我一直在阅读Keras文档,以构建我自己的MLP网络,实现MLP反向传播。我熟悉sklearn中的MLPClassifier,但我想学习Keras
深入学习。以下是第一次尝试。网络有3层1输入(功能= 64),1输出和1隐藏。总数是(64,64,1)。输入为numpy
矩阵X
,125K样本(64 dim),y
是1D numpy
二进制类(1,-1):
# Keras imports
from keras.models import Sequential
from sklearn.model_selection import train_test_split
from keras.layers import Dense, Dropout, Activation
from keras.initializers import RandomNormal, VarianceScaling, RandomUniform
from keras.optimizers import SGD, Adam, Nadam, RMSprop
# System imports
import sys
import os
import numpy as np
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
def train_model(X, y, num_streams, num_stages):
'''
STEP1: Initialize the Model
'''
tr_X, ts_X, tr_y, ts_y = train_test_split(X, y, train_size=.8)
model = initialize_model(num_streams, num_stages)
'''
STEP2: Train the Model
'''
model.compile(loss='binary_crossentropy',
optimizer=Adam(lr=1e-3),
metrics=['accuracy'])
model.fit(tr_X, tr_y,
validation_data=(ts_X, ts_y),
epochs=3,
batch_size=200)
def initialize_model(num_streams, num_stages):
model = Sequential()
hidden_units = 2 ** (num_streams + 1)
# init = VarianceScaling(scale=5.0, mode='fan_in', distribution='normal')
init_bound1 = np.sqrt(3.5 / ((num_stages + 1) + num_stages))
init_bound2 = np.sqrt(3.5 / ((num_stages + 1) + hidden_units))
init_bound3 = np.sqrt(3.5 / (hidden_units + 1))
# drop_out = np.random.uniform(0, 1, 3)
# This is the input layer (that's why you have to state input_dim value)
model.add(Dense(num_stages,
input_dim=num_stages,
activation='relu',
kernel_initializer=RandomUniform(minval=-init_bound1, maxval=init_bound1)))
model.add(Dense(hidden_units,
activation='relu',
kernel_initializer=RandomUniform(minval=-init_bound2, maxval=init_bound2)))
# model.add(Dropout(drop_out[1]))
# This is the output layer
model.add(Dense(1,
activation='sigmoid',
kernel_initializer=RandomUniform(minval=-init_bound3, maxval=init_bound3)))
return model
问题是,在使用X
时,使用相同的数据集y
和MLPClassifier sklearn
可以获得99%的准确率。但是,Keras的准确性很差,如下所示:
Train on 100000 samples, validate on 25000 samples
Epoch 1/3
100000/100000 [==============================] - 1s - loss: -0.5358 - acc: 0.0022 - val_loss: -0.7322 - val_acc: 0.0000e+00
Epoch 2/3
100000/100000 [==============================] - 1s - loss: -0.6353 - acc: 0.0000e+00 - val_loss: -0.7385 - val_acc: 0.0000e+00
Epoch 3/3
100000/100000 [==============================] - 1s - loss: -0.7720 - acc: 9.0000e-05 - val_loss: -0.9474 - val_acc: 5.2000e-04
我不明白为什么?我在这里错过了什么吗?任何帮助表示赞赏。
答案 0 :(得分:0)
在训练模型之前,将标记数据转换为一个热门代码进行检查。
有关一个热门代码查看https://machinelearningmastery.com/why-one-hot-encode-data-in-machine-learning/
的详细信息答案 1 :(得分:0)
我认为问题在于您使用的是sigmoid
输出图层(绑定到[0,1])但是您的类是(1,-1),您需要更改输出值或使用{ {1}}。
此外,keras图层可能具有与sklearn不同的默认参数,请务必查看文档中的参数。
最后一件事,对于tanh
尝试kernel_initializer
,这是一个很好的默认设置。