tensorflow2.0——compile-fit实现多输入复合模型

将经过卷积层处理后的feature与非图像特征进行融合

import numpy as np

import tensorflow as tf
import pandas as pd
from my_cnn_net import My_cnn
import os

#   设置GPU相关底层配置
physical_devices = tf.config.experimental.list_physical_devices('GPU')
assert len(physical_devices) > 0, "Not enough GPU hardware devices available"
tf.config.experimental.set_memory_growth(physical_devices[0], True)


path5 = 'D:学姐项目\Nowcastinghubei_CIKM\test_v2\train_nonimg_feature.npy'

data5 = np.load(path5)
print('data5.shape:{}'.format(data5.shape))
trn_pic = pd.read_csv('D:学姐项目\Nowcastinghubei_CIKM\test_v2\' + 'train_SAM_IND.csv', encoding='gbk')
Y = np.reshape(10 * trn_pic.label.values, [-1, 1])
trn_img = np.load('D:学姐项目\Nowcastinghubei_CIKM\test_v2\' + 'train_global_3layer_image.npy')
trn_img = 1.5 * 1e-4 * trn_img ** 2

#   超参数
learn_rate = 1.5e-3
print('lr:{}'.format(learn_rate))
epochs = 200000
data_length = 2300
test_length = 200
train_length = data_length - test_length
# bat_size = train_length
bat_size = 500
my_drop = 0.15
print('mydrop:{}'.format(my_drop))

all_db = tf.data.Dataset.from_tensor_slices((data5[:train_length], trn_img[:train_length], Y[:train_length])).batch(train_length)
train_db = tf.data.Dataset.from_tensor_slices((data5[:train_length], trn_img[:train_length], Y[:train_length])).batch(
    bat_size)
test_db = tf.data.Dataset.from_tensor_slices((data5[train_length:], trn_img[train_length:], Y[train_length:])).batch(
    test_length)

sample = next(iter(train_db))
print('sample[0]:{},sample[1]:{}'.format(sample[0].shape, sample[1].shape))

#   构建复合网络
#   定义输入形状
img_input = tf.keras.layers.Input(shape=(101,101,3))
noimg_input = tf.keras.layers.Input(shape=(43,))

#   卷积模型
#   1
my_conv1 = tf.keras.layers.Conv2D(6, kernel_size=[4, 4], activation=tf.nn.relu,kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=0.1, seed=12),bias_initializer=tf.keras.initializers.constant(0.1))(img_input)
my_conv1 = tf.keras.layers.BatchNormalization()(my_conv1)
my_conv1 = tf.keras.layers.MaxPool2D(pool_size=[3, 3], strides=3, padding='same')(my_conv1)
my_conv1 = tf.keras.layers.Dropout(my_drop)(my_conv1)

#   2
my_conv2 = tf.keras.layers.Conv2D(12, kernel_size=[4, 4], activation=tf.nn.relu,kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=0.1, seed=12),bias_initializer=tf.keras.initializers.constant(0.1))(my_conv1)
my_conv2 = tf.keras.layers.BatchNormalization()(my_conv2)
my_conv2 = tf.keras.layers.MaxPool2D(pool_size=[3, 3], strides=3, padding='same')(my_conv2)
my_conv2 = tf.keras.layers.Dropout(my_drop)(my_conv2)

#   3
my_conv3 = tf.keras.layers.Conv2D(24, kernel_size=[4, 4], activation=tf.nn.relu,kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=0.1, seed=12),bias_initializer=tf.keras.initializers.constant(0.1))(my_conv2)
my_conv3 = tf.keras.layers.BatchNormalization()(my_conv3)
my_conv3 = tf.keras.layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same')(my_conv3)
my_conv3 = tf.keras.layers.Dropout(my_drop)(my_conv3)

#   4
my_conv4 = tf.keras.layers.Conv2D(48, kernel_size=[4, 4], activation=tf.nn.relu,kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=0.1, seed=12),bias_initializer=tf.keras.initializers.constant(0.1))(my_conv3)
my_conv4 = tf.keras.layers.BatchNormalization()(my_conv4)
my_conv4 = tf.keras.layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same')(my_conv4)
my_conv4 = tf.keras.layers.Dropout(my_drop)(my_conv4)
my_conv4 = tf.keras.layers.Flatten()(my_conv4)
# my_conv4 = tf.squeeze(my_conv4)

#   复合模型
concat = tf.keras.layers.concatenate([noimg_input,my_conv4])
out = tf.keras.layers.Dense(45, activation='relu')(concat)
out = tf.keras.layers.Dense(20, activation='relu')(out)
out = tf.keras.layers.Dense(1, activation='relu',bias_initializer=tf.keras.initializers.constant(11.0))(out)

sum_model = tf.keras.models.Model(inputs = [noimg_input,img_input],outputs = out)
sum_model.summary()
# sum_model = tf.keras.models.Model(inputs = img_input,outputs = my_conv4)

#   定义优化器
opt = tf.keras.optimizers.Adam(lr=learn_rate)
sum_model.compile(loss=tf.losses.MSE, optimizer=opt)

sum_model.fit([data5[:train_length], trn_img[:train_length]], [Y[:train_length]],
                   validation_data = ([data5[train_length:], trn_img[train_length:]],[Y[train_length:]]),
                   epochs = epochs,batch_size=bat_size)

网络结构：