主页 / 问题 / 944642
Accepted
Pavel
Asked:2020-02-14 14:46:47 +0000 UTC2020-02-14 14:46:47 +0000 UTC

训练模型时loss和acc的值相同

  • 772

同事，不要告诉我可能是什么问题：

我正在尝试训练一个模型来预测公司的违约情况。有一个包含 100,000 家公司和 37 个特征的数据集。

将数据集拆分为训练和测试

from sklearn.model_selection import train_test_split
train, test = train_test_split(df_findataset, test_size=0.2)
X_train = train.iloc[:,0:36]
Y_train = train.iloc[:,-1] # (0 - нет дефолта, 1- дефолтная компания)

x_test = test.iloc[:,0:36]
y_test = test.iloc[:,-1]

创建了一个模型：

x_train = tf.keras.utils.normalize(x_train, axis=1)
x_test = tf.keras.utils.normalize(x_test, axis=1)

model = tf.keras.models.Sequential()
model.add(tf.keras.layers.Dense(50, input_dim=36, activation='relu'))
model.add(tf.keras.layers.Dense(2, activation='relu'))
model.add(tf.keras.layers.Dense(1, activation='sigmoid'))

model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])

model.fit(x_train, y_train, epochs=10, batch_size=10)

训练时，发出以下内容 - 一切都一样:(

Epoch 1/10
108960/108960 [==============================] - 14s 125us/step - loss: 0.1534 - acc: 0.9645
Epoch 2/10
108960/108960 [==============================] - 13s 123us/step - loss: 0.1534 - acc: 0.9645
Epoch 3/10
108960/108960 [==============================] - 14s 125us/step - loss: 0.1534 - acc: 0.9645
Epoch 4/10
108960/108960 [==============================] - 14s 131us/step - loss: 0.1534 - acc: 0.9645
Epoch 5/10
108960/108960 [==============================] - 15s 139us/step - loss: 0.1534 - acc: 0.9645
Epoch 6/10
108960/108960 [==============================] - 15s 136us/step - loss: 0.1534 - acc: 0.9645
Epoch 7/10
108960/108960 [==============================] - 15s 139us/step - loss: 0.1534 - acc: 0.9645
Epoch 8/10
108960/108960 [==============================] - 16s 143us/step - loss: 0.1534 - acc: 0.9645
Epoch 9/10
108960/108960 [==============================] - 14s 129us/step - loss: 0.1534 - acc: 0.9645
Epoch 10/10
108960/108960 [==============================] - 15s 140us/step - loss: 0.1534 - acc: 0.9645
Out[152]:
<tensorflow.python.keras.callbacks.History at 0x7fed4cb9feb8>

告诉我它可能与什么相关联:(((数据似乎已被尽可能多地清理(尽我所能:())

pandas

1 个回答

  1. Best Answer

    试试这样：

    import pandas as pd
import numpy as np
import tensorflow as tf
from keras import Sequential
from keras.layers import *
from keras.callbacks import *
from keras.models import save_model, load_model
from sklearn.model_selection import train_test_split

filename = r'C:\work\ML\SO\944642-Keras_NN\fin2018.csv'

# let's set 'Регистрационный номер' as an index as 
df = pd.read_csv(filename, index_col=0)

X_train, X_test, Y_train, Y_test = \
  train_test_split(
    tf.keras.utils.normalize(df.drop(['bankrt_status'],axis=1)),
    #df.drop(['Регистрационный номер','bankrt_status'],axis=1),
    df['bankrt_status'], 
    test_size=0.2)


model = Sequential()
model.add(Dense(64, input_dim=X_train.shape[1], activation='relu'))
model.add(Dense(16, activation='relu'))
model.add(Dropout(0.2))
model.add(Dense(1, activation='sigmoid'))

model.compile(loss='binary_crossentropy', optimizer='adamax', metrics=['accuracy'])


model_fn = r'c:\temp\model.h5'
# Keras callbacks
early_stop = EarlyStopping(monitor='val_acc', min_delta=0.0001,
                           patience=5, verbose=1, mode='auto')
chkpt = ModelCheckpoint(str(model_fn), 
                        monitor='val_acc', 
                        verbose=1, 
                        save_best_only=True, 
                        mode='auto')
callbacks = [early_stop, chkpt]


model.fit(X_train, Y_train, validation_split=0.15, epochs=20, batch_size=32,
          callbacks=callbacks)

model = load_model(model_fn)

score, acc = model.evaluate(X_test, Y_test)
print(f"test score: {score:0.3f}\ttest accuracy: {acc:0.4f}")

    教育：

    Train on 92616 samples, validate on 16344 samples
Epoch 1/20
92616/92616 [==============================] - 3s 31us/step - loss: 0.1314 - acc: 0.9646 - val_loss: 0.1185 - val_acc: 0.9637

Epoch 00001: val_acc improved from -inf to 0.96372, saving model to c:\temp\model.h5
Epoch 2/20
92616/92616 [==============================] - 3s 33us/step - loss: 0.1208 - acc: 0.9651 - val_loss: 0.1175 - val_acc: 0.9641

Epoch 00002: val_acc improved from 0.96372 to 0.96415, saving model to c:\temp\model.h5
Epoch 3/20
92616/92616 [==============================] - 3s 28us/step - loss: 0.1192 - acc: 0.9656 - val_loss: 0.1163 - val_acc: 0.9645

Epoch 00003: val_acc improved from 0.96415 to 0.96451, saving model to c:\temp\model.h5
Epoch 4/20
92616/92616 [==============================] - 3s 29us/step - loss: 0.1178 - acc: 0.9655 - val_loss: 0.1160 - val_acc: 0.9645

Epoch 00004: val_acc did not improve from 0.96451
Epoch 5/20
92616/92616 [==============================] - 2s 27us/step - loss: 0.1163 - acc: 0.9655 - val_loss: 0.1152 - val_acc: 0.9646

Epoch 00005: val_acc improved from 0.96451 to 0.96464, saving model to c:\temp\model.h5
Epoch 6/20
92616/92616 [==============================] - 3s 31us/step - loss: 0.1155 - acc: 0.9657 - val_loss: 0.1149 - val_acc: 0.9646

Epoch 00006: val_acc did not improve from 0.96464
Epoch 7/20
92616/92616 [==============================] - 2s 27us/step - loss: 0.1145 - acc: 0.9657 - val_loss: 0.1146 - val_acc: 0.9646

Epoch 00007: val_acc did not improve from 0.96464
Epoch 8/20
92616/92616 [==============================] - 2s 27us/step - loss: 0.1143 - acc: 0.9654 - val_loss: 0.1158 - val_acc: 0.9647

Epoch 00008: val_acc improved from 0.96464 to 0.96470, saving model to c:\temp\model.h5
Epoch 9/20
92616/92616 [==============================] - 3s 29us/step - loss: 0.1132 - acc: 0.9657 - val_loss: 0.1150 - val_acc: 0.9649

Epoch 00009: val_acc improved from 0.96470 to 0.96488, saving model to c:\temp\model.h5
Epoch 10/20
92616/92616 [==============================] - 3s 34us/step - loss: 0.1129 - acc: 0.9658 - val_loss: 0.1140 - val_acc: 0.9648

Epoch 00010: val_acc did not improve from 0.96488
Epoch 11/20
92616/92616 [==============================] - 2s 27us/step - loss: 0.1131 - acc: 0.9657 - val_loss: 0.1135 - val_acc: 0.9647

Epoch 00011: val_acc did not improve from 0.96488
Epoch 12/20
92616/92616 [==============================] - 2s 26us/step - loss: 0.1124 - acc: 0.9657 - val_loss: 0.1142 - val_acc: 0.9648

Epoch 00012: val_acc did not improve from 0.96488
Epoch 13/20
92616/92616 [==============================] - 2s 26us/step - loss: 0.1122 - acc: 0.9656 - val_loss: 0.1136 - val_acc: 0.9648

Epoch 00013: val_acc did not improve from 0.96488
Epoch 14/20
92616/92616 [==============================] - 2s 26us/step - loss: 0.1119 - acc: 0.9658 - val_loss: 0.1142 - val_acc: 0.9648

Epoch 00014: val_acc did not improve from 0.96488
Epoch 00014: early stopping
27241/27241 [==============================] - 0s 14us/step
test score: 0.107       test accuracy: 0.9674

    考试：

    test score: 0.107       test accuracy: 0.9674
    • 2

相关问题