时间戳

首页 > 解决方案 > LSTM 3D 输入 3D 输出尺寸不匹配

keras - LSTM 3D 输入 3D 输出尺寸不匹配

问题描述

我在 Keras 有一个简单的 LSTM 网络：

def create_model(x):
    lstm_model = Sequential()
    lstm_model.add(LSTM(100, input_dim=x.shape[2], input_length=x.shape[1]))
    lstm_model.add(Dense(1,activation='sigmoid'))
    lstm_model.compile(loss='mean_squared_error', optimizer='adam')
    return lstm_model

我正在尝试对具有以下形状的数据进行训练：

训练数据输入：(100, 2784, 6)，训练数据输出：(100, 2784, 1)
验证数据输入：(50, 27, 6)，验证数据输出：(50, 27, 1)
测试数据输入： (50, 27, 6)，测试数据输出：(50, 27, 1)

model.fit(train_x, train_y, validation_data=(validation_x, validation_y), epochs=EPOCHS, batch_size=BATCH_SIZE, shuffle=False, callbacks=[...])

我一直在维度上失败，要么是因为 a）验证数据与训练数据的形状不同，要么 b）因为 y 形状不正确

我究竟做错了什么？

为方便起见 PS 独立代码

import numpy as np
import keras
from keras.models import Sequential
from keras.layers import LSTM, Dense

x_train=np.arange(1670400)
x_train=x_train.reshape((100, 2784, 6))
y_train=np.arange(278400)
y_train=y_train.reshape((100, 2784, 1))

x_val=np.arange(8100)
x_val=x_val.reshape((50, 27, 6))
y_val=np.arange(1350)
y_val=y_val.reshape((50, 27, 1))

x_test=np.arange(8100)
x_test=x_test.reshape((50, 27, 6))
y_test=np.arange(1350)
y_test=y_test.reshape((50, 27, 1))


def create_model(x):
    lstm_model = Sequential()
    lstm_model.add(LSTM(100, input_dim=x.shape[2], input_length=x.shape[1]))
    lstm_model.add(Dense(1,activation='sigmoid'))
    lstm_model.compile(loss='mean_squared_error', optimizer='adam')
    return lstm_model

model=create_model(x_train)
model.fit(x_train, y_train, validation_data=(x_val, y_val), epochs=10, batch_size=32, shuffle=False)

标签： keras

解决方案

此修改有效：
更改为batch_input_shape并将return_sequences=True添加到 LSTM 层

import numpy as np
import keras
from keras.models import Sequential
from keras.layers import LSTM, Dense

x_train=np.arange(1670400)
x_train=x_train.reshape((100, 2784, 6))
y_train=np.arange(278400)
y_train=y_train.reshape((100, 2784, 1))

x_val=np.arange(8100)
x_val=x_val.reshape((50, 27, 6))
y_val=np.arange(1350)
y_val=y_val.reshape((50, 27, 1))

x_test=np.arange(8100)
x_test=x_test.reshape((50, 27, 6))
y_test=np.arange(1350)
y_test=y_test.reshape((50, 27, 1))

def create_model():
    lstm_model = Sequential()
    lstm_model.add(LSTM(100, batch_input_shape=(None,None,6), return_sequences=True))
    lstm_model.add(Dense(1, activation='sigmoid'))
    lstm_model.compile(loss='mean_squared_error', optimizer='adam')
    return lstm_model

model=create_model()
model.fit(x_train, y_train, validation_data=(x_val, y_val), epochs=10, batch_size=32, shuffle=False)

推荐阅读