为什么我的keras模型根本不训练？

Question

我的代码是：

from keras.models import Sequential
from keras.layers import Dense, Dropout, Masking
import numpy as np
import pandas as pd

dataset = pd.read_csv("data/train.csv", header=0)
dataset = dataset.fillna(0)

X = dataset.drop(columns=['YearRemodAdd', "Id", "SalePrice"], axis=1)
Y = dataset[['SalePrice']]

X = pd.get_dummies(X, columns=["MSSubClass", "MSZoning",
                               "Street", "Alley", "LotShape",
                               "LandContour", "Utilities", "LotConfig",
                               "LandSlope", "Neighborhood", "Condition1",
                               "Condition2", "BldgType", "HouseStyle",
                               "YearBuilt", "RoofStyle", "RoofMatl",
                               "Exterior1st", "Exterior2nd", "MasVnrType",
                               "ExterQual", "ExterCond", "Foundation",
                               "BsmtQual", "BsmtCond", "BsmtExposure",
                               "BsmtFinType1", "BsmtFinType2", "Heating",
                               "HeatingQC", "CentralAir", "Electrical",
                               "KitchenQual", "Functional", "FireplaceQu",
                               "GarageType", "GarageFinish", "GarageQual",
                               "GarageCond", "PavedDrive", "PoolQC",
                               "Fence", "MiscFeature", "MoSold",
                               "YrSold", "SaleType", "SaleCondition"])

Ymax = Y['SalePrice'].max()
Y = Y['SalePrice'].apply(lambda x: float(x) / Ymax)

input_units = X.shape[1]
print(X)
print(Y)

model = Sequential()
model.add(Dense(input_units, input_dim=input_units, activation='relu'))
model.add(Dense(input_units, activation='relu'))
model.add(Dense(input_units, activation='relu'))

model.add(Dense(1, activation='sigmoid'))

model.compile(loss='binary_crossentropy',
              optimizer='adam', metrics=['mse'])
model.fit(X, Y, epochs=250, batch_size=50,
          shuffle=True, validation_split=0.05, verbose=2)

scores = model.evaluate(X, Y)
print("\n%s: %.2f%%" % (model.metrics_names[1], scores[1]*100))

我的数据像：

Id,MSSubClass,MSZoning,LotFrontage,LotArea,Street,Alley,LotShape,LandContour,Utilities,LotConfig,LandSlope,Neighborhood,Condition1,Condition2,BldgType,HouseStyle,OverallQual,OverallCond,YearBuilt,YearRemodAdd,RoofStyle,RoofMatl,Exterior1st,Exterior2nd,MasVnrType,MasVnrArea,ExterQual,ExterCond,Foundation,BsmtQual,BsmtCond,BsmtExposure,BsmtFinType1,BsmtFinSF1,BsmtFinType2,BsmtFinSF2,BsmtUnfSF,TotalBsmtSF,Heating,HeatingQC,CentralAir,Electrical,1stFlrSF,2ndFlrSF,LowQualFinSF,GrLivArea,BsmtFullBath,BsmtHalfBath,FullBath,HalfBath,BedroomAbvGr,KitchenAbvGr,KitchenQual,TotRmsAbvGrd,Functional,Fireplaces,FireplaceQu,GarageType,GarageYrBlt,GarageFinish,GarageCars,GarageArea,GarageQual,GarageCond,PavedDrive,WoodDeckSF,OpenPorchSF,EnclosedPorch,3SsnPorch,ScreenPorch,PoolArea,PoolQC,Fence,MiscFeature,MiscVal,MoSold,YrSold,SaleType,SaleCondition,SalePrice
1,60,RL,65,8450,Pave,NA,Reg,Lvl,AllPub,Inside,Gtl,CollgCr,Norm,Norm,1Fam,2Story,7,5,2003,2003,Gable,CompShg,VinylSd,VinylSd,BrkFace,196,Gd,TA,PConc,Gd,TA,No,GLQ,706,Unf,0,150,856,GasA,Ex,Y,SBrkr,856,854,0,1710,1,0,2,1,3,1,Gd,8,Typ,0,NA,Attchd,2003,RFn,2,548,TA,TA,Y,0,61,0,0,0,0,NA,NA,NA,0,2,2008,WD,Normal,208500
2,20,RL,80,9600,Pave,NA,Reg,Lvl,AllPub,FR2,Gtl,Veenker,Feedr,Norm,1Fam,1Story,6,8,1976,1976,Gable,CompShg,MetalSd,MetalSd,None,0,TA,TA,CBlock,Gd,TA,Gd,ALQ,978,Unf,0,284,1262,GasA,Ex,Y,SBrkr,1262,0,0,1262,0,1,2,0,3,1,TA,6,Typ,1,TA,Attchd,1976,RFn,2,460,TA,TA,Y,298,0,0,0,0,0,NA,NA,NA,0,5,2007,WD,Normal,181500

我的结果是：

Epoch 123/250
 - 0s - loss: 3.8653 - mean_squared_error: 0.0687 - val_loss: 3.8064 - val_mean_squared_error: 0.0639
Epoch 124/250

像2个纪元之后就卡在那里了。 我该怎么做才能防止它卡得这么快？

Answer 1

看来您正在研究回归问题（即预测连续值）。 至少需要考虑两件事：

1. 正如@Mitiku在注释部分中提到的，数据中有一些NA （即丢失）值。 这是使损失变得nan的原因之一。 删除具有NA值的行，或者将NA值替换为特定值（例如0）。有关处理缺失数据的更多信息，请参见此答案 。

2. 使用accuracy作为回归问题的度量标准没有意义，因为它仅对分类任务有效。 而是使用诸如mse （即均方误差）或mae （即平均绝对误差）之类的回归度量。

请在代码中应用以上两点，然后报告培训的进行情况，我将根据需要更新此答案。