WiggleGAN / WiggleGAN.py
Rodrigo_Cobo
added thesis
cc6c676
raw
history blame
36.3 kB
import utils, torch, time, os, pickle
import numpy as np
import torch.nn as nn
import torch.cuda as cu
import torch.optim as optim
import pickle
from torchvision import transforms
from torchvision.utils import save_image
from utils import augmentData, RGBtoL, LtoRGB
from PIL import Image
from dataloader import dataloader
from torch.autograd import Variable
import matplotlib.pyplot as plt
import random
from datetime import date
from statistics import mean
from architectures import depth_generator_UNet, \
depth_discriminator_noclass_UNet
class WiggleGAN(object):
def __init__(self, args):
# parameters
self.epoch = args.epoch
self.sample_num = 100
self.nCameras = args.cameras
self.batch_size = args.batch_size
self.save_dir = args.save_dir
self.result_dir = args.result_dir
self.dataset = args.dataset
self.log_dir = args.log_dir
self.gpu_mode = args.gpu_mode
self.model_name = args.gan_type
self.input_size = args.input_size
self.class_num = (args.cameras - 1) * 2 # un calculo que hice en paint
self.sample_num = self.class_num ** 2
self.imageDim = args.imageDim
self.epochVentaja = args.epochV
self.cantImages = args.cIm
self.visdom = args.visdom
self.lambdaL1 = args.lambdaL1
self.depth = args.depth
self.name_wiggle = args.name_wiggle
self.clipping = args.clipping
self.WGAN = False
if (self.clipping > 0):
self.WGAN = True
self.seed = str(random.randint(0, 99999))
self.seed_load = args.seedLoad
self.toLoad = False
if (self.seed_load != "-0000"):
self.toLoad = True
self.zGenFactor = args.zGF
self.zDisFactor = args.zDF
self.bFactor = args.bF
self.CR = False
if (self.zGenFactor > 0 or self.zDisFactor > 0 or self.bFactor > 0):
self.CR = True
self.expandGen = args.expandGen
self.expandDis = args.expandDis
self.wiggleDepth = args.wiggleDepth
self.wiggle = False
if (self.wiggleDepth > 0):
self.wiggle = True
# load dataset
self.onlyGen = args.lrD <= 0
if not self.wiggle:
self.data_loader = dataloader(self.dataset, self.input_size, self.batch_size, self.imageDim, split='train',
trans=not self.CR)
self.data_Validation = dataloader(self.dataset, self.input_size, self.batch_size, self.imageDim,
split='validation')
self.dataprint = self.data_Validation.__iter__().__next__()
data = self.data_loader.__iter__().__next__().get('x_im')
if not self.onlyGen:
self.D = depth_discriminator_noclass_UNet(input_dim=3, output_dim=1, input_shape=data.shape,
class_num=self.class_num,
expand_net=self.expandDis, depth = self.depth, wgan = self.WGAN)
self.D_optimizer = optim.Adam(self.D.parameters(), lr=args.lrD, betas=(args.beta1, args.beta2))
self.data_Test = dataloader(self.dataset, self.input_size, self.batch_size, self.imageDim, split='test')
self.dataprint_test = self.data_Test.__iter__().__next__()
# networks init
self.G = depth_generator_UNet(input_dim=4, output_dim=3, class_num=self.class_num, expand_net=self.expandGen, depth = self.depth)
self.G_optimizer = optim.Adam(self.G.parameters(), lr=args.lrG, betas=(args.beta1, args.beta2))
if self.gpu_mode:
self.G.cuda()
if not self.wiggle and not self.onlyGen:
self.D.cuda()
self.BCE_loss = nn.BCELoss().cuda()
self.CE_loss = nn.CrossEntropyLoss().cuda()
self.L1 = nn.L1Loss().cuda()
self.MSE = nn.MSELoss().cuda()
self.BCEWithLogitsLoss = nn.BCEWithLogitsLoss().cuda()
else:
self.BCE_loss = nn.BCELoss()
self.CE_loss = nn.CrossEntropyLoss()
self.MSE = nn.MSELoss()
self.L1 = nn.L1Loss()
self.BCEWithLogitsLoss = nn.BCEWithLogitsLoss()
print('---------- Networks architecture -------------')
utils.print_network(self.G)
if not self.wiggle and not self.onlyGen:
utils.print_network(self.D)
print('-----------------------------------------------')
temp = torch.zeros((self.class_num, 1))
for i in range(self.class_num):
temp[i, 0] = i
temp_y = torch.zeros((self.sample_num, 1))
for i in range(self.class_num):
temp_y[i * self.class_num: (i + 1) * self.class_num] = temp
self.sample_y_ = torch.zeros((self.sample_num, self.class_num)).scatter_(1, temp_y.type(torch.LongTensor), 1)
if self.gpu_mode:
self.sample_y_ = self.sample_y_.cuda()
if (self.toLoad):
self.load()
def train(self):
if self.visdom:
random.seed(time.time())
today = date.today()
vis = utils.VisdomLinePlotter(env_name='Cobo_depth_Train-Plots_' + str(today) + '_' + self.seed)
visValidation = utils.VisdomLinePlotter(env_name='Cobo_depth_Train-Plots_' + str(today) + '_' + self.seed)
visEpoch = utils.VisdomLineTwoPlotter(env_name='Cobo_depth_Train-Plots_' + str(today) + '_' + self.seed)
visImages = utils.VisdomImagePlotter(env_name='Cobo_depth_Images_' + str(today) + '_' + self.seed)
visImagesTest = utils.VisdomImagePlotter(env_name='Cobo_depth_ImagesTest_' + str(today) + '_' + self.seed)
visLossGTest = utils.VisdomLinePlotter(env_name='Cobo_depth_Train-Plots_' + str(today) + '_' + self.seed)
visLossGValidation = utils.VisdomLinePlotter(env_name='Cobo_depth_Train-Plots_' + str(today) + '_' + self.seed)
visLossDTest = utils.VisdomLinePlotter(env_name='Cobo_depth_Train-Plots_' + str(today) + '_' + self.seed)
visLossDValidation = utils.VisdomLinePlotter(env_name='Cobo_depth_Train-Plots_' + str(today) + '_' + self.seed)
self.train_hist = {}
self.epoch_hist = {}
self.details_hist = {}
self.train_hist['D_loss_train'] = []
self.train_hist['G_loss_train'] = []
self.train_hist['D_loss_Validation'] = []
self.train_hist['G_loss_Validation'] = []
self.train_hist['per_epoch_time'] = []
self.train_hist['total_time'] = []
self.details_hist['G_T_Comp_im'] = []
self.details_hist['G_T_BCE_fake_real'] = []
self.details_hist['G_T_Cycle'] = []
self.details_hist['G_zCR'] = []
self.details_hist['G_V_Comp_im'] = []
self.details_hist['G_V_BCE_fake_real'] = []
self.details_hist['G_V_Cycle'] = []
self.details_hist['D_T_BCE_fake_real_R'] = []
self.details_hist['D_T_BCE_fake_real_F'] = []
self.details_hist['D_zCR'] = []
self.details_hist['D_bCR'] = []
self.details_hist['D_V_BCE_fake_real_R'] = []
self.details_hist['D_V_BCE_fake_real_F'] = []
self.epoch_hist['D_loss_train'] = []
self.epoch_hist['G_loss_train'] = []
self.epoch_hist['D_loss_Validation'] = []
self.epoch_hist['G_loss_Validation'] = []
##Para poder tomar el promedio por epoch
iterIniTrain = 0
iterFinTrain = 0
iterIniValidation = 0
iterFinValidation = 0
maxIter = self.data_loader.dataset.__len__() // self.batch_size
maxIterVal = self.data_Validation.dataset.__len__() // self.batch_size
if (self.WGAN):
one = torch.tensor(1, dtype=torch.float).cuda()
mone = one * -1
else:
self.y_real_ = torch.ones(self.batch_size, 1)
self.y_fake_ = torch.zeros(self.batch_size, 1)
if self.gpu_mode:
self.y_real_, self.y_fake_ = self.y_real_.cuda(), self.y_fake_.cuda()
print('training start!!')
start_time = time.time()
for epoch in range(self.epoch):
if (epoch < self.epochVentaja):
ventaja = True
else:
ventaja = False
self.G.train()
if not self.onlyGen:
self.D.train()
epoch_start_time = time.time()
# TRAIN!!!
for iter, data in enumerate(self.data_loader):
x_im = data.get('x_im')
x_dep = data.get('x_dep')
y_im = data.get('y_im')
y_dep = data.get('y_dep')
y_ = data.get('y_')
# x_im = imagenes normales
# x_dep = profundidad de images
# y_im = imagen con el angulo cambiado
# y_ = angulo de la imagen = tengo que tratar negativos
# Aumento mi data
if (self.CR):
x_im_aug, y_im_aug = augmentData(x_im, y_im)
x_im_vanilla = x_im
if self.gpu_mode:
x_im_aug, y_im_aug = x_im_aug.cuda(), y_im_aug.cuda()
if iter >= maxIter:
break
if self.gpu_mode:
x_im, y_, y_im, x_dep, y_dep = x_im.cuda(), y_.cuda(), y_im.cuda(), x_dep.cuda(), y_dep.cuda()
# update D network
if not ventaja and not self.onlyGen:
for p in self.D.parameters(): # reset requires_grad
p.requires_grad = True # they are set to False below in netG update
self.D_optimizer.zero_grad()
# Real Images
D_real, D_features_real = self.D(y_im, x_im, y_dep, y_) ## Es la funcion forward `` g(z) x
# Fake Images
G_, G_dep = self.G( y_, x_im, x_dep)
D_fake, D_features_fake = self.D(G_, x_im, G_dep, y_)
# Losses
# GAN Loss
if (self.WGAN): # de WGAN
D_loss_real_fake_R = - torch.mean(D_real)
D_loss_real_fake_F = torch.mean(D_fake)
#D_loss_real_fake_R = - D_loss_real_fake_R_positive
else: # de Gan normal
D_loss_real_fake_R = self.BCEWithLogitsLoss(D_real, self.y_real_)
D_loss_real_fake_F = self.BCEWithLogitsLoss(D_fake, self.y_fake_)
D_loss = D_loss_real_fake_F + D_loss_real_fake_R
if self.CR:
# Fake Augmented Images bCR
x_im_aug_bCR, G_aug_bCR = augmentData(x_im_vanilla, G_.data.cpu())
if self.gpu_mode:
G_aug_bCR, x_im_aug_bCR = G_aug_bCR.cuda(), x_im_aug_bCR.cuda()
D_fake_bCR, D_features_fake_bCR = self.D(G_aug_bCR, x_im_aug_bCR, G_dep, y_)
D_real_bCR, D_features_real_bCR = self.D(y_im_aug, x_im_aug, y_dep, y_)
# Fake Augmented Images zCR
G_aug_zCR, G_dep_aug_zCR = self.G(y_, x_im_aug, x_dep)
D_fake_aug_zCR, D_features_fake_aug_zCR = self.D(G_aug_zCR, x_im_aug, G_dep_aug_zCR, y_)
# bCR Loss (*)
D_loss_real = self.MSE(D_features_real, D_features_real_bCR)
D_loss_fake = self.MSE(D_features_fake, D_features_fake_bCR)
D_bCR = (D_loss_real + D_loss_fake) * self.bFactor
# zCR Loss
D_zCR = self.MSE(D_features_fake, D_features_fake_aug_zCR) * self.zDisFactor
D_CR_losses = D_bCR + D_zCR
#D_CR_losses.backward(retain_graph=True)
D_loss += D_CR_losses
self.details_hist['D_bCR'].append(D_bCR.detach().item())
self.details_hist['D_zCR'].append(D_zCR.detach().item())
else:
self.details_hist['D_bCR'].append(0)
self.details_hist['D_zCR'].append(0)
self.train_hist['D_loss_train'].append(D_loss.detach().item())
self.details_hist['D_T_BCE_fake_real_R'].append(D_loss_real_fake_R.detach().item())
self.details_hist['D_T_BCE_fake_real_F'].append(D_loss_real_fake_F.detach().item())
if self.visdom:
visLossDTest.plot('Discriminator_losses',
['D_T_BCE_fake_real_R','D_T_BCE_fake_real_F', 'D_bCR', 'D_zCR'], 'train',
self.details_hist)
#if self.WGAN:
# D_loss_real_fake_F.backward(retain_graph=True)
# D_loss_real_fake_R_positive.backward(mone)
#else:
# D_loss_real_fake.backward()
D_loss.backward()
self.D_optimizer.step()
#WGAN
if (self.WGAN):
for p in self.D.parameters():
p.data.clamp_(-self.clipping, self.clipping) #Segun paper si el valor es muy chico lleva al banishing gradient
# Si se aplicaria la mejora en las WGANs tendiramos que sacar los batch normalizations de la red
# update G network
self.G_optimizer.zero_grad()
G_, G_dep = self.G(y_, x_im, x_dep)
if not ventaja and not self.onlyGen:
for p in self.D.parameters():
p.requires_grad = False # to avoid computation
# Fake images
D_fake, _ = self.D(G_, x_im, G_dep, y_)
if (self.WGAN):
G_loss_fake = -torch.mean(D_fake) #de WGAN
else:
G_loss_fake = self.BCEWithLogitsLoss(D_fake, self.y_real_)
# loss between images (*)
#G_join = torch.cat((G_, G_dep), 1)
#y_join = torch.cat((y_im, y_dep), 1)
G_loss_Comp = self.L1(G_, y_im)
if self.depth:
G_loss_Comp += self.L1(G_dep, y_dep)
G_loss_Dif_Comp = G_loss_Comp * self.lambdaL1
reverse_y = - y_ + 1
reverse_G, reverse_G_dep = self.G(reverse_y, G_, G_dep)
G_loss_Cycle = self.L1(reverse_G, x_im)
if self.depth:
G_loss_Cycle += self.L1(reverse_G_dep, x_dep)
G_loss_Cycle = G_loss_Cycle * self.lambdaL1/2
if (self.CR):
# Fake images augmented
G_aug, G_dep_aug = self.G(y_, x_im_aug, x_dep)
D_fake_aug, _ = self.D(G_aug, x_im, G_dep_aug, y_)
if (self.WGAN):
G_loss_fake = - (torch.mean(D_fake)+torch.mean(D_fake_aug))/2
else:
G_loss_fake = ( self.BCEWithLogitsLoss(D_fake, self.y_real_) +
self.BCEWithLogitsLoss(D_fake_aug,self.y_real_)) / 2
# loss between images (*)
#y_aug_join = torch.cat((y_im_aug, y_dep), 1)
#G_aug_join = torch.cat((G_aug, G_dep_aug), 1)
G_loss_Comp_Aug = self.L1(G_aug, y_im_aug)
if self.depth:
G_loss_Comp_Aug += self.L1(G_dep_aug, y_dep)
G_loss_Dif_Comp = (G_loss_Comp + G_loss_Comp_Aug)/2 * self.lambdaL1
G_loss = G_loss_fake + G_loss_Dif_Comp + G_loss_Cycle
self.details_hist['G_T_BCE_fake_real'].append(G_loss_fake.detach().item())
self.details_hist['G_T_Comp_im'].append(G_loss_Dif_Comp.detach().item())
self.details_hist['G_T_Cycle'].append(G_loss_Cycle.detach().item())
self.details_hist['G_zCR'].append(0)
else:
G_loss = self.L1(G_, y_im)
if self.depth:
G_loss += self.L1(G_dep, y_dep)
G_loss = G_loss * self.lambdaL1
self.details_hist['G_T_Comp_im'].append(G_loss.detach().item())
self.details_hist['G_T_BCE_fake_real'].append(0)
self.details_hist['G_T_Cycle'].append(0)
self.details_hist['G_zCR'].append(0)
G_loss.backward()
self.G_optimizer.step()
self.train_hist['G_loss_train'].append(G_loss.detach().item())
if self.onlyGen:
self.train_hist['D_loss_train'].append(0)
iterFinTrain += 1
if self.visdom:
visLossGTest.plot('Generator_losses',
['G_T_Comp_im', 'G_T_BCE_fake_real', 'G_zCR','G_T_Cycle'],
'train', self.details_hist)
vis.plot('loss', ['D_loss_train', 'G_loss_train'], 'train', self.train_hist)
##################Validation####################################
with torch.no_grad():
self.G.eval()
if not self.onlyGen:
self.D.eval()
for iter, data in enumerate(self.data_Validation):
# Aumento mi data
x_im = data.get('x_im')
x_dep = data.get('x_dep')
y_im = data.get('y_im')
y_dep = data.get('y_dep')
y_ = data.get('y_')
# x_im = imagenes normales
# x_dep = profundidad de images
# y_im = imagen con el angulo cambiado
# y_ = angulo de la imagen = tengo que tratar negativos
# x_im = torch.Tensor(list(x_im))
# x_dep = torch.Tensor(x_dep)
# y_im = torch.Tensor(y_im)
# print(y_.shape[0])
if iter == maxIterVal:
# print ("Break")
break
# print (y_.type(torch.LongTensor).unsqueeze(1))
# print("y_vec_", y_vec_)
# print ("z_", z_)
if self.gpu_mode:
x_im, y_, y_im, x_dep, y_dep = x_im.cuda(), y_.cuda(), y_im.cuda(), x_dep.cuda(), y_dep.cuda()
# D network
if not ventaja and not self.onlyGen:
# Real Images
D_real, _ = self.D(y_im, x_im, y_dep,y_) ## Es la funcion forward `` g(z) x
# Fake Images
G_, G_dep = self.G(y_, x_im, x_dep)
D_fake, _ = self.D(G_, x_im, G_dep, y_)
# Losses
# GAN Loss
if (self.WGAN): # de WGAN
D_loss_real_fake_R = - torch.mean(D_real)
D_loss_real_fake_F = torch.mean(D_fake)
else: # de Gan normal
D_loss_real_fake_R = self.BCEWithLogitsLoss(D_real, self.y_real_)
D_loss_real_fake_F = self.BCEWithLogitsLoss(D_fake, self.y_fake_)
D_loss_real_fake = D_loss_real_fake_F + D_loss_real_fake_R
D_loss = D_loss_real_fake
self.train_hist['D_loss_Validation'].append(D_loss.item())
self.details_hist['D_V_BCE_fake_real_R'].append(D_loss_real_fake_R.item())
self.details_hist['D_V_BCE_fake_real_F'].append(D_loss_real_fake_F.item())
if self.visdom:
visLossDValidation.plot('Discriminator_losses',
['D_V_BCE_fake_real_R','D_V_BCE_fake_real_F'], 'Validation',
self.details_hist)
# G network
G_, G_dep = self.G(y_, x_im, x_dep)
if not ventaja and not self.onlyGen:
# Fake images
D_fake,_ = self.D(G_, x_im, G_dep, y_)
#Loss GAN
if (self.WGAN):
G_loss = -torch.mean(D_fake) # porWGAN
else:
G_loss = self.BCEWithLogitsLoss(D_fake, self.y_real_) #de GAN NORMAL
self.details_hist['G_V_BCE_fake_real'].append(G_loss.item())
#Loss comparation
#G_join = torch.cat((G_, G_dep), 1)
#y_join = torch.cat((y_im, y_dep), 1)
G_loss_Comp = self.L1(G_, y_im)
if self.depth:
G_loss_Comp += self.L1(G_dep, y_dep)
G_loss_Comp = G_loss_Comp * self.lambdaL1
reverse_y = - y_ + 1
reverse_G, reverse_G_dep = self.G(reverse_y, G_, G_dep)
G_loss_Cycle = self.L1(reverse_G, x_im)
if self.depth:
G_loss_Cycle += self.L1(reverse_G_dep, x_dep)
G_loss_Cycle = G_loss_Cycle * self.lambdaL1/2
G_loss += G_loss_Comp + G_loss_Cycle
self.details_hist['G_V_Comp_im'].append(G_loss_Comp.item())
self.details_hist['G_V_Cycle'].append(G_loss_Cycle.detach().item())
else:
G_loss = self.L1(G_, y_im)
if self.depth:
G_loss += self.L1(G_dep, y_dep)
G_loss = G_loss * self.lambdaL1
self.details_hist['G_V_Comp_im'].append(G_loss.item())
self.details_hist['G_V_BCE_fake_real'].append(0)
self.details_hist['G_V_Cycle'].append(0)
self.train_hist['G_loss_Validation'].append(G_loss.item())
if self.onlyGen:
self.train_hist['D_loss_Validation'].append(0)
iterFinValidation += 1
if self.visdom:
visLossGValidation.plot('Generator_losses', ['G_V_Comp_im', 'G_V_BCE_fake_real','G_V_Cycle'],
'Validation', self.details_hist)
visValidation.plot('loss', ['D_loss_Validation', 'G_loss_Validation'], 'Validation',
self.train_hist)
##Vis por epoch
if ventaja or self.onlyGen:
self.epoch_hist['D_loss_train'].append(0)
self.epoch_hist['D_loss_Validation'].append(0)
else:
#inicioTr = (epoch - self.epochVentaja) * (iterFinTrain - iterIniTrain)
#inicioTe = (epoch - self.epochVentaja) * (iterFinValidation - iterIniValidation)
self.epoch_hist['D_loss_train'].append(mean(self.train_hist['D_loss_train'][iterIniTrain: -1]))
self.epoch_hist['D_loss_Validation'].append(mean(self.train_hist['D_loss_Validation'][iterIniValidation: -1]))
self.epoch_hist['G_loss_train'].append(mean(self.train_hist['G_loss_train'][iterIniTrain:iterFinTrain]))
self.epoch_hist['G_loss_Validation'].append(
mean(self.train_hist['G_loss_Validation'][iterIniValidation:iterFinValidation]))
if self.visdom:
visEpoch.plot('epoch', epoch,
['D_loss_train', 'G_loss_train', 'D_loss_Validation', 'G_loss_Validation'],
self.epoch_hist)
self.train_hist['D_loss_train'] = self.train_hist['D_loss_train'][-1:]
self.train_hist['G_loss_train'] = self.train_hist['G_loss_train'][-1:]
self.train_hist['D_loss_Validation'] = self.train_hist['D_loss_Validation'][-1:]
self.train_hist['G_loss_Validation'] = self.train_hist['G_loss_Validation'][-1:]
self.train_hist['per_epoch_time'] = self.train_hist['per_epoch_time'][-1:]
self.train_hist['total_time'] = self.train_hist['total_time'][-1:]
self.details_hist['G_T_Comp_im'] = self.details_hist['G_T_Comp_im'][-1:]
self.details_hist['G_T_BCE_fake_real'] = self.details_hist['G_T_BCE_fake_real'][-1:]
self.details_hist['G_T_Cycle'] = self.details_hist['G_T_Cycle'][-1:]
self.details_hist['G_zCR'] = self.details_hist['G_zCR'][-1:]
self.details_hist['G_V_Comp_im'] = self.details_hist['G_V_Comp_im'][-1:]
self.details_hist['G_V_BCE_fake_real'] = self.details_hist['G_V_BCE_fake_real'][-1:]
self.details_hist['G_V_Cycle'] = self.details_hist['G_V_Cycle'][-1:]
self.details_hist['D_T_BCE_fake_real_R'] = self.details_hist['D_T_BCE_fake_real_R'][-1:]
self.details_hist['D_T_BCE_fake_real_F'] = self.details_hist['D_T_BCE_fake_real_F'][-1:]
self.details_hist['D_zCR'] = self.details_hist['D_zCR'][-1:]
self.details_hist['D_bCR'] = self.details_hist['D_bCR'][-1:]
self.details_hist['D_V_BCE_fake_real_R'] = self.details_hist['D_V_BCE_fake_real_R'][-1:]
self.details_hist['D_V_BCE_fake_real_F'] = self.details_hist['D_V_BCE_fake_real_F'][-1:]
##Para poder tomar el promedio por epoch
iterIniTrain = 1
iterFinTrain = 1
iterIniValidation = 1
iterFinValidation = 1
self.train_hist['per_epoch_time'].append(time.time() - epoch_start_time)
if epoch % 10 == 0:
self.save(str(epoch))
with torch.no_grad():
if self.visdom:
self.visualize_results(epoch, dataprint=self.dataprint, visual=visImages)
self.visualize_results(epoch, dataprint=self.dataprint_test, visual=visImagesTest)
else:
imageName = self.model_name + '_' + 'Train' + '_' + str(self.seed) + '_' + str(epoch)
self.visualize_results(epoch, dataprint=self.dataprint, name= imageName)
self.visualize_results(epoch, dataprint=self.dataprint_test, name= imageName)
self.train_hist['total_time'].append(time.time() - start_time)
print("Avg one epoch time: %.2f, total %d epochs time: %.2f" % (np.mean(self.train_hist['per_epoch_time']),
self.epoch, self.train_hist['total_time'][0]))
print("Training finish!... save training results")
self.save()
#utils.generate_animation(self.result_dir + '/' + self.dataset + '/' + self.model_name + '/' + self.model_name,
# self.epoch)
#utils.loss_plot(self.train_hist, os.path.join(self.save_dir, self.dataset, self.model_name), self.model_name)
def visualize_results(self, epoch, dataprint, visual="", name= "test"):
with torch.no_grad():
self.G.eval()
#if not os.path.exists(self.result_dir + '/' + self.dataset + '/' + self.model_name):
# os.makedirs(self.result_dir + '/' + self.dataset + '/' + self.model_name)
# print("sample z: ",self.sample_z_,"sample y:", self.sample_y_)
##Podria hacer un loop
# .zfill(4)
#newSample = None
#print(dataprint.shape)
#newSample = torch.tensor([])
#se que es ineficiente pero lo hago cada 10 epoch nomas
newSample = []
iter = 1
for x_im,x_dep in zip(dataprint.get('x_im'), dataprint.get('x_dep')):
if (iter > self.cantImages):
break
#x_im = (x_im + 1) / 2
#imgX = transforms.ToPILImage()(x_im)
#imgX.show()
x_im_input = x_im.repeat(2, 1, 1, 1)
x_dep_input = x_dep.repeat(2, 1, 1, 1)
sizeImage = x_im.shape[2]
sample_y_ = torch.zeros((self.class_num, 1, sizeImage, sizeImage))
for i in range(self.class_num):
if(int(i % self.class_num) == 1):
sample_y_[i] = torch.ones(( 1, sizeImage, sizeImage))
if self.gpu_mode:
sample_y_, x_im_input, x_dep_input = sample_y_.cuda(), x_im_input.cuda(), x_dep_input.cuda()
G_im, G_dep = self.G(sample_y_, x_im_input, x_dep_input)
newSample.append(x_im.squeeze(0))
newSample.append(x_dep.squeeze(0).expand(3, -1, -1))
if self.wiggle:
im_aux, im_dep_aux = G_im, G_dep
for i in range(0, 2):
index = i
for j in range(0, self.wiggleDepth):
# print(i,j)
if (j == 0 and i == 1):
# para tomar el original
im_aux, im_dep_aux = G_im, G_dep
newSample.append(G_im.cpu()[0].squeeze(0))
newSample.append(G_im.cpu()[1].squeeze(0))
elif (i == 1):
# por el problema de las iteraciones proximas
index = 0
# imagen generada
x = im_aux[index].unsqueeze(0)
x_dep = im_dep_aux[index].unsqueeze(0)
y = sample_y_[i].unsqueeze(0)
if self.gpu_mode:
y, x, x_dep = y.cuda(), x.cuda(), x_dep.cuda()
im_aux, im_dep_aux = self.G(y, x, x_dep)
newSample.append(im_aux.cpu()[0])
else:
newSample.append(G_im.cpu()[0])
newSample.append(G_im.cpu()[1])
newSample.append(G_dep.cpu()[0].expand(3, -1, -1))
newSample.append(G_dep.cpu()[1].expand(3, -1, -1))
# sadadas
iter+=1
if self.visdom:
visual.plot(epoch, newSample, int(len(newSample) /self.cantImages))
else:
utils.save_wiggle(newSample, self.cantImages, name)
##TENGO QUE HACER QUE SAMPLES TENGAN COMO MAXIMO self.class_num * self.class_num
# utils.save_images(newSample[:, :, :, :], [image_frame_dim * cantidadIm , image_frame_dim * (self.class_num+2)],
# self.result_dir + '/' + self.dataset + '/' + self.model_name + '/' + self.model_name + '_epoch%04d' % epoch + '.png')
def show_plot_images(self, images, cols=1, titles=None):
"""Display a list of images in a single figure with matplotlib.
Parameters
---------
images: List of np.arrays compatible with plt.imshow.
cols (Default = 1): Number of columns in figure (number of rows is
set to np.ceil(n_images/float(cols))).
titles: List of titles corresponding to each image. Must have
the same length as titles.
"""
# assert ((titles is None) or (len(images) == len(titles)))
n_images = len(images)
if titles is None: titles = ['Image (%d)' % i for i in range(1, n_images + 1)]
fig = plt.figure()
for n, (image, title) in enumerate(zip(images, titles)):
a = fig.add_subplot(np.ceil(n_images / float(cols)), cols, n + 1)
# print(image)
image = (image + 1) * 255.0
# print(image)
# new_im = Image.fromarray(image)
# print(new_im)
if image.ndim == 2:
plt.gray()
# print("spi imshape ", image.shape)
plt.imshow(image)
a.set_title(title)
fig.set_size_inches(np.array(fig.get_size_inches()) * n_images)
plt.show()
def joinImages(self, data):
nData = []
for i in range(self.class_num):
nData.append(data)
nData = np.array(nData)
nData = torch.tensor(nData.tolist())
nData = nData.type(torch.FloatTensor)
return nData
def save(self, epoch=''):
save_dir = os.path.join(self.save_dir, self.dataset, self.model_name)
if not os.path.exists(save_dir):
os.makedirs(save_dir)
torch.save(self.G.state_dict(),
os.path.join(save_dir, self.model_name + '_' + self.seed + '_' + epoch + '_G.pkl'))
if not self.onlyGen:
torch.save(self.D.state_dict(),
os.path.join(save_dir, self.model_name + '_' + self.seed + '_' + epoch + '_D.pkl'))
with open(os.path.join(save_dir, self.model_name + '_history_ '+self.seed+'.pkl'), 'wb') as f:
pickle.dump(self.train_hist, f)
def load(self):
save_dir = os.path.join(self.save_dir, self.dataset, self.model_name)
self.G.load_state_dict(torch.load(os.path.join(save_dir, self.model_name + '_' + self.seed_load + '_G.pkl')))
if not self.wiggle:
self.D.load_state_dict(torch.load(os.path.join(save_dir, self.model_name + '_' + self.seed_load + '_D.pkl')))
def wiggleEf(self):
seed, epoch = self.seed_load.split('_')
if self.visdom:
visWiggle = utils.VisdomImagePlotter(env_name='Cobo_depth_wiggle_' + seed)
self.visualize_results(epoch=epoch, dataprint=self.dataprint_test, visual=visWiggle)
else:
self.visualize_results(epoch=epoch, dataprint=self.dataprint_test, visual=None, name = self.name_wiggle)
def recreate(self):
dataloader_recreate = dataloader(self.dataset, self.input_size, self.batch_size, self.imageDim, split='score')
with torch.no_grad():
self.G.eval()
accum = 0
for data_batch in dataloader_recreate.__iter__():
#{'x_im': x1, 'x_dep': x1_dep, 'y_im': x2, 'y_dep': x2_dep, 'y_': torch.ones(1, self.imageDim, self.imageDim)}
left,left_depth,right,right_depth,direction = data_batch.values()
if self.gpu_mode:
left,left_depth,right,right_depth,direction = left.cuda(),left_depth.cuda(),right.cuda(),right_depth.cuda(),direction.cuda()
G_right, G_right_dep = self.G( direction, left, left_depth)
reverse_direction = direction * 0
G_left, G_left_dep = self.G( reverse_direction, right, right_depth)
for index in range(0,self.batch_size):
image_right = (G_right[index] + 1.0)/2.0
image_right_dep = (G_right_dep[index] + 1.0)/2.0
image_left = (G_left[index] + 1.0)/2.0
image_left_dep = (G_left_dep[index] + 1.0)/2.0
save_image(image_right, os.path.join("results","recreate_dataset","CAM1","n_{num:0{width}}.png".format(num = index+accum, width = 4)))
save_image(image_right_dep, os.path.join("results","recreate_dataset","CAM1","d_{num:0{width}}.png".format(num = index+accum, width = 4)))
save_image(image_left, os.path.join("results","recreate_dataset","CAM0","n_{num:0{width}}.png".format(num = index+accum, width = 4)))
save_image(image_left_dep, os.path.join("results","recreate_dataset","CAM0","d_{num:0{width}}.png".format(num = index+accum, width = 4)))
accum+= self.batch_size