我有一个仅包含数字1,0或2的2-D数组。该数组有150行和124列，我尝试使用VAE稍加修改。 但是结果和acc非常差，请帮忙
from __future__ import print_function
import numpy as np
from scipy.stats import norm
from keras.layers import Input, Dense, Lambda
from keras.models import Model
from keras import backend as K
from keras import metrics, losses
from keras.datasets import mnist
class VAE:
def __init__(self, batch_size=100, original_dim =1600, latent_dim = 24, epochs=50, root="IsingMC/", epsilon=0.5):
'''
#Reference
- Auto-Encoding Variational Bayes
https://arxiv.org/abs/1312.6114
This code is taken from Keras VAE tutorial available at https://blog.keras.io/building-autoencoders-in-keras.html
Parameters
----------
batch_size : int, default=100
Size of batches for gradient descent
original_dim : int, default =1600
Number of features
latent_dim: int, default = 100
Dimensionality of the latent space
epochs: int, default = 50
Number of epochs for training
'''
self.batch_size = batch_size
self.original_dim = original_dim
self.latent_dim = latent_dim
self.intermediate_dim = 256
self.epochs = epochs
self.epsilon_std = epsilon
def sampling(self, args):
''' Sampling from the latent variables using the means and log-variances'''
z_mean, z_log_var = args
epsilon = K.random_normal(shape=(K.shape(z_mean)[0], self.latent_dim), mean=0.,
stddev=self.epsilon_std)
return z_mean + K.exp(z_log_var / 2) * epsilon
def build(self):
""" This class method constructs the VAE model
"""
original_dim = self.original_dim
latent_dim = self.latent_dim
intermediate_dim = self.intermediate_dim
# encoder
self.x = Input(shape=(original_dim,))
h = Dense(intermediate_dim, activation='relu')(self.x)
self.z_mean = Dense(latent_dim)(h)
self.z_log_var = Dense(latent_dim)(h)
# note that "output_shape" isn't necessary with the TensorFlow backend
z = Lambda(self.sampling, output_shape=(latent_dim,))([self.z_mean, self.z_log_var])
# we instantiate these layers separately so as to reuse them later
decoder_h = Dense(intermediate_dim, activation='relu')
decoder_mean = Dense(original_dim, activation='sigmoid')
h_decoded = decoder_h(z)
x_decoded_mean = decoder_mean(h_decoded)
#decoder
decoder_input = Input(shape=(latent_dim,))
_h_decoded = decoder_h(decoder_input)
_x_decoded_mean = decoder_mean(_h_decoded)
self.generator = Model(decoder_input, _x_decoded_mean)
# end-to-end VAE model
self.vae = Model(self.x, x_decoded_mean)
# encoder, from inputs to latent space
self.encoder = Model(self.x, self.z_mean)
# decoder
#self.decoder = Model(decoder_input, _x_decoded_mean)
# Compute VAE loss
self.vae.compile(optimizer='rmsprop', loss=self.vae_loss,metrics=['acc'])
# Prints a summary of the architecture used
self.vae.summary()
def vae_loss(self, x, x_decoded_mean):
xent_loss = losses.binary_crossentropy(x, x_decoded_mean)
kl_loss = - 0.5 * K.mean(1 + self.z_log_var - K.square(self.z_mean) - K.exp(self.z_log_var), axis=-1)
return xent_loss + kl_loss
def train(self, x_train, x_test):
from sklearn.preprocessing import minmax_scale
x_train = x_train.reshape((len(x_train), np.prod(x_train.shape[1:]))) # flatten each sample out
x_test = x_test.reshape((len(x_test), np.prod(x_test.shape[1:])))
x_train = minmax_scale(x_train) # this step is required in order to use cross-entropy loss for reconstruction
x_test = minmax_scale(x_train) # scaling features in 0,1 interval
self.vae.fit(x_train, x_train,
shuffle=True,
epochs=self.epochs,
batch_size=self.batch_size,
validation_data=(x_test, x_test)
)
# build a model to project inputs on the latent space
#encoder = Model(self.x, self.z_mean)
def predict_latent(self, xnew):
# build a model to project inputs on the latent space
return self.encoder.predict(xnew)
def generate_decoding(self, znew):
# Generate new fantasy particles
return self.generator.predict(znew)
from pandas import read_csv
from pandas import DataFrame
import keras
import pandas as pd
import numpy as np
from keras.models import Input, Model, Sequential
from keras.layers import Dense,Dropout,Embedding
def columns(df,startwith):
return [col for col in df if col.startswith(startwith)]
data = read_csv('./data0115/d1.csv')
data.fillna(2,inplace=True)
x_col = columns(data,'Q2_')
x=data[x_col].values
x_train,x_test = x,x
to_ca = True
copy = x_test.copy()
if to_ca:
x = keras.utils.to_categorical(x_train,num_classes=3)
x= x.reshape((x.shape[0],x.shape[1]*x.shape[2]))
x_train,x_test = x,x
v = VAE(original_dim=len(x[0]))
v.build()
第35/50集 500/500 [==============================]-0s 24us / step-损耗：0.4306-acc：0.0000e + 00-val_loss：0.4308-val_acc：0.0000e + 00 时代36/50 500/500 [=============================]-0s 24us / step-损耗：0.4293-acc：0.0000e + 00-val_loss：0.4294-val_acc：0.0000e + 00 时代37/50 500/500 [==============================]-0s 24us / step-损耗：0.4292-acc：0.0000e + 00-val_loss：0.4287-val_acc：0.0000e + 00 时代38/50 500/500 [==============================]-0s 24us / step-损耗：0.4282-acc：0.0000e + 00-val_loss：0.4285-val_acc：0.0000e + 00 时代39/50 500/500 [==============================]-0s 24us / step-损耗：0.4262-acc：0.0000e + 00-val_loss：0.4265-val_acc：0.0000e + 00 时代40/50 500/500 [=============================]-0s 24us / step-损耗：0.4250-acc：0.0000e + 00-val_loss：0.4256-val_acc：0.0000e + 00 时代41/50 500/500 [=============================]-0s 24us / step-损耗：0.4271-acc：0.0000e + 00-val_loss：0.4247-val_acc：0.0000e + 00 时代42/50 500/500 [=============================]-0s 24us / step-损耗：0.4226-acc：0.0000e + 00-val_loss：0.4252-val_acc：0.0000e + 00 时代43/50 500/500 [=============================]-0s 25us / step-损耗：0.4214-acc：0.0000e + 00-val_loss：0.4233-val_acc：0.0000e + 00 时代44/50 500/500 [=============================]-0s 25us / step-损耗：0.4237-acc：0.0000e + 00-val_loss：0.4209-val_acc：0.0000e + 00 纪元45/50 500/500 [==============================]-0s 24us / step-损耗：0.4203-acc：0.0000e + 00-val_loss：0.4179-val_acc：0.0000e + 00 时代46/50 500/500 [=============================]-0s 24us / step-损耗：0.4211-acc：0.0000e + 00-val_loss：0.4209-val_acc：0.0000e + 00 时代47/50 500/500 [=============================]-0s 25us / step-损耗：0.4220-acc：0.0000e + 00-val_loss：0.4179-val_acc：0.0000e + 00 时代48/50 500/500 [=============================]-0s 24us / step-损耗：0.4172-acc：0.0000e + 00-val_loss：0.4180-val_acc：0.0000e + 00 时代49/50 500/500 [=============================]-0s 24us / step-损耗：0.4166-acc：0.0000e + 00-val_loss：0.4154-val_acc：0.0000e + 00 时代50/50 500/500 [=============================]-0s 25us / step-损耗：0.4173-acc：0.0000e + 00-val_loss：0.4176-val_acc：0.0000e + 00