Question

我正在使用Django项目的网站上工作。我要以基于关键字的价格为基础，并根据从ebay api数据库调用（关键字）获得的数据做出未来的价格预测和机器学习。我正在使用来自tensorflow的keras基于sklearn.model_selection的train_test_split进行价格预测。我的所有功能和目标（销售价格）数组均在-1和1之间缩小，平均值为0和std。开发。的1。所有的预测都回到1.0，我不知道为什么。看起来我在“过度拟合”模型。我想知道是否有人可以帮助我。我将在神经网络文件下方显示。

neuralNetwork.py

import pandas as pd
import numpy as np
from tensorflow import keras
from sklearn.preprocessing import MinMaxScaler
from sklearn.model_selection import train_test_split

class Neural_Network:

    def neural_network(self, n_df):
        df = n_df.copy()
        df = df.replace('^\s*$', np.nan, regex=True)
        #df['itemId'] = df['itemId'].astype(int)
        df['listingType'] = pd.get_dummies(df['listingType'])
        df['endPrice'] = df['endPrice'].astype(float)
        df['shippingServiceCost'] = df['shippingServiceCost'].astype(float)
        #df['shippingServiceCost'] = df['shippingServiceCost'].interpolate()
        df['shippingServiceCost'] = df['shippingServiceCost'].fillna(df['shippingServiceCost'].mean())
        df['bidCount'] = df['bidCount'].astype(np.float)
        #df['bidCount'] = df['bidCount'].interpolate()
        df['bidCount'] = df['bidCount'].fillna(df['bidCount'].mean())
        df['watchCount'] = df['watchCount'].astype(np.float)
        #df['watchCount'] = df['watchCount'].interpolate()
        df['watchCount'] = df['watchCount'].fillna(df['watchCount'].mean())
        df['returnsAccepted'] = pd.get_dummies(df['returnsAccepted'])
        df['handlingTime'] = df['handlingTime'].astype(int)
        df['sellerUserName'] = pd.get_dummies(df['sellerUserName'])
        df['feedbackScore'] = df['feedbackScore'].astype(int)
        df['positiveFeedbackPercent'] = df['positiveFeedbackPercent'].astype(float)
        df['topRatedSeller'] = pd.get_dummies(df['topRatedSeller'])
        df['endDate'] = pd.get_dummies(df['endDate'])
        print('\nnull values in dataframe are:\n', df.isnull().any())
        features_df = df.drop(['itemId','title','endPrice','location','endTime','startTime','endTimeOfDay'], axis=1)
        num_of_cols = len(features_df.columns)
        features = features_df.values
        target = df.endPrice.values
        print('\ntarget values:\n', target)
        print('\nfeatures values:\n', features)
        print('\ntarget shape:\n', target.shape)
        print('\nfeatures shape:\n', features.shape)
        X_train, X_test, y_train, y_test = train_test_split(features, target, test_size=0.3, random_state=124)
        print('\nTRAIN TEST SPLIT EXECUTED\n')
        X_train = MinMaxScaler(feature_range=(-1,1)).fit_transform(X_train)
        X_test = MinMaxScaler(feature_range=(-1,1)).fit_transform(X_test)
        print('\nX_train and X_test scaled\n')
        y_train = y_train.reshape(-1,1)
        y_test = y_test.reshape(-1,1)
        y_train = MinMaxScaler(feature_range=(-1,1)).fit_transform(y_train)
        y_test = MinMaxScaler(feature_range=(-1,1)).fit_transform(y_test)
        y_train = y_train.reshape(-1)
        y_test = y_test.reshape(-1)
        print('\nshape of X_train:\n', X_train.shape)
        print('\nshape of X_test:\n', X_test.shape)
        print('\nshape of y_train:\n', y_train.shape)
        print('\nshape of y_test:\n', y_test.shape)
        model = keras.Sequential()
        input_layer = keras.layers.Dense(16, input_dim=num_of_cols, activation='sigmoid')
        model.add(input_layer)
        hidden_layer = keras.layers.Dense(num_of_cols, input_dim=16, activation='sigmoid')
        model.add(hidden_layer)
        output_layer = keras.layers.Dense(1, input_dim=num_of_cols, activation='softmax')
        model.add(output_layer)
        model.compile(loss='binary_crossentropy', optimizer='rmsprop', metrics=['accuracy'])
        history = model.fit(X_train, y_train, validation_split=0.2, batch_size=32, epochs=100, shuffle=True)
        predictions = model.predict(X_test, verbose=0, steps=1)
        print('\npredictions shape:\n', predictions.shape)
        pred_nn_df = pd.DataFrame({'predictions':pd.Series(np.round(predictions.reshape(-1),2)),'actual_sell_prices':pd.Series(y_test)})
        return pred_nn_df, history

我将显示数据框列和值的示例，但我无法发布图像并将其复制/粘贴到stackoverflow最终显示出一团糟。因此，我假设您足够了解ebay，可以想象神经网络文件中使用的功能通常具有哪些值。我尝试过在互联网上寻找有类似问题的人，却找不到任何有效的方法。我可以肯定我以某种方式过度拟合了模型。

“预测”数组的示例（每次我得到的输出）：
[1.0 1.0 1.0 1.0 ... 1.0 1.0]
“实际售价”数组的示例（缩小到（-1,1）之间）：
[-0.104930 -0.866221 0.279235 ... 1.000000 -0.201099]

有什么想法吗？

Answer 1

在输出层将softmax与一个神经元一起使用只会产生一个恒定的1.0值。您需要使用适当的激活功能，可能是tanh，因为它在[-1, 1]范围内输出。

我的tensorflow keras模型每次预测总是给出1.0。运作不正常

1 个答案: