LLNet模型实现-训练模型

最新推荐文章于 2024-04-22 19:55:48 发布

Hit_HSW

最新推荐文章于 2024-04-22 19:55:48 发布

阅读量1.3k

点赞数

分类专栏：算法实现机器学习 tensorflow 深度学习文章标签： LLNet模型稀疏自动编码

本文链接：https://blog.csdn.net/hit1524468/article/details/80430672

版权

本文档介绍了LLNet模型的实现，这是一个用于低光照图像增强的深度自动编码器。该模型包含多个层，并使用了Xavier初始化方法。在预训练阶段，通过计算数据损失、KL散度损失和稀疏损失来优化权重；在微调阶段，只优化最后一层的权重。

摘要由CSDN通过智能技术生成

# Ref: LLNet: Deep Autoencoders for Low-light Image Enhancement
# 
# CVPR 2014
# 
# Author: HSW
# Date: 2018-05-11

import tensorflow as tf
import numpy as np 

class LLNet_Model(object):

    def __init__(self, beta_pretrain, lambda_pretrain, lambda_finetune, transfer_function=tf.nn.sigmoid, LLnet_Shape=(289, 847, 578, 289), sparseCoef = 0.05):
        ''' LLNet Model Inputs is 17 x 17 image Patch: 289 pixels '''
        self.LLnet_Shape     = LLnet_Shape
        # Inputs + Outputs
        self.inputs = tf.placeholder(tf.float32, [None, self.LLnet_Shape[0]])
        self.labels = tf.placeholder(tf.float32, [None, self.LLnet_Shape[0]])
        # other initial Params 
        self.transfer        = transfer_function
        self.sparseCoef      = sparseCoef 
        self.beta_pretrain   = beta_pretrain
        self.lambda_pretrain = lambda_pretrain
        self.lambda_finetune = lambda_finetune 
        self.weights         = self.initial_weights()

     self.saver           = tf.train.Saver()

def initial_weights(self): ''' Create LLNet weights and biads ''' all_weights = dict() # Layer1 all_weights['w1'] = tf.Variable(self.xavier_init(self.LLnet_Shape[0], self.LLnet_Shape[1])) all_weights['b1'] = tf.Variable(tf.ones([self.LLnet_Shape[1]], dtype=tf.float32)) # Layer2 all_weights['w2'] = tf.Variable(self.xavier_init(self.LLnet_Shape[1], self.LLnet_Shape[2])) all_weights['b2'] = tf.Variable(tf.ones([self.LLnet_Shape[2]], dtype=tf.float32)) # Layer3 all_weights['w3'] = tf.Variable(self.xavier_init(self.LLnet_Shape[2], self.LLnet_Shape[3])) all_weights['b3'] = tf.Variable(tf.ones([self.LLnet_Shape[3]], dtype=tf.float32)) # Layer4 all_weights['w4'] = tf.Variable(self.xavier_init(self.LLnet_Shape[3],