这种“合并”模式与顺序模型不兼容。我认为使用函数式 keras API 更容易keras.Model https://keras.io/models/model/代替keras.Sequential
(主要差异的简短解释 https://jovianlin.io/keras-models-sequential-vs-functional/):
import tensorflow as tf
from tensorflow import keras
import numpy as np
num_data = np.random.random(size=(10,3))
multi_hot_encode_data = np.random.randint(0,2, 30).reshape(10,3)
target = np.eye(3)[np.random.randint(0,3, 10)]
# Use Input layers, specify input shape (dimensions except first)
inp_multi_hot = keras.layers.Input(shape=(multi_hot_encode_data.shape[1],))
inp_num_data = keras.layers.Input(shape=(num_data.shape[1],))
# Bind nulti_hot to embedding layer
emb = keras.layers.Embedding(input_dim=multi_hot_encode_data.shape[1], output_dim=2)(inp_multi_hot)
# Also you need flatten embedded output of shape (?,3,2) to (?, 6) -
# otherwise it's not possible to concatenate it with inp_num_data
flatten = keras.layers.Flatten()(emb)
# Concatenate two layers
conc = keras.layers.Concatenate()([flatten, inp_num_data])
dense1 = keras.layers.Dense(3, activation=tf.nn.relu, )(conc)
# Creating output layer
out = keras.layers.Dense(3, activation=tf.nn.softmax)(dense1)
model = keras.Model(inputs=[inp_multi_hot, inp_num_data], outputs=out)
model.compile(optimizer=tf.train.RMSPropOptimizer(0.01),
loss=keras.losses.categorical_crossentropy,
metrics=[keras.metrics.categorical_accuracy])
- 您应该在连接嵌入层之前压平嵌入层的输出,或者 numeric_data 应该具有兼容的形状和至少三个维度
- 在各层之后定义功能模型。输入和输出可以是单层或可迭代的层
输出model.summary
:
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_5 (InputLayer) (None, 3) 0
__________________________________________________________________________________________________
embedding_2 (Embedding) (None, 3, 2) 6 input_5[0][0]
__________________________________________________________________________________________________
flatten (Flatten) (None, 6) 0 embedding_2[0][0]
__________________________________________________________________________________________________
input_6 (InputLayer) (None, 3) 0
__________________________________________________________________________________________________
concatenate_2 (Concatenate) (None, 9) 0 flatten[0][0]
input_6[0][0]
__________________________________________________________________________________________________
dense (Dense) (None, 3) 30 concatenate_2[0][0]
__________________________________________________________________________________________________
dense_1 (Dense) (None, 3) 12 dense[0][0]
==================================================================================================
Total params: 48
Trainable params: 48
Non-trainable params: 0
__________________________________________________________________________________________________
此外,它也成功适配:
model.fit([multi_hot_encode_data, num_data], target)
Epoch 1/1
10/10 [==============================] - 0s 34ms/step - loss: 1.0623 - categorical_accuracy: 0.3000