relu"))
>>> model.add(layers.MaxPooling2D(2))
# 然后再次查看
>>> model.summary()
Model: "sequential_8"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d (Conv2D) (None, 123, 123, 32) 2432
conv2d_1 (Conv2D) (None, 121, 121, 32) 9248
max_pooling2d (MaxPooling2D (None, 40, 40, 32) 0
)
conv2d_2 (Conv2D) (None, 38, 38, 32) 9248
conv2d_3 (Conv2D) (None, 36, 36, 32) 9248
max_pooling2d_1 (MaxPooling (None, 12, 12, 32) 0
2D)
conv2d_4 (Conv2D) (None, 10, 10, 32) 9248
conv2d_5 (Conv2D) (None, 8, 8, 32) 9248
max_pooling2d_2 (MaxPooling (None, 4, 4, 32) 0
2D)
=================================================================
Total params: 48,672
Trainable params: 48,672
Non-trainable params: 0
# 此时输出为 4x4 的 feature map，然后添加 global max pooling
>>> model.add(layers.GlobalMaxPooling2D())
# 以及最终的分类层
>>> model.add(layers.Dense(10))
最终模型如下：
>>> model.summary()
Model: "sequential_8"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d (Conv2D) (None, 123, 123, 32) 2432
conv2d_1 (Conv2D) (None, 121, 121, 32) 9248
max_pooling2d (MaxPooling2D (None, 40, 40, 32) 0
)
conv2d_2 (Conv2D) (None, 38, 38, 32) 9248
conv2d_3 (Conv2D) (None, 36, 36, 32) 9248
max_pooling2d_1 (MaxPooling (None, 12, 12, 32) 0
2D)
conv2d_4 (Conv2D) (None, 10, 10, 32) 9248
conv2d_5 (Conv2D) (None, 8, 8, 32) 9248
max_pooling2d_2 (MaxPooling (None, 4, 4, 32) 0
2D)
global_max_pooling2d (Globa (None, 32) 0
lMaxPooling2D)
dense_13 (Dense) (None, 10) 330
=================================================================
Total params: 49,002
Trainable params: 49,002
Non-trainable params: 0
准备好模型后：
序列模型构建好后，其行为和 函数 API 模型类似，即每层都有 input 和 output 属性。这些属性可用来做一些简单的事情，比如用来快速创建模型，提取 Sequential 模型中所有中间层的输出：
initial_model = keras.Sequential(
[
keras.Input(shape=(250, 250, 3)),
layers.Conv2D(32, 5, strides=2, activation="relu"),
layers.Conv2D(32, 3, activation="relu"),
layers.Conv2D(32, 3, activation="relu"),
]
)
feature_extractor = keras.Model(
inputs=initial_model.inputs,
outputs=[layer.output for layer in initial_model.layers],
)
# Call feature extractor on test input.
x = tf.ones((1, 250, 250, 3))
features = feature_extractor(x)
也可以只提取某一层的特征：
initial_model = keras.Sequential(
[
keras.Input(shape=(250, 250, 3)),
layers.Conv2D(32, 5, strides=2, activation="relu"),
layers.Conv2D(32, 3, activation="relu", name="my_intermediate_layer"), # 提取该层的输出
layers.Conv2D(32, 3, activation="relu"),
]
)
feature_extractor = keras.Model(
inputs=initial_model.inputs,
outputs=initial_model.get_layer(name="my_intermediate_layer").output,
)
# Call feature extractor on test input.
x = tf.ones((1, 250, 250, 3))
features = feature_extractor(x)
迁移学习冻结模型的底层，只训练顶层。下面是两个常见的涉及序列模型的迁移学习范本。
首先，假设你有一个序列模型，需要冻结除最后一层外的所有层，此时只需要迭代模型的 model.layers（除最后一层），对每层设置 layer.trainable = False。如下：
model = keras.Sequential([
keras.Input(shape=(784)),
layers.Dense(32, activation='relu'),
layers.Dense(32, activation='relu'),
layers.Dense(32, activation='relu'),
layers.Dense(10),
])
# 加载预训练的权重
model.load_weights(...)
# 冻结最后一层外的所有层
for layer in model.layers[:-1]:
layer.trainable = False
# 重新编译和训练，此时只会更新最后一层的权重
model.compile(...)
model.fit(...)
另一种是使用序列模型堆叠预先训练过的模型和一些新初始化的分类层。如下：
# 加载预训练的基础卷积层
base_model = keras.applications.Xception(
weights='imagenet',
include_top=False,
pooling='avg')
# 冻结基础模型
base_model.trainable = False
# 使用序列模型在顶部添加可训练的分类器
model = keras.Sequential([
base_model,
layers.Dense(1000),
])
# 编译，训练
model.compile(...)
model.fit(...)
使用迁移学习会经常使用这两种模式。