keras最后一层如何输入单个变量?

问题描述 投票:0回答:2

我试图在 Keras 前馈神经网络的最后一步中包含一个变量。我似乎只有在包含 2 列而不是仅包含一列时才能使其正常工作。这是我的代码示例:

首先我准备主要输入数据集:

import pandas as pd
from keras.models import Model
from keras.layers import Dense, Input, Concatenate
from keras.optimizers import Adam

iris = pd.read_csv('https://raw.githubusercontent.com/mwaskom/seaborn-data/master/iris.csv')
iris.loc[:, 'target'] = (iris.species == "setosa").map(int)
train_x = iris.drop(columns=['target', 'species'])
train_y = iris['target'].map(int)

然后我将

train_x
分成两个单独的数据帧,并将它们输入到不同位置的网络中:

feature_x = train_x.drop(columns='petal_width')
single_feature_x = train_x[['petal_width']]

input_x = Input(shape=feature_x.shape, name='feature_input')
single_input_x = Input(shape=single_feature_x.shape, name='single_input')

x = Dense(4, activation='relu')(input_x)

concat_feat = Concatenate(axis=-1, name='concat_fc')([x, single_input_x])

outputs = Dense(1, activation='sigmoid')(concat_feat)

model = Model(inputs=[input_x, single_input_x], outputs=outputs)
model.compile(loss='binary_crossentropy',
              optimizer=Adam(lr=0.001))

model.fit({'feature_input': feature_x,
           'single_input': single_feature_x},
          train_y,
          epochs=100,
          batch_size=512,
          verbose=1)

这会引发错误:

ValueError: Shape must be rank 2 but is rank 3 for '{{node model_5/concat_fc/concat}} = ConcatV2[N=2, T=DT_FLOAT, Tidx=DT_INT32](model_5/dense_10/Relu, model_5/Cast_1, model_5/concat_fc/concat/axis)' with input shapes: [?,4], [?,1,1], [].

但是,如果我添加这一行,它运行得很好:

feature_x = train_x.drop(columns='petal_width')
single_feature_x = train_x[['petal_width']]
# Add a constant column so the shape becomes (?,2)
single_feature_x.loc[:, 'constant'] = 0

为什么这适用于两列,但不适用于一列?

python tensorflow keras
2个回答
1
投票

您只需正确指定输入形状即可。对于 2D 数据,您只需要传递特征 dim。样品尺寸不是必需的。您只需将输入更正为:

input_x = Input(shape=feature_x.shape[1], name='feature_input')
single_input_x = Input(shape=single_feature_x.shape[1], name='single_input')

这里是正在运行的笔记本


0
投票

嗨,我们如何将数据增强应用于此结构?你有尝试过吗?

© www.soinside.com 2019 - 2024. All rights reserved.