RNN LSTM 输入形状应该是什么？

Question

我对 LSTM 输入大小的工作原理感到困惑。

我有一个场景，我试图根据时间序列中的体重和身高来预测一个人的体重。

我似乎无法弄清楚我在尺寸方面哪里出错了：

from datetime import datetime
import numpy as np
import pandas as pd
import math
from keras.layers import Dense, Activation, LSTM, Input, concatenate
from keras.models import Model

def create_dataset(dataset, window_length=1):
    dataX, dataY = [], []
    for i in range(len(dataset)-window_length-1):
        # print("dataX from {} to {}".format(i, i+window_length))
        # print("dataY from {}".format(i+window_length))
        dataX.append(dataset[i:(i+window_length)])
        dataY.append(dataset[(i + window_length):])


    return np.array(dataX), np.array(dataY)

def buildModel(dataLength, labelLength):

    weight = Input(shape=(dataLength, 1), name="weight")
    height = Input(shape=(dataLength, 1), name="height")

    weightLayers = LSTM(64, return_sequences=False)(weight)
    heightLayers = LSTM(64, return_sequences=False)(height)

    output = concatenate([ weightLayers, heightLayers ])

    output = Dense(labelLength, activation="linear", name="weightedAverage_output")(output)

    model = Model(
        inputs=[weight, height],
        outputs=[output]
        )

    model.compile(optimizer="rmsprop", loss="mse")

    return model

bogus = {
    "weight": range(100,200),
    "height": range(150,250)
}

dataset = pd.DataFrame(bogus)



train_size      = int(len(dataset) * 0.90)
test_size       = len(dataset) - train_size
train, test     = dataset[:train_size], dataset[-test_size:]

# print("*" * 30)
# print(train.head())
# print(train.tail())
# print("==> {}".format(len(train)))
# print("*" * 30)
# print(test.head())
# print(test.tail())
# print("==> {}".format(len(test)))
# input(">")

height_train = np.array(train["height"].values.tolist()).reshape((-1, 1)).astype('float32')
weight_train = np.array(train["weight"].values.tolist()).reshape((-1, 1)).astype('float32')

height_test = np.array(test["height"].values.tolist()).reshape((-1, 1)).astype('float32')
weight_test = np.array(test["weight"].values.tolist()).reshape((-1, 1)).astype('float32')

x_train_height, y_train_height  = create_dataset(height_train, 60)
x_train_weight, y_train_weight  = create_dataset(weight_train, 60)

x_test_height, y_test_height  = create_dataset(height_test, 60)
x_test_weight, y_test_weight  = create_dataset(weight_test, 60)



model = buildModel(60,4)
model.fit(
    [ 
        x_train_weight,
        x_train_height,
    ],

    [ 
        y_train_weight
    ],

    validation_data=(
        [
            x_test_weight,
            x_test_height,
        ],
        [
            y_test_weight
        ],
    ),

    epochs=1,
    batch_size=3000,
    callbacks=[
        # board.createTensorboardConfig("log/graph"),
    ]
)

我收到此错误：

ValueError: 检查目标时出错：预期 weightedAverage_output 的形状为 (None, 4) 但得到的数组的形状为 (29, 1)

很确定我在输入和输出维度的某个地方出错了。

有什么想法吗？

Answer 1

首先您需要更改最终输出层的尺寸：

output = Dense(1, activation="linear", name="weightedAverage_output")(output)

其次，您需要更改输入维度以包含时间步长：

timesteps = 1
weight = Input(shape=(dataLength,timesteps,1), name="weight")
height = Input(shape=(dataLength,timesteps,1), name="height")

时间步长可以在 (0,inf) 之间。

您可以参考这篇文章以进一步了解时间步：https://machinelearningmastery.com/use-timesteps-lstm-networks-time-series-forecasting/