快捷導(dǎo)航

解讀tf.keras.layers模塊中的函數(shù)

更新時(shí)間：2023年02月21日 10:27:04 作者：. 小王同學(xué).

這篇文章主要介紹了tf.keras.layers模塊中的函數(shù)，具有很好的參考價(jià)值，希望對(duì)大家有所幫助。如有錯(cuò)誤或未考慮完全的地方，望不吝賜教

tf.keras.layers模塊中的函數(shù)

from __future__ import print_function as _print_function

import sys as _sys

from . import experimental
from tensorflow.python.keras.engine.base_layer import Layer
from tensorflow.python.keras.engine.input_layer import Input
from tensorflow.python.keras.engine.input_layer import InputLayer
from tensorflow.python.keras.engine.input_spec import InputSpec
from tensorflow.python.keras.feature_column.dense_features_v2 import DenseFeatures
from tensorflow.python.keras.layers.advanced_activations import ELU
from tensorflow.python.keras.layers.advanced_activations import LeakyReLU
from tensorflow.python.keras.layers.advanced_activations import PReLU
from tensorflow.python.keras.layers.advanced_activations import ReLU
from tensorflow.python.keras.layers.advanced_activations import Softmax
from tensorflow.python.keras.layers.advanced_activations import ThresholdedReLU
from tensorflow.python.keras.layers.convolutional import Conv1D
from tensorflow.python.keras.layers.convolutional import Conv1D as Convolution1D
from tensorflow.python.keras.layers.convolutional import Conv1DTranspose
from tensorflow.python.keras.layers.convolutional import Conv1DTranspose as Convolution1DTranspose
from tensorflow.python.keras.layers.convolutional import Conv2D
from tensorflow.python.keras.layers.convolutional import Conv2D as Convolution2D
from tensorflow.python.keras.layers.convolutional import Conv2DTranspose
from tensorflow.python.keras.layers.convolutional import Conv2DTranspose as Convolution2DTranspose
from tensorflow.python.keras.layers.convolutional import Conv3D
from tensorflow.python.keras.layers.convolutional import Conv3D as Convolution3D
from tensorflow.python.keras.layers.convolutional import Conv3DTranspose
from tensorflow.python.keras.layers.convolutional import Conv3DTranspose as Convolution3DTranspose
from tensorflow.python.keras.layers.convolutional import Cropping1D
from tensorflow.python.keras.layers.convolutional import Cropping2D
from tensorflow.python.keras.layers.convolutional import Cropping3D
from tensorflow.python.keras.layers.convolutional import DepthwiseConv2D
from tensorflow.python.keras.layers.convolutional import SeparableConv1D
from tensorflow.python.keras.layers.convolutional import SeparableConv1D as SeparableConvolution1D
from tensorflow.python.keras.layers.convolutional import SeparableConv2D
from tensorflow.python.keras.layers.convolutional import SeparableConv2D as SeparableConvolution2D
from tensorflow.python.keras.layers.convolutional import UpSampling1D
from tensorflow.python.keras.layers.convolutional import UpSampling2D
from tensorflow.python.keras.layers.convolutional import UpSampling3D
from tensorflow.python.keras.layers.convolutional import ZeroPadding1D
from tensorflow.python.keras.layers.convolutional import ZeroPadding2D
from tensorflow.python.keras.layers.convolutional import ZeroPadding3D
from tensorflow.python.keras.layers.convolutional_recurrent import ConvLSTM2D
from tensorflow.python.keras.layers.core import Activation
from tensorflow.python.keras.layers.core import ActivityRegularization
from tensorflow.python.keras.layers.core import Dense
from tensorflow.python.keras.layers.core import Dropout
from tensorflow.python.keras.layers.core import Flatten
from tensorflow.python.keras.layers.core import Lambda
from tensorflow.python.keras.layers.core import Masking
from tensorflow.python.keras.layers.core import Permute
from tensorflow.python.keras.layers.core import RepeatVector
from tensorflow.python.keras.layers.core import Reshape
from tensorflow.python.keras.layers.core import SpatialDropout1D
from tensorflow.python.keras.layers.core import SpatialDropout2D
from tensorflow.python.keras.layers.core import SpatialDropout3D
from tensorflow.python.keras.layers.dense_attention import AdditiveAttention
from tensorflow.python.keras.layers.dense_attention import Attention
from tensorflow.python.keras.layers.embeddings import Embedding
from tensorflow.python.keras.layers.local import LocallyConnected1D
from tensorflow.python.keras.layers.local import LocallyConnected2D
from tensorflow.python.keras.layers.merge import Add
from tensorflow.python.keras.layers.merge import Average
from tensorflow.python.keras.layers.merge import Concatenate
from tensorflow.python.keras.layers.merge import Dot
from tensorflow.python.keras.layers.merge import Maximum
from tensorflow.python.keras.layers.merge import Minimum
from tensorflow.python.keras.layers.merge import Multiply
from tensorflow.python.keras.layers.merge import Subtract
from tensorflow.python.keras.layers.merge import add
from tensorflow.python.keras.layers.merge import average
from tensorflow.python.keras.layers.merge import concatenate
from tensorflow.python.keras.layers.merge import dot
from tensorflow.python.keras.layers.merge import maximum
from tensorflow.python.keras.layers.merge import minimum
from tensorflow.python.keras.layers.merge import multiply
from tensorflow.python.keras.layers.merge import subtract
from tensorflow.python.keras.layers.noise import AlphaDropout
from tensorflow.python.keras.layers.noise import GaussianDropout
from tensorflow.python.keras.layers.noise import GaussianNoise
from tensorflow.python.keras.layers.normalization import LayerNormalization
from tensorflow.python.keras.layers.normalization_v2 import BatchNormalization
from tensorflow.python.keras.layers.pooling import AveragePooling1D
from tensorflow.python.keras.layers.pooling import AveragePooling1D as AvgPool1D
from tensorflow.python.keras.layers.pooling import AveragePooling2D
from tensorflow.python.keras.layers.pooling import AveragePooling2D as AvgPool2D
from tensorflow.python.keras.layers.pooling import AveragePooling3D
from tensorflow.python.keras.layers.pooling import AveragePooling3D as AvgPool3D
from tensorflow.python.keras.layers.pooling import GlobalAveragePooling1D
from tensorflow.python.keras.layers.pooling import GlobalAveragePooling1D as GlobalAvgPool1D
from tensorflow.python.keras.layers.pooling import GlobalAveragePooling2D
from tensorflow.python.keras.layers.pooling import GlobalAveragePooling2D as GlobalAvgPool2D
from tensorflow.python.keras.layers.pooling import GlobalAveragePooling3D
from tensorflow.python.keras.layers.pooling import GlobalAveragePooling3D as GlobalAvgPool3D
from tensorflow.python.keras.layers.pooling import GlobalMaxPooling1D
from tensorflow.python.keras.layers.pooling import GlobalMaxPooling1D as GlobalMaxPool1D
from tensorflow.python.keras.layers.pooling import GlobalMaxPooling2D
from tensorflow.python.keras.layers.pooling import GlobalMaxPooling2D as GlobalMaxPool2D
from tensorflow.python.keras.layers.pooling import GlobalMaxPooling3D
from tensorflow.python.keras.layers.pooling import GlobalMaxPooling3D as GlobalMaxPool3D
from tensorflow.python.keras.layers.pooling import MaxPooling1D
from tensorflow.python.keras.layers.pooling import MaxPooling1D as MaxPool1D
from tensorflow.python.keras.layers.pooling import MaxPooling2D
from tensorflow.python.keras.layers.pooling import MaxPooling2D as MaxPool2D
from tensorflow.python.keras.layers.pooling import MaxPooling3D
from tensorflow.python.keras.layers.pooling import MaxPooling3D as MaxPool3D
from tensorflow.python.keras.layers.recurrent import AbstractRNNCell
from tensorflow.python.keras.layers.recurrent import RNN
from tensorflow.python.keras.layers.recurrent import SimpleRNN
from tensorflow.python.keras.layers.recurrent import SimpleRNNCell
from tensorflow.python.keras.layers.recurrent import StackedRNNCells
from tensorflow.python.keras.layers.recurrent_v2 import GRU
from tensorflow.python.keras.layers.recurrent_v2 import GRUCell
from tensorflow.python.keras.layers.recurrent_v2 import LSTM
from tensorflow.python.keras.layers.recurrent_v2 import LSTMCell
from tensorflow.python.keras.layers.serialization import deserialize
from tensorflow.python.keras.layers.serialization import serialize
from tensorflow.python.keras.layers.wrappers import Bidirectional
from tensorflow.python.keras.layers.wrappers import TimeDistributed
from tensorflow.python.keras.layers.wrappers import Wrapper

del _print_function

匯總tf.keras模型層layers

tf.keras.layers.Dense()：密集連接層。參數(shù)個(gè)數(shù) = 輸入層特征數(shù)× 輸出層特征數(shù)(weight)＋輸出層特征數(shù)(bias)
tf.keras.layers.Activation()：激活函數(shù)層。一般放在Dense層后面，等價(jià)于在Dense層中指定activation。
tf.keras.layers.Dropout()：隨機(jī)置零層。訓(xùn)練期間以一定幾率將輸入置0，一種正則化手段。
tf.keras.layers.BatchNormalization()：批標(biāo)準(zhǔn)化層。通過線性變換將輸入批次縮放平移到穩(wěn)定的均值和標(biāo)準(zhǔn)差?？梢栽鰪?qiáng)模型對(duì)輸入不同分布的適應(yīng)性，加快模型訓(xùn)練速度，有輕微正則化效果。一般在激活函數(shù)之前使用。
tf.keras.layers.SpatialDropout2D()：空間隨機(jī)置零層。訓(xùn)練期間以一定幾率將整個(gè)特征圖置0，一種正則化手段，有利于避免特征圖之間過高的相關(guān)性。
tf.keras.layers.Input()：輸入層。通常使用Functional API方式構(gòu)建模型時(shí)作為第一層。
tf.keras.layers.DenseFeature()：特征列接入層，用于接收一個(gè)特征列列表并產(chǎn)生一個(gè)密集連接層。
tf.keras.layers.Flatten()：壓平層，用于將多維張量壓成一維。
tf.keras.layers.Reshape()：形狀重塑層，改變輸入張量的形狀。
tf.keras.layers.Concatenate()：拼接層，將多個(gè)張量在某個(gè)維度上拼接。
tf.keras.layers.Add()：加法層。
tf.keras.layers.Subtract()：減法層。
tf.keras.layers.Maximum()：取最大值層。
tf.keras.layers.Minimum()：取最小值層。

卷積網(wǎng)絡(luò)相關(guān)層

tf.keras.layers.Conv1D()：普通一維卷積，常用于文本。參數(shù)個(gè)數(shù) = 輸入通道數(shù)×卷積核尺寸(如3)×卷積核個(gè)數(shù)
tf.keras.layers.Conv2D()：普通二維卷積，常用于圖像。參數(shù)個(gè)數(shù) = 輸入通道數(shù)×卷積核尺寸(如3乘3)×卷積核個(gè)數(shù)
tf.keras.layers.Conv3D()：普通三維卷積，常用于視頻。參數(shù)個(gè)數(shù) = 輸入通道數(shù)×卷積核尺寸(如3乘3乘3)×卷積核個(gè)數(shù)
tf.keras.layers.SeparableConv2D()：二維深度可分離卷積層。不同于普通卷積同時(shí)對(duì)區(qū)域和通道操作，深度可分離卷積先操作區(qū)域，再操作通道。即先對(duì)每個(gè)通道做獨(dú)立卷即先操作區(qū)域，再用1乘1卷積跨通道組合即再操作通道。參數(shù)個(gè)數(shù) = 輸入通道數(shù)×卷積核尺寸 + 輸入通道數(shù)×1×1×輸出通道數(shù)。深度可分離卷積的參數(shù)數(shù)量一般遠(yuǎn)小于普通卷積，效果一般也更好。
tf.keras.layers.DepthwiseConv2D()：二維深度卷積層。僅有SeparableConv2D前半部分操作，即只操作區(qū)域，不操作通道，一般輸出通道數(shù)和輸入通道數(shù)相同，但也可以通過設(shè)置depth_multiplier讓輸出通道為輸入通道的若干倍數(shù)。輸出通道數(shù) = 輸入通道數(shù) × depth_multiplier。參數(shù)個(gè)數(shù) = 輸入通道數(shù)×卷積核尺寸× depth_multiplier。
tf.keras.layers.Conv2DTranspose()：二維卷積轉(zhuǎn)置層，俗稱反卷積層。并非卷積的逆操作，但在卷積核相同的情況下，當(dāng)其輸入尺寸是卷積操作輸出尺寸的情況下，卷積轉(zhuǎn)置的輸出尺寸恰好是卷積操作的輸入尺寸。
tf.keras.layers.LocallyConnected2D():二維局部連接層。類似Conv2D，唯一的差別是沒有空間上的權(quán)值共享，所以其參數(shù)個(gè)數(shù)遠(yuǎn)高于二維卷積。
tf.keras.layers.MaxPooling2D():二維最大池化層。也稱作下采樣層。池化層無(wú)參數(shù)，主要作用是降維。
tf.keras.layers.AveragePooling2D():二維平均池化層。
tf.keras.layers.GlobalMaxPool2D():全局最大池化層。每個(gè)通道僅保留一個(gè)值。一般從卷積層過渡到全連接層時(shí)使用，是Flatten的替代方案。
tf.keras.layers.GlobalAvgPool2D():全局平均池化層。每個(gè)通道僅保留一個(gè)值。

示例代碼一、搭建LeNet-5神經(jīng)網(wǎng)絡(luò)

import tensorflow as tf
from tensorflow.keras import datasets, layers, optimizers, Sequential, metrics, losses
 
# 1.數(shù)據(jù)集準(zhǔn)備
(x, y), (x_val, y_val) = datasets.mnist.load_data()  # 加載數(shù)據(jù)集，返回的是兩個(gè)元組，分別表示訓(xùn)練集和測(cè)試集
x = tf.convert_to_tensor(x, dtype=tf.float32) / 255.  # 轉(zhuǎn)換為張量，并縮放到0~1
y = tf.convert_to_tensor(y, dtype=tf.int32)  # 轉(zhuǎn)換為張量（標(biāo)簽）
print(x.shape, y.shape)
train_dataset = tf.data.Dataset.from_tensor_slices((x, y))  # 構(gòu)建數(shù)據(jù)集對(duì)象
train_dataset = train_dataset.batch(32).repeat(10)  # 設(shè)置批量訓(xùn)練的batch為32，要將訓(xùn)練集重復(fù)訓(xùn)練10遍
 
# 2.搭建網(wǎng)絡(luò)
network = Sequential([  # 搭建網(wǎng)絡(luò)容器
    layers.Conv2D(6, kernel_size=3, strides=1),  # 第一個(gè)卷積層，6個(gè)3*3*1卷積核
    layers.MaxPooling2D(pool_size=2, strides=2),  # 池化層，卷積核2*2，步長(zhǎng)2
    layers.ReLU(),  # 激活函數(shù)
    layers.Conv2D(16, kernel_size=3, strides=1),  # 第二個(gè)卷積層，16個(gè)3*3*6卷積核
    layers.MaxPooling2D(pool_size=2, strides=2),  # 池化層
    layers.ReLU(),  # 激活函數(shù)
    layers.Flatten(),  # 拉直，方便全連接層處理
    layers.Dense(120, activation='relu'),  # 全連接層，120個(gè)節(jié)點(diǎn)
    layers.Dense(84, activation='relu'),  # 全連接層，84個(gè)節(jié)點(diǎn)
    layers.Dense(10)  # 輸出層，10個(gè)節(jié)點(diǎn)
])
network.build(input_shape=(None, 28, 28, 1))  # 定義輸入,batch_size=32,輸入圖片大小是28*28,通道數(shù)為1。
network.summary()  # 顯示出每層的待優(yōu)化參數(shù)量
 
# 3.模型訓(xùn)練（計(jì)算梯度，迭代更新網(wǎng)絡(luò)參數(shù)）
optimizer = optimizers.SGD(lr=0.01)  # 聲明采用批量隨機(jī)梯度下降方法，學(xué)習(xí)率=0.01
acc_meter = metrics.Accuracy()  # 新建accuracy測(cè)量器
for step, (x, y) in enumerate(train_dataset):  # 一次輸入batch組數(shù)據(jù)進(jìn)行訓(xùn)練
    with tf.GradientTape() as tape:  # 構(gòu)建梯度記錄環(huán)境
        x = tf.reshape(x, (32, 28, 28, 1))  # 將輸入拉直，[b,28,28]->[b,784]
        # x = tf.extand_dims(x, axis=3)
        out = network(x)  # 輸出[b, 10]
        y_onehot = tf.one_hot(y, depth=10)  # one-hot編碼
        loss = tf.square(out - y_onehot)
        loss = tf.reduce_sum(loss) / 32  # 定義均方差損失函數(shù)，注意此處的32對(duì)應(yīng)為batch的大小
        grads = tape.gradient(loss, network.trainable_variables)  # 計(jì)算網(wǎng)絡(luò)中各個(gè)參數(shù)的梯度
        optimizer.apply_gradients(zip(grads, network.trainable_variables))  # 更新網(wǎng)絡(luò)參數(shù)
        acc_meter.update_state(tf.argmax(out, axis=1), y)  # 比較預(yù)測(cè)值與標(biāo)簽，并計(jì)算精確度（寫入數(shù)據(jù)，進(jìn)行求精度）
 
    if step % 200 == 0:  # 每200個(gè)step，打印一次結(jié)果
        print('Step', step, ': Loss is: ', float(loss), ' Accuracy: ', acc_meter.result().numpy())  # 讀取數(shù)據(jù)
        acc_meter.reset_states()  # 清零測(cè)量器l