我正在使用带有TensorFlow后端的Keras进行神经网络回归问题。我需要一个RMS Error值来确定我的模型的性能。不幸的是，这不是Keras的开箱即用指标，因此我制定了自定义指标。到现在为止还挺好。我的问题是，我在控制台中的RMS错误值与使用model.evaluate（）和打印到文件时的值不同。关于这是为什么的任何想法？
#! /usr/bin/env python
'''
Trains 7D QuaLiKiz-NN with a single output (efiTG)
'''
from __future__ import print_function
import keras
from keras.models import Sequential
from keras.layers import Dense, Dropout, BatchNormalization
from keras.optimizers import RMSprop, adam, Adam
from keras.initializers import TruncatedNormal
from keras import regularizers
from keras import backend as K
import pandas
import numpy
import sys
import os
# Define new Metric: rmse = Root Mean Square Error
def rmse(y_true, y_pred):
return K.mean(K.sum(K.square( y_true-y_pred )))
# Gets the current file name. Useful for procedurally generating output/log
files.
file_name = os.path.basename(sys.argv[0][:-3])
# Define neural network parameters
batch_size = 10
#num_classes = 1
epochs = 1
# Load Data (which is in HDF5 or .h5 format)
store = pandas.HDFStore("unstable_training_gen2_7D_nions0_flat_filter7.h5")
target_df = store['efiITG_GB'].to_frame() # This one is relatively easy to
train
input_df = store['input']
# Puts inputs and outputs in the same pandas dataframe.
# Also only keeps overlapping entries.
joined_dataFrame = target_df.join(input_df)
# Remove all negative values
joined_dataFrame = joined_dataFrame[(joined_dataFrame['efiITG_GB']>0)
& (joined_dataFrame['Ati']>0)
& (joined_dataFrame['Ate']>0)
& (joined_dataFrame['An']>0)
& (joined_dataFrame['qx']>0)
& (joined_dataFrame['smag']>0)
& (joined_dataFrame['x']>0)
& (joined_dataFrame['Ti_Te']>0)]
# Shuffles dataset
shuffled_joined_dataFrame =
joined_dataFrame.reindex(numpy.random.permutation(
joined_dataFrame.index))
# Normalizes data (no standardization necessary due to dataset design)
shuffled_joined_dataFrame =
shuffled_joined_dataFrame/shuffled_joined_dataFrame \
.max().astype(numpy.float64)
# Creates a pandas dataframe for the outputs
shuffled_clean_output_df = shuffled_joined_dataFrame['efiITG_GB']
# Creates a pandas dataframe for the inputs
shuffled_clean_input_df = shuffled_joined_dataFrame.drop('efiITG_GB',
axis=1)
# Creates training dataset (90% of total data) for outputs
y_train = shuffled_clean_output_df.iloc[:int(
numpy.round(len(shuffled_clean_output_df)*0.9))]
# Creates training dataset (90% of total data) for inputs
x_train = shuffled_clean_input_df.iloc[:int(
numpy.round(len(shuffled_clean_input_df)*0.9))]
# Creates testing dataset (10% of total data) for outputs
y_test = shuffled_clean_output_df.iloc[int(
numpy.round(len(shuffled_clean_output_df)*0.9)):]
# Creates testing dataset (10% of total data) for inputs
x_test = shuffled_clean_input_df.iloc[int(
numpy.round(len(shuffled_clean_input_df)*0.9)):]
# Deletes pandas dataframes that are no longer needed
del target_df, input_df
# Closes the HDFStore. This is good practice.
store.close()
# Define neural network
model = Sequential()
model.add(Dense(30,
activation='tanh',
kernel_initializer=TruncatedNormal(mean=0.0, stddev=0.05,
seed=None),
kernel_regularizer=regularizers.l2(0.0000001),
input_shape=(7,)))
model.add(Dense(30,
activation='tanh',
kernel_initializer=TruncatedNormal(mean=0.0, stddev=0.05,
seed=None),
kernel_regularizer=regularizers.l2(0.0000001)))
model.add(Dense(30,
activation='tanh',
kernel_initializer=TruncatedNormal(mean=0.0, stddev=0.05,
seed=None),
kernel_regularizer=regularizers.l2(0.0000001)))
model.add(Dense(1,
activation='linear'))
#model.add(keras.layers.normalization.BatchNormalization())
model.summary()
# Add CallBacks (including TensorBoard)
tbCallBack = keras.callbacks.TensorBoard(
log_dir='TensorBoard_logs/' + str(file_name), write_graph=True)
EarlyStoppingCallBack = keras.callbacks.EarlyStopping(
monitor='val_loss', min_delta=0, patience=5, verbose=0, mode='auto')
model.compile(loss='mean_squared_error', #categorical_crossentropy
#optimizer='adam',
optimizer=Adam(lr=0.001, beta_1=0.9, beta_2=0.999),
metrics=['accuracy', "mae", "mean_squared_error", rmse])
history = model.fit(x_train, y_train,
batch_size=batch_size,
epochs=epochs,
shuffle=True,
verbose=2,
validation_data=(x_test, y_test),
callbacks=[tbCallBack, EarlyStoppingCallBack])
score = model.evaluate(x_test, y_test, verbose=0)
print('Test loss:', score[0])
print('Test accuracy:', score[1])
print("score")
print(score)
print("model.metrics_names")
print(model.metrics_names)
# Create output file
OutputFile = open("./Loss-Values/" +str(file_name) +".txt", "w+")
OutputFile.write("Test loss: " + str(score[0]) + "\n")
OutputFile.write("Test accuracy: " + str(score[1]) + "\n")
OutputFile.write("val_mean_absolute_error: " +str(score[2]) + "\n")
OutputFile.write("val_mean_squared_error: " +str(score[3]) + "\n")
OutputFile.write("RMSE: " +str(score[4]) + "\n")
OutputFile.close()
# creates a HDF5 file 'my_model.h5'
model.save("./Saved-Networks/" + str(file_name) +".h5")
我只是为了调试目的而训练了1个纪元。我的控制台输出是：
Epoch 1/1
- 76s - loss: 0.0224 - acc: 1.5288e-06 - mean_absolute_error: 0.1145 -
mean_squared_error: 0.0224 - rmse: 0.2237 - val_loss: 0.0116 - val_acc:
0.0000e+00 - val_mean_absolute_error: 0.0787 - val_mean_squared_error:
0.0116 - val_rmse: 0.1156
Test loss: 0.0115768997629
Test accuracy: 0.0
score
[0.011576899762892088, 0.0, 0.078728127306243859, 0.01156135053880431,
0.36995773298057744]
model.metrics_names
['loss', 'acc', 'mean_absolute_error', 'mean_squared_error', 'rmse']
[Finished in 108.322s]
如您所见，
的值[＆＃39;遗失＆＃39; ，＆＃39; acc＆＃39; ，＆＃39; mean_absolute_error＆＃39; ，＆＃39; mean_squared_error＆＃39; ，＆＃39; rmse＆＃39;]
很好并且同意
中的值大纪元1/1
- 76s - 损失：0.0224 - acc：1.5288e-06 - mean_absolute_error：0.1145 -
mean_squared_error：0.0224 - rmse：0.2237 - val_loss：0.0116 - val_acc：
0.0000e + 00 - val_mean_absolute_error：0.0787 - val_mean_squared_error：
0.0116 - val_rmse：0.1156
虽然价值
[＆＃39;丢失＆＃39;＆＃39; acc＆＃39;，＆＃39; mean_absolute_error＆＃39;，＆＃39; mean_squared_error＆＃39;，＆＃39; rmse ＆＃39;]
是不同的。任何想法为什么这个自定义Keras指标 rmse 表现不同？
P.S。我的输出文件是：
Test loss: 0.0115768997629
Test accuracy: 0.0
val_mean_absolute_error: 0.0787281273062
val_mean_squared_error: 0.0115613505388
RMSE: 0.369957732981
答案 0 :(得分：0)
K.mean（K.sum（K.square（y_true-y_pred）））
不是rmse。它是相当的正方形。
您不需要NEW,SUCCESS,ERROR电话。你还需要一个平方根。