实体嵌入与 1d cnn keras R 连接答案

【问题标题】：Entity Embedding concatenate with 1d cnn keras R实体嵌入与 1d cnn keras R 连接
【发布时间】：2019-12-20 17:52:04
【问题描述】：

我正在尝试使用 keras 拟合多输入模型。数据由数值和分类组成，所以我定义了两个输入分支，分类使用实体嵌入，数值使用 1D CNN，这是我的数据：

dim(Y_train)       = c(1000,1)
dim(X_categorical) = c(1000,20)
dim(X_numerical)   = c(1000, 21, 50)

这是我的模型：

library(data.table)
library(keras)
unique_cat_num <- vector()
for(cat in ctgrc_cols){
  unique_cat_num = append(unique_cat_num,length(unique(myDT[,get(cat)])))
}
embed_input <- list()
  for(cat in ctgrc_vars){
    embed_input = append(embed_input,layer_input(shape = 1, name = paste0("input_",cat)))
  }

  embed_dim_vec <- vector()
  for(i in 1:length(ctgrc_vars)){
    embed_dim_vec = append(embed_dim_vec,min((unique_cat_len[i] + 1) %/% 2, 50))
  }

  embed_out <- list()
  for(i in 1:length(ctgrc_vars)){
    layout    = layer_embedding(object = embed_input[[i]], input_dim = unique_cat_len[i] + 1, output_dim = embed_dim_vec[i], name = paste0("embedding_",ctgrc_vars[i]))
    layout    = layer_flatten(layout)
    embed_out = append(embed_out, layout)
  }
Conv_input  <- layer_input(shape = c(21, 50))

conv_flat <- Conv_input %>%
      layer_conv_1d(filters = 64, kernel_size = 3, activation = "relu") %>% 
      layer_conv_1d(filters = 64, kernel_size = 3, activation = "relu") %>% 
      layer_max_pooling_1d() %>% 
      layer_conv_1d(filters = 32, kernel_size = 3, activation = "relu") %>% 
      layer_conv_1d(filters = 32, kernel_size = 3, activation = "relu") %>% 
      layer_global_max_pooling_1d()%>%
      layer_dropout(0.3) %>% 
      layer_dense(units = 32, activation = "relu") %>% 
      layer_flatten() 

output <- layer_concatenate(list(embed_out, conv_flat)) %>% 
    layer_dense(units = 32, activation = "relu") %>% 
    layer_dropout(0.3) %>%
    layer_dense(units = out_neuron, activation = "sigmoid")

model <- keras_model(inputs = list(embed_input, Conv_input), outputs = output)

当我尝试尝试该模型时，我收到此值错误：

ValueError: A `Concatenate` layer requires inputs with matching shapes except for the concat axis. Got inputs shapes: [[(None, 14), (None, 2), (None, 6), (None, 3), (None, 2), (None, 2), (None, 17), (None, 4), (None, 5), (None, 5), (None, 5), (None, 5), (None, 5), (None, 5), (None, 5), (None, 5), 
                                                                                                                          Called from: py_call_impl(callable, dots$args, dots$keywords)

有什么建议吗？

【问题讨论】：

标签： r python-3.x keras

【解决方案1】：

嵌入列表的问题结果！ Entity embedding -> embed_input 是一个列表，因此使用 list(embed_input,Conv_input) 结果是一个嵌入列表，而 keras 不想要一个平面列表。和 embed_out 是一样的。

【讨论】：