fine_tuning.py

# -*- coding: utf-8 -*-
"""
Copyright (c) 2018 Image Processing Research Group of University Federico II of Naples ('GRIP-UNINA').
All rights reserved. This work should only be used for nonprofit purposes.
"""
from inputprep import input_preparation
import theano
from downgrade_images import downgrade_images
import numpy as np
from PNN_finetuning_model import ConvLayer, Network
import os

def fine_tuning(I_MS,I_PAN,model,epochs,ft_dir_out):
    """
    single image fine tuning
    model:       dictionary with all necessary paramters
    epochs:      #epochs of fine tuning
    ft_dir_out:  directory of output for fine tuned network
    """
    
    # load parameters
    residual=model['residual']
    pretrained_lr=model['lr']
    cost=model['cost']
    regol=model['regol']
    patch_size=model['patch_size']
    net_scope=model['net_scope']
    ratio=model['ratio']
    sensor=model['sensor']    
    
    # net building    
    layer=[]
    for i in range(0,len(model['layers']),2):
        layer.append(ConvLayer(model['layers'][i], model['layers'][i+1]))
    net=Network(layer)
    
    
    if not os.path.exists(ft_dir_out):
        os.makedirs(ft_dir_out)
               
    I_MS =np.expand_dims(I_MS,axis=0)
    I_PAN =np.expand_dims(I_PAN,axis=0)
    
    # scope-adaptative learning rate
    img_dim=I_MS.shape[2]
    lr=pretrained_lr*((patch_size-net_scope+1)**2)/((img_dim-net_scope+1)**2)
    
    # downgrade images 
    I_MS_LR,I_PAN_LR=downgrade_images(np.squeeze(I_MS),np.squeeze(I_PAN),ratio,sensor)
    
    
    I_PAN_LR=np.expand_dims(I_PAN_LR,axis=0)
    I_MS=I_MS/(2**np.float32(model['L']))
    
        
    I_input=input_preparation(I_MS_LR,I_PAN_LR,model)
    I_input =np.expand_dims(I_input,axis=0)
    
    I_in=theano.shared(np.asarray((I_input),dtype=theano.config.floatX))
    I_ref=theano.shared(np.asarray((I_MS),dtype=theano.config.floatX))
    
    if residual:
        I_ref=I_ref-I_in[:,:-1,:,:]
    
    #training
    loss=np.zeros((2,1))
    net.SGD(I_in,I_ref,epochs, lr,residual,model,cost,regol,ft_dir_out,loss)