def setup(opts):
model = Transformer()
model.load_state_dict(torch.load(opts["checkpoint"]))
model.eval()
if torch.cuda.is_available():
print("GPU Mode")
model.cuda()
else:
print("CPU Mode")
model.float()
return model
def init_model():
model = Transformer()
model.load_state_dict(
torch.load(
os.path.join('./pretrained_model', 'Hayao' + '_net_G_float.pth')))
model.eval()
model.cuda(0)
return model
def build_network(self):
print('=' * 50)
print('Building network...')
#self.T = Transformer(4, cfg.img_channels, nb_features=64)
self.T = Transformer()
self.D = Discriminator(input_features=cfg.img_channels)
if cfg.cuda_use:
self.T.cuda()
self.D.cuda()
self.opt_T = torch.optim.Adam(self.T.parameters(), lr=cfg.t_lr)
self.opt_D = torch.optim.SGD(self.D.parameters(), lr=cfg.d_lr)
self.self_regularization_loss = nn.L1Loss(size_average=False)
self.local_adversarial_loss = nn.CrossEntropyLoss(size_average=True)
self.delta = cfg.delta
self.vgg = Vgg16(requires_grad=False)
network.netutils.init_vgg16("./models/")
self.vgg.load_state_dict(torch.load(os.path.join("./models/", "vgg16.weight")))
self.vgg.cuda()
def load_models(s3, bucket):
styles = ["Hosoda", "Hayao", "Shinkai", "Paprika"]
models = {}
for style in styles:
model = Transformer()
response = s3.get_object(Bucket=bucket, Key=f"models/{style}_net_G_float.pth")
state = torch.load(BytesIO(response["Body"].read()))
model.load_state_dict(state)
model.eval()
models[style] = model
return models
import torch
import torchvision.transforms as transforms
import cv2
import os
import matplotlib.pyplot as plt
from network.Transformer import Transformer
model = Transformer()
model.load_state_dict(torch.load('pretrained_model/Hayao_net_G_float.pth'))
model.eval()
print('Model loaded!')
img_size = 700
img_path = 'test_img/9.jpg'
img = cv2.imread(img_path)
T = transforms.Compose([
transforms.ToPILImage(),
transforms.Resize(img_size, 2),
transforms.ToTensor()
])
img_input = T(img).unsqueeze(0)
img_input = -1 + 2 * img_input
img_output = model(img_input)
img_output = (img_output.squeeze().detach().numpy() + 1.) / 2.
img_output = img_output.transpose([1, 2, 0])
parser = argparse.ArgumentParser()
parser.add_argument('--input_dir', type=str, default = 'test_img')
parser.add_argument('--load_size', type=int, default = 450)
parser.add_argument('--model_path', type=str, default = './pretrained_model')
parser.add_argument('--style', type=str, default = 'Hayao')
parser.add_argument('--output_dir', type=str, default = 'test_output')
parser.add_argument('--gpu', type=int, default = 0)
opt = parser.parse_args()
valid_ext = ['.jpg', '.png']
if not os.path.exists(opt.output_dir): os.mkdir(opt.output_dir)
# load pretrained model
model = Transformer()
model.load_state_dict(torch.load(os.path.join(opt.model_path, opt.style + '_net_G_float.pth')))
model.eval()
if opt.gpu > -1:
print('GPU mode')
model.cuda()
else:
print('CPU mode')
model.float()
for files in os.listdir(opt.input_dir):
torch.cuda.empty_cache()
gc.collect()
ext = os.path.splitext(files)[1]
if ext not in valid_ext:
class Main(object):
def __init__(self):
# network
self.T = None
self.D = None
self.opt_T = None
self.opt_D = None
self.self_regularization_loss = None
self.local_adversarial_loss = None
self.delta = None
# data
self.anime_train_loader = None
self.animeblur_train_loader = None
self.real_loader = None
self.anime_train_loader = None
self.animeblur_train_loader = None
def build_network(self):
print('=' * 50)
print('Building network...')
#self.T = Transformer(4, cfg.img_channels, nb_features=64)
self.T = Transformer()
self.D = Discriminator(input_features=cfg.img_channels)
if cfg.cuda_use:
self.T.cuda()
self.D.cuda()
self.opt_T = torch.optim.Adam(self.T.parameters(), lr=cfg.t_lr)
self.opt_D = torch.optim.SGD(self.D.parameters(), lr=cfg.d_lr)
self.self_regularization_loss = nn.L1Loss(size_average=False)
self.local_adversarial_loss = nn.CrossEntropyLoss(size_average=True)
self.delta = cfg.delta
self.vgg = Vgg16(requires_grad=False)
network.netutils.init_vgg16("./models/")
self.vgg.load_state_dict(torch.load(os.path.join("./models/", "vgg16.weight")))
self.vgg.cuda()
def load_data(self):
print('=' * 50)
print('Loading data...')
transform = transforms.Compose([
# transforms.Grayscale,
transforms.Scale((cfg.img_width, cfg.img_height)),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])
anime_train_folder = torchvision.datasets.ImageFolder(root=cfg.anime_path, transform=transform)
animeblur_train_folder = torchvision.datasets.ImageFolder(root=cfg.animeblur_path, transform=transform)
real_train_folder = torchvision.datasets.ImageFolder(root=cfg.real_path, transform=transform)
self.anime_train_loader = Data.DataLoader(anime_train_folder, batch_size=cfg.batch_size, shuffle=True,
pin_memory=True)
self.animeblur_train_loader = Data.DataLoader(animeblur_train_folder, batch_size=cfg.batch_size, shuffle=True,
pin_memory=True)
self.real_train_loader = Data.DataLoader(real_train_folder, batch_size=cfg.batch_size, shuffle=True,
pin_memory=True)
print('anime_train_batch %d' % len(self.anime_train_loader))
print('animeblur_train_batch %d' % len(self.animeblur_train_loader))
real_folder = torchvision.datasets.ImageFolder(root=cfg.real_path, transform=transform)
# real_folder.imgs = real_folder.imgs[:2000]
self.real_loader = Data.DataLoader(real_folder, batch_size=cfg.batch_size, shuffle=True,
pin_memory=True)
print('real_batch %d' % len(self.real_loader))
def pre_train_t(self):
print('=' * 50)
#device = torch.device("cuda" if args.cuda else "cpu")
if cfg.ref_pre_path:
print('Loading t_pre from %s' % cfg.ref_pre_path)
self.T.load_state_dict(torch.load(cfg.ref_pre_path))
return
print('pre-training the refiner network %d times...' % cfg.t_pretrain)
mse_loss = torch.nn.MSELoss()
for index in range(cfg.t_pretrain):
#print("aaaaaaaa")
anime_image_batch, _ = iter(self.anime_train_loader).next()
anime_image_batch = Variable(anime_image_batch).cuda()
animeblur_image_batch, _ = iter(self.animeblur_train_loader).next()
animeblur_image_batch = Variable(animeblur_image_batch).cuda()
real_image_batch, _ = iter(self.real_train_loader).next()
real_image_batch = Variable(real_image_batch).cuda()
#print(real_image_batch.size())
self.T.train()
transreal_image_batch = self.T(real_image_batch)
#################################
real_features = self.vgg(real_image_batch).relu4_3
transreal_features = self.vgg(transreal_image_batch).relu4_3
#t_loss = self.self_regularization_loss( real_image_batch, transreal_image_batch )
#t_loss = mse_loss(real_features, transreal_features)
loss = torch.abs(transreal_features - real_features)
#t_loss = loss.sum() / (cfg.batch_size * loss.mean())
t_loss = loss.mean()
print(t_loss.size())
print(t_loss)
#################################
# t_loss = torch.div(t_loss, cfg.batch_size)
t_loss = torch.mul(t_loss, self.delta)
self.opt_T.zero_grad()
t_loss.backward()
self.opt_T.step()
# log every `log_interval` steps
if (index % cfg.t_pre_per == 0) or (index == cfg.t_pretrain - 1):
# figure_name = 'refined_image_batch_pre_train_step_{}.png'.format(index)
print('[%d/%d] (R)reg_loss: %.4f' % (index, cfg.t_pretrain, t_loss.data[0]))
anime_image_batch, _ = iter(self.anime_train_loader).next()
anime_image_batch = Variable(anime_image_batch, volatile=True).cuda()
animeblur_image_batch, _ = iter(self.animeblur_train_loader).next()
animeblur_image_batch = Variable(animeblur_image_batch, volatile=True).cuda()
real_image_batch, _ = iter(self.real_loader).next()
real_image_batch = Variable(real_image_batch, volatile=True).cuda()
self.T.eval()
ref_image_batch = self.T(real_image_batch)
figure_path = os.path.join(cfg.train_res_path, 'refined_image_batch_pre_train_%d.png' % index)
generate_img_batch(anime_image_batch.data.cpu(), ref_image_batch.data.cpu(),
real_image_batch.data.cpu(), figure_path)
self.T.train()
print('Save t_pre to models/t_pre.pkl')
torch.save(self.T.state_dict(), 'models/t_pre.pkl')
def pre_train_d(self):
print('=' * 50)
if cfg.disc_pre_path:
print('Loading D_pre from %s' % cfg.disc_pre_path)
self.D.load_state_dict(torch.load(cfg.disc_pre_path))
return
print('pre-training the discriminator network %d times...' % cfg.t_pretrain)
self.D.train()
self.T.eval()
for index in range(cfg.d_pretrain):
real_image_batch, _ = iter(self.real_loader).next()
real_image_batch = Variable(real_image_batch).cuda()
anime_image_batch, _ = iter(self.anime_train_loader).next()
anime_image_batch = Variable(anime_image_batch).cuda()
animeblur_image_batch, _ = iter(self.anime_train_loader).next()
animeblur_image_batch = Variable(animeblur_image_batch).cuda()
assert real_image_batch.size(0) == anime_image_batch.size(0)
assert real_image_batch.size(0) == anime_image_batch.size(0)
d_real_pred = self.D(real_image_batch).view(-1, 2)
d_anime_y = Variable(torch.zeros(d_real_pred.size(0)).type(torch.LongTensor)).cuda()
# real to fake
d_real_y = Variable(torch.ones(d_real_pred.size(0)).type(torch.LongTensor)).cuda()
# blur is fake
d_blur_y = Variable(torch.ones(d_real_pred.size(0)).type(torch.LongTensor)).cuda()
# ============ real image D ====================================================
# self.D.train()
d_anime_pred = self.D(anime_image_batch).view(-1, 2)
acc_anime = calc_acc(d_anime_pred, 'real')
d_loss_anime = self.local_adversarial_loss(d_anime_pred, d_anime_y)
# d_loss_real = torch.div(d_loss_real, cfg.batch_size)
# ============ anime image D ====================================================
# self.T.eval()
real_image_batch = self.T(real_image_batch)
# self.D.train()
d_real_pred = self.D(real_image_batch).view(-1, 2)
acc_real = calc_acc(d_real_pred, 'refine')
d_loss_real = self.local_adversarial_loss(d_real_pred, d_real_y)
# d_loss_ref = torch.div(d_loss_ref, cfg.batch_size)
# =========== blue image D =============
d_animeblur_pred = self.D(animeblur_image_batch).view(-1, 2)
acc_blur = calc_acc(d_animeblur_pred, 'refine')
d_loss_animeblur = self.local_adversarial_loss(d_animeblur_pred, d_blur_y)
d_loss = d_loss_anime + d_loss_animeblur + d_loss_real
self.opt_D.zero_grad()
d_loss.backward()
self.opt_D.step()
if (index % cfg.d_pre_per == 0) or (index == cfg.d_pretrain - 1):
print('[%d/%d] (D)d_loss:%f acc_anime:%.2f%% acc_real:%.2f%% acc_blur:%.2f%%'
% (index, cfg.d_pretrain, d_loss.data[0], acc_anime, acc_real, acc_blur))
print('Save D_pre to models/D_pre.pkl')
torch.save(self.D.state_dict(), 'models/D_pre.pkl')
def train(self):
print('=' * 50)
print('Training...')
#self.D.load_state_dict(torch.load("models/D_620.pkl"))
#self.T.load_state_dict(torch.load("models/T_620.pkl"))
'''
image_history_buffer = ImageHistoryBuffer((0, cfg.img_channels, cfg.img_height, cfg.img_width),
cfg.buffet_size * 10, cfg.batch_size)
'''
for step in range(cfg.train_steps):
print('Step[%d/%d]' % (step, cfg.train_steps))
# ========= train the T =========
self.D.eval()
self.T.train()
for p in self.D.parameters():
p.requires_grad = False
total_t_loss = 0.0
total_t_loss_reg_scale = 0.0
total_t_loss_adv = 0.0
total_acc_adv = 0.0
for index in range(cfg.k_t):
real_image_batch, _ = iter(self.real_loader).next()
real_image_batch = Variable(real_image_batch).cuda()
#real_image_batch, _ = iter(self.real_loader).next()
#real_image_batch = Variable(real_image_batch).cuda()
d_real_pred = self.D(real_image_batch).view(-1, 2)
d_real_y = Variable(torch.zeros(d_real_pred.size(0)).type(torch.LongTensor)).cuda()
transreal_image_batch = self.T(real_image_batch)
d_transreal_pred = self.D(transreal_image_batch).view(-1, 2)
acc_adv = calc_acc(d_transreal_pred, 'real')
t_loss_adv = self.local_adversarial_loss(d_transreal_pred, d_real_y)
#--------================================================================
real_features = self.vgg(real_image_batch).relu4_3
transreal_features = self.vgg(transreal_image_batch).relu4_3
#t_loss = self.self_regularization_loss( real_image_batch, transreal_image_batch )
#t_loss = mse_loss(real_features, transreal_features)
loss = torch.abs(transreal_features - real_features)
#t_loss_reg = loss.sum() / cfg.batch_size
t_loss_reg = loss.mean()
#--------================================================================
#t_loss_reg = self.self_regularization_loss( self.vgg(real_image_batch).relu4_3 , self.vgg(transreal_image_batch).relu4_3 )
t_loss_reg_scale = torch.mul(t_loss_reg, self.delta)
t_loss = t_loss_adv + t_loss_reg_scale
self.opt_T.zero_grad()
self.opt_D.zero_grad()
t_loss.backward()
self.opt_T.step()
total_t_loss += t_loss
total_t_loss_reg_scale += t_loss_reg_scale
total_t_loss_adv += t_loss_adv
total_acc_adv += acc_adv
mean_t_loss = total_t_loss / cfg.k_t
mean_t_loss_reg_scale = total_t_loss_reg_scale / cfg.k_t
mean_t_loss_adv = total_t_loss_adv / cfg.k_t
mean_acc_adv = total_acc_adv / cfg.k_t
print('(R)t_loss:%.4f t_loss_reg:%.4f, t_loss_adv:%f(%.2f%%)'
% (mean_t_loss.data[0], mean_t_loss_reg_scale.data[0], mean_t_loss_adv.data[0], mean_acc_adv))
# ========= train the D =========
self.T.eval()
self.D.train()
for p in self.D.parameters():
p.requires_grad = True
for index in range(cfg.k_d):
real_image_batch, _ = iter(self.real_loader).next()
anime_image_batch, _ = iter(self.anime_train_loader).next()
animeblur_image_batch, _ = iter(self.animeblur_train_loader).next()
assert real_image_batch.size(0) == anime_image_batch.size(0)
real_image_batch = Variable(real_image_batch).cuda()
anime_image_batch = Variable(anime_image_batch).cuda()
animeblur_image_batch = Variable(anime_image_batch).cuda()
d_anime_y = Variable(torch.zeros(d_real_pred.size(0)).type(torch.LongTensor)).cuda()
d_real_y = Variable(torch.ones(d_real_pred.size(0)).type(torch.LongTensor)).cuda()
d_blur_y = Variable(torch.ones(d_real_pred.size(0)).type(torch.LongTensor)).cuda()
d_anime_pred = self.D(anime_image_batch).view(-1, 2)
acc_anime = calc_acc(d_anime_pred, 'real')
d_loss_anime = self.local_adversarial_loss(d_anime_pred, d_anime_y)
real_image_batch = self.T(real_image_batch)
d_real_pred = self.D(real_image_batch).view(-1, 2)
acc_real = calc_acc(d_real_pred, 'refine')
d_loss_real = self.local_adversarial_loss(d_real_pred, d_real_y)
d_animeblur_pred = self.D(animeblur_image_batch).view(-1, 2)
acc_blur = calc_acc(d_animeblur_pred, 'refine')
d_loss_animeblur = self.local_adversarial_loss(d_animeblur_pred, d_blur_y)
d_loss = d_loss_real + d_loss_anime + d_loss_animeblur
self.D.zero_grad()
d_loss.backward()
self.opt_D.step()
print('(D)d_loss:%.4f anime_loss:%.4f(%.2f%%) real_loss:%.4f(%.2f%%) blur_loss:%.4f(%.2f%%)'
% (d_loss.data[0] / 2, d_loss_anime.data[0], acc_anime, d_loss_real.data[0], acc_real, d_loss_animeblur.data[0], acc_blur))
if step % cfg.save_per == 0:
print('Save two model dict.')
torch.save(self.D.state_dict(), cfg.D_path % step)
torch.save(self.T.state_dict(), cfg.T_path % step)
real_image_batch, _ = iter(self.real_loader).next()
real_image_batch = Variable(real_image_batch, volatile=True).cuda()
anime_image_batch, _ = iter(self.anime_train_loader).next()
anime_image_batch = Variable(anime_image_batch, volatile=True).cuda()
animeblur_image_batch, _ = iter(self.animeblur_train_loader).next()
animeblur_image_batch = Variable(animeblur_image_batch, volatile=True).cuda()
self.T.eval()
realtrans_image_batch = self.T(real_image_batch)
self.generate_batch_train_image(real_image_batch, realtrans_image_batch, animeblur_image_batch, step_index=step)
def generate_batch_train_image(self, anime_image_batch, ref_image_batch, real_image_batch, step_index=-1):
print('=' * 50)
print('Generating a batch of training images...')
self.T.eval()
pic_path = os.path.join(cfg.train_res_path, 'step_%d.png' % step_index)
generate_img_batch(anime_image_batch.cpu().data, ref_image_batch.cpu().data, real_image_batch.cpu().data, pic_path)
print('=' * 50)
def imageConverter(input_dir='input_img',
load_size=1080,
model_path='./pretrained_model',
style='Hayao',
output_dir='Output_img',
input_file='4--24.jpg'):
gpu = -1
file_name = input_file
ext = os.path.splitext(file_name)
if not os.path.exists(output_dir): os.mkdir(output_dir)
# load pretrained model
model = Transformer()
model.load_state_dict(
torch.load(os.path.join(model_path, style + '_net_G_float.pth')))
model.eval()
#check if gpu available
if gpu > -1:
print('GPU mode')
model.cuda()
else:
# print('CPU mode')
model.float()
# load image
input_image = Image.open(os.path.join(input_dir, file_name)).convert("RGB")
# resize image, keep aspect ratio
h = input_image.size[0]
w = input_image.size[1]
# TODO should change this usage and make it more elegant
ratio = h * 1.0 / w
if w > 1080 or h > 1080:
load_size = 1080
if load_size != -1:
if ratio > 1:
h = int(load_size)
w = int(h * 1.0 / ratio)
else:
w = int(load_size)
h = int(w * ratio)
input_image = input_image.resize((h, w), Image.BICUBIC)
input_image = np.asarray(input_image)
# RGB -> BGR
input_image = input_image[:, :, [2, 1, 0]]
input_image = transforms.ToTensor()(input_image).unsqueeze(0)
# preprocess, (-1, 1)
input_image = -1 + 2 * input_image
if gpu > -1:
input_image = Variable(input_image, requires_grad=False).cuda()
else:
input_image = Variable(input_image, requires_grad=False).float()
# forward
output_image = model(input_image)
output_image = output_image[0]
# BGR -> RGB
output_image = output_image[[2, 1, 0], :, :]
print(output_image.shape)
# deprocess, (0, 1)
output_image = output_image.data.cpu().float() * 0.5 + 0.5
# save
final_name = file_name[:-4] + '_' + style + '.jpg'
output_path = os.path.join(output_dir, final_name)
vutils.save_image(output_image, output_path)
return final_name
def main():
if not os.path.exists(opt.output_dir):
os.mkdir(opt.output_dir)
# load pretrained model
model = Transformer()
model.load_state_dict(
torch.load("{dir}/{name}".format(
**{
"dir": opt.model_path,
"name": "{}_net_G_float.pth".format(opt.style)
})))
model.eval()
if opt.gpu > -1:
print("GPU mode")
model.cuda()
else:
print("CPU mode")
model.float()
for filename in os.listdir(opt.input_dir):
ext = os.path.splitext(filename)[1]
if ext not in valid_ext:
continue
print(filename)
# load image
if ext == ".gif":
if not os.path.exists("tmp"):
os.mkdir("tmp")
else:
shutil.rmtree("tmp")
os.mkdir("tmp")
input_gif = Image.open(os.path.join(opt.input_dir, filename))
for nframe in range(input_gif.n_frames):
print(" {} / {}".format(nframe, input_gif.n_frames), end="\r")
input_gif.seek(nframe)
output_image = convert_image(
model,
input_gif.split()[0].convert("RGB"))
save(image=output_image,
name="tmp/{name}_{nframe:04d}.jpg".format(
**{
"dir": opt.output_dir,
"name": "{}_{}".format(filename[:-4], opt.style),
"nframe": nframe
}))
jpg_to_gif(input_gif, filename)
shutil.rmtree("tmp")
else:
input_image = Image.open(os.path.join(opt.input_dir,
filename)).convert("RGB")
output_image = convert_image(model, input_image)
# save
save(image=output_image,
name="{dir}/{name}.jpg".format(
**{
"dir": opt.output_dir,
"name": "{}_{}".format(filename[:-4], opt.style)
}))
print("Done!")
parser = argparse.ArgumentParser()
parser.add_argument('--input_dir', default = 'test_img')
parser.add_argument('--load_size', default = 450)
parser.add_argument('--model_path', default = './pretrained_model')
parser.add_argument('--style', default = 'Hayao')
parser.add_argument('--output_dir', default = 'test_output')
parser.add_argument('--gpu', type=int, default = 0)
opt = parser.parse_args()
valid_ext = ['.jpg', '.png']
if not os.path.exists(opt.output_dir): os.mkdir(opt.output_dir)
# load pretrained model
model = Transformer()
model.load_state_dict(torch.load(os.path.join(opt.model_path, opt.style + '_net_G_float.pth')))
model.eval()
if opt.gpu > -1:
print('GPU mode')
model.cuda()
else:
print('CPU mode')
model.float()
for files in os.listdir(opt.input_dir):
ext = os.path.splitext(files)[1]
if ext not in valid_ext:
continue
# load image
