|
import torch.nn.functional as F |
|
from models.networks.base_network import BaseNetwork |
|
import torch |
|
import torch.nn as nn |
|
from .generator_module.StyTR import StyTr |
|
from torchvision.utils import save_image |
|
|
|
from .generator_module.transformerEncoder import TransformerEncoder |
|
from .generator_module.HEtransformerEncoder import HeTransformerEncoder |
|
from .generator_module.schedule import CosineAnnealingWarmUpLR |
|
from .generator_module.DecoderCNN import Decoder_MV, vgg_structures,decoder_stem |
|
from .generator_module.transformer_decoder import TransformerDecoder |
|
|
|
class TSITGenerator(BaseNetwork): |
|
@staticmethod |
|
def modify_commandline_options(parser, is_train): |
|
parser.set_defaults(norm_G='spectralfadesyncbatch3x3') |
|
parser.add_argument('--num_upsampling_layers', |
|
choices=('normal', 'more', 'most'), default='more', |
|
help="If 'more', adds upsampling layer between the two middle resnet blocks." |
|
"If 'most', also add one more upsampling + resnet layer at the end of the generator." |
|
"We only use 'more' as the default setting.") |
|
parser.add_argument('--lr_decay', type=float, default=1e-4, help='learning rate decay') |
|
parser.add_argument('--lr_stytr2', type=float, default=5e-4, help='initial learning rate for adam') |
|
return parser |
|
|
|
def __init__(self, opt): |
|
super().__init__() |
|
self.vgg_path = r'models/networks/experiments/vgg_normalised.pth' |
|
self.CNNdecoder = Decoder_MV(d_model=768,seq_input=True) |
|
self.HeTransEncoder=HeTransformerEncoder(img_size=224,patch_size=2,in_chans=3,embed_dim=192,depths=[2, 2, 2],nhead=[3, 6, 12],strip_width=[2, 4, 7],drop_path_rate=0.,patch_norm=True) |
|
self.TransEncoder= TransformerEncoder(img_size=224,patch_size=2,in_chans=3,embed_dim=192,depths=[2, 2, 2],nhead=[3, 6, 12],strip_width=[2, 4, 7],drop_path_rate=0.,patch_norm=True) |
|
self.TransDecoder = TransformerDecoder(nlayer=3,d_model=768,nhead=8,mlp_ratio=4,qkv_bias=False,attn_drop=0.,drop=0.,drop_path=0.,act_layer=nn.GELU,norm_layer=nn.LayerNorm,norm_first=True) |
|
self.vgg=vgg_structures |
|
|
|
self.content_weight= opt.content_weight |
|
self.style_weight = opt.style_weight |
|
self.lr_decay = opt.lr_decay |
|
self.lr_stytr2 = opt.lr_stytr2 |
|
self.device="cuda" if torch.cuda.is_available() else "cpu" |
|
self.iters_count=0 |
|
self.save_dir= opt.save_dir |
|
self.opt=opt |
|
self.mode=opt.mode |
|
|
|
def SetUp_model(self, TransEncoder,HeTransEncoder ,CNN_decoder,TransDecoder,vgg,alpha,mode): |
|
vgg.load_state_dict(torch.load(self.vgg_path)) |
|
|
|
|
|
|
|
network=StyTr(TransEncoder,HeTransEncoder,CNN_decoder,TransDecoder,vgg,alpha,mode) |
|
network.train() |
|
network.to(self.device) |
|
return network |
|
|
|
def forward(self): |
|
network = self.SetUp_model(self.TransEncoder,self.HeTransEncoder,self.CNNdecoder, self.TransDecoder, self.vgg,self.opt.alpha,self.mode) |
|
return network |
|
|
|
|
