如何在 keras 中取輸入並行層

Question

我正在嘗試構建一個 model 用於對 12 導聯心電圖信號進行分類。 我想要這樣，每條線索首先經過不同的 conv1D 層，然后我想將它們全部連接起來。 我不知道如何在將輸入提供給 model 的同時拆分輸入。 這是我嘗試過的，但得到了錯誤：

input = keras.Input(shape=(1000, 12))

conv1=(layers.Conv1D(32,(7),activation='relu'))
conv1=conv1(input[:,:,0])

conv2=(layers.Conv1D(32,(7),activation='relu'))
conv2=conv2(input[:,:,1])

conv3=(layers.Conv1D(32,(7),activation='relu'))
conv3=conv3(input[:,:,2])

conv4=(layers.Conv1D(32,(7),activation='relu'))
conv4=conv4(input[:,:,3])

conv5=(layers.Conv1D(32,(7),activation='relu'))
conv5=conv5(input[:,:,4])

conv6=(layers.Conv1D(32,(7),activation='relu'))
conv6=conv6(input[:,:,5])

conv7=(layers.Conv1D(32,(7),activation='relu'))
conv7=conv7(input[:,:,6])

conv8=(layers.Conv1D(32,(7),activation='relu'))
conv8=conv8(input[:,:,7])

conv9=(layers.Conv1D(32,(7),activation='relu'))
conv9=conv9(input[:,:,8])

conv10=(layers.Conv1D(32,(7),activation='relu'))
conv10=conv10(input[:,:,9])

conv11=(layers.Conv1D(32,(7),activation='relu'))
conv11=conv11(input[:,:,10])

conv12=(layers.Conv1D(32,(7),activation='relu'))
conv12=conv12(input[:,:,11])

conv13=tf.keras.layers.concatenate([conv1,conv2,conv3,conv4,conv5,conv6,conv7,conv8,conv9,conv10,conv11,conv12], axis=2)

錯誤說：

ValueError: Input 0 of layer conv1d_6 is incompatible with the layer: expected ndim=3, found ndim=2. Full shape received: [None, 1000]

誰能幫我解決這個問題？

Answer 1

錯誤消息解釋了錯誤， input[:,:,11]將張量從 3 維折疊到 2 維，因此您需要使用input[:,:,11:12]或tf.expand_dims(input[:,:,11], axis=-1)以保留最后一個維度。

import tensorflow as tf
from tensorflow.keras.layers import Conv1D

n_layers = 12
n_filters = 32
k = 7
activation_fn = 'relu'


input = tf.keras.Input(shape=(1000, n_layers))
layers = {
    f'conv{i}': Conv1D(n_filters, (k), activation=activation_fn) for i in range(n_layers)
}

channels = [layers[f'conv{i}'](input[:,:,i:i+1]) for i in range(n_layers)]

convcat = tf.keras.layers.concatenate(channels, axis=2)