tf 卷积函数 Conv2D & Conv1D

最新推荐文章于 2020-12-20 14:15:07 发布

意慢

最新推荐文章于 2020-12-20 14:15:07 发布

阅读量883

点赞数 1

CC 4.0 BY-SA版权

分类专栏： tensorflow

本文链接：https://blog.youkuaiyun.com/ouyang20110913/article/details/109912170

tensorflow 专栏收录该内容

2 篇文章

订阅专栏

本文通过构建一个多输入多输出的模型，介绍了TensorFlow中Conv2D和Conv1D函数的应用。模型包含两个输入：32x32的RGB图像和包含20个10维单词向量的时间序列数据。经过卷积层处理后，使用全局池化层降低维度，并通过全连接层进行最终的分类和评分预测。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

tf 卷积函数 Conv2D & Conv1D

import tensorflow as tf

# 每一张图片的尺寸是 32*32，通道数是 3
image_input = tf.keras.Input(shape=(32, 32, 3), name='img_input') 
# 每一个句子20个单词，每个单词向量的长度是10
timeseries_input = tf.keras.Input(shape=(20, 10), name='ts_input') 

# 第1个参数表示过滤器的个数（输出的通道数），第2个参数表示2维过滤器的长和宽（tuple2或者整数），output_shape: (30,30,3)
x1 = tf.keras.layers.Conv2D(3, 3)(image_input) 
x1 = tf.keras.layers.GlobalMaxPooling2D()(x1)  # output_shape: (3,)

# 第1个参数表示过滤器的个数（输出的通道数），第2个参数表示1维过滤器的长和宽（tuple1或者整数），output_shape: (18,3)
x2 = tf.keras.layers.Conv1D(3, 3)(timeseries_input)  
x2 = tf.keras.layers.GlobalMaxPooling1D()(x2)  # output_shape: (3,)

x = tf.keras.layers.concatenate([x1, x2]) # output_shape: (6,)

score_output = tf.keras.layers.Dense(1, name='score_output')(x)  # shape: (1,)
class_output = tf.keras.layers.Dense(5, name='class_output')(x)  # shape: (5,)

model = tf.keras.Model(inputs=[image_input, timeseries_input],
                    outputs=[score_output, class_output])

print(model.summary())

tf.keras.utils.plot_model(model, 'multi_input_and_output_model.png', show_shapes=True,dpi=500)

输出结果：

Model: "model"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
img_input (InputLayer)          [(None, 32, 32, 3)]  0                                            
__________________________________________________________________________________________________
ts_input (InputLayer)           [(None, 20, 10)]     0                                            
__________________________________________________________________________________________________
conv2d (Conv2D)                 (None, 30, 30, 3)    84          img_input[0][0]                  
__________________________________________________________________________________________________
conv1d (Conv1D)                 (None, 18, 3)        93          ts_input[0][0]                   
__________________________________________________________________________________________________
global_max_pooling2d (GlobalMax (None, 3)            0           conv2d[0][0]                     
__________________________________________________________________________________________________
global_max_pooling1d (GlobalMax (None, 3)            0           conv1d[0][0]                     
__________________________________________________________________________________________________
concatenate (Concatenate)       (None, 6)            0           global_max_pooling2d[0][0]       
                                                                 global_max_pooling1d[0][0]       
__________________________________________________________________________________________________
score_output (Dense)            (None, 1)            7           concatenate[0][0]                
__________________________________________________________________________________________________
class_output (Dense)            (None, 5)            35          concatenate[0][0]                
==================================================================================================
Total params: 219
Trainable params: 219
Non-trainable params: 0
__________________________________________________________________________________________________
None

model-image