st_helper_maxiter.py

import time
import math
import sys
from glob import glob
import shutil
import os

import torch
torch.manual_seed(0)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
from torch.autograd import Variable
import torch.nn.functional as F
import torch.optim as optim
from imageio import imread, imwrite

import utils
from utils import *
from vgg_pt import *
from pyr_lap import *
from stylize_objectives import objective_class

def style_transfer(stylized_im, content_im, style_path, output_path, scl, long_side, mask, content_weight=0., use_guidance=False, regions=0, coords=0, lr=2e-3):
    
    print('[starting style transfer]')
    
    REPORT_INTERVAL = 50
    RESAMPLE_FREQ = 1
    RESAMPLE_INCREASE_FREQ = 150
    MAX_ITER = 250
    save_ind = 0

    use_pyr=True

    #temp_name = './'+output_path.split('/')[-1].split('.')[0]+'_temp.png'
    temp_name = './'+output_path.split('/')[-1].split('.')[0]+'_'+str(os.getpid())+'.png'

    ### Keep track of current output image for GUI ###
    canvas = aug_canvas(stylized_im, scl, 0)
    imwrite(temp_name, canvas)
    shutil.move(temp_name, output_path)

    #### Define feature extractor ###
    cnn = utils.to_device(Vgg16_pt())

    phi = lambda x: cnn.forward(x)
    phi2 = lambda x, y, z: cnn.forward_cat(x,z,samps=y,forward_func=cnn.forward)


    #### Optimize over laplaccian pyramid instead of pixels directly ####


    ### Define Optimizer ###
    if use_pyr:
        s_pyr = dec_lap_pyr(stylized_im,5)
        s_pyr = [Variable(li.data,requires_grad=True) for li in s_pyr]
    else:
        s_pyr = [Variable(stylized_im.data,requires_grad=True)]

    optimizer =  optim.RMSprop(s_pyr,lr=lr)

    ### Pre-Extract Content Features ###
    z_c = phi(content_im)

    ### Pre-Extract Style Features from a Folder###
    paths = glob(style_path+'*')[::3]

    ### Create Objective Object ###
    objective_wrapper = 0
    objective_wrapper = objective_class(objective='remd_dp_g')
    

    z_s_all = []
    for ri in range(len(regions[1])):
        z_s, style_ims = load_style_folder(phi2, paths, regions,ri, n_samps=-1, subsamps=1000, scale=long_side, inner=5)
        z_s_all.append(z_s)

    ### Extract guidance features if required ###
    gs = np.array([0.])
    if use_guidance:
        gs = load_style_guidance(phi, style_path, coords[:,2:], scale=long_side)


    ### Randomly choose spatial locations to extract features from ###
    if use_pyr:
        stylized_im = syn_lap_pyr(s_pyr)
    else:
        stylized_im = s_pyr[0]

    for ri in range(len(regions[0])):
        

        r_temp = regions[0][ri]
        r_temp = torch.from_numpy(r_temp).unsqueeze(0).unsqueeze(0).contiguous()
        #r = F.upsample(r_temp,(stylized_im.size(3),stylized_im.size(2)),mode='bilinear')[0,0,:,:].numpy()    
        r = F.interpolate(r_temp,(stylized_im.size(3),stylized_im.size(2)),mode='bilinear',align_corners=True)[0,0,:,:].numpy()    

        if r.max()<0.1:
            r = np.greater(r+1.,0.5)
        else:
            r = np.greater(r,0.5)

        objective_wrapper.init_inds(z_c, z_s_all,r,ri)

    if use_guidance:
        objective_wrapper.init_g_inds(coords, stylized_im)

    scl_start = time.time()
    #checkframe=0
    
    for i in range(MAX_ITER):
        
        ### zero out gradients and compute output image from pyramid ##
        optimizer.zero_grad()
        if use_pyr:
            stylized_im = syn_lap_pyr(s_pyr)
        else:
            stylized_im = s_pyr[0]

        ## Dramatically Resample Large Set of Spatial Locations ##
        if i==0 or i%(RESAMPLE_FREQ*10) == 0:
            for ri in range(len(regions[0])):
                
                r_temp = regions[0][ri]
                r_temp = torch.from_numpy(r_temp).unsqueeze(0).unsqueeze(0).contiguous()
                #r = F.upsample(r_temp,(stylized_im.size(3),stylized_im.size(2)),mode='bilinear')[0,0,:,:].numpy()     
                r = F.interpolate(r_temp,(stylized_im.size(3),stylized_im.size(2)),mode='bilinear',align_corners=True)[0,0,:,:].numpy()     

                if r.max()<0.1:
                    r = np.greater(r+1.,0.5)
                else:
                    r = np.greater(r,0.5)

                objective_wrapper.init_inds(z_c, z_s_all,r,ri)

        ## Subsample spatial locations to compute loss over ##
        if i==0 or i%RESAMPLE_FREQ == 0:
            objective_wrapper.shuffle_feature_inds()
        
        ## Extract Features from Current Output
        z_x = phi(stylized_im)

        ## Compute Objective and take gradient step ##
        ell = objective_wrapper.eval(z_x, z_c, z_s_all, gs, 0., content_weight=content_weight,moment_weight=1.0)

        ell.backward()
        optimizer.step()
            
        ## Periodically save output image for GUI ###
        if (i+1)%REPORT_INTERVAL == 0:
            
            #output iter to see evolution
            #checkframe=checkframe+1
            #check_name = './'+output_path.split('/')[-1].split('.')[0]+'_scale'+str(scl)+'.'+str(checkframe)+'.png'
            #print ('check name : ',check_name)
            #imwrite(check_name, canvas)
            
            canvas = aug_canvas(stylized_im, scl, i)
            imwrite(temp_name, canvas)
            
            shutil.move(temp_name, output_path)

        ### Periodically Report Loss and Save Current Image ###
        if (i+1)%REPORT_INTERVAL == 0:
            print('             [st_helper] scale: ',scl,' iters: ' , i+1 , ell.item() ,ell)
            #print((i+1),ell)
            save_ind += 1

    print('             scale : ',scl,' took: ', int(time.time()-scl_start), 'Seconds')
    return stylized_im, ell