udithhaputhanthri

G=Generator()
filepath='gdrive/My Drive/pix2pixmodel/G_L1.pth'
G.load_state_dict(torch.load(filepath,map_location=torch.device('cpu')))
G.eval()

for x,y in dataloader:
  z=torch.randn((batch_size,1,128,128)).to(device)
  generated_imgs=G(x[:5],z[:5])
  real_imgs=x[:5]
  imgs=torch.cat([generated_imgs,real_imgs,y[:5]],0).data.cpu()

udithhaputhanthri

import zipfile
import os
if not os.path.isfile('celeba.zip'):
  !mkdir data_faces && wget https://s3-us-west-1.amazonaws.com/udacity-dlnfd/datasets/celeba.zip 
  with zipfile.ZipFile("celeba.zip","r") as zip_ref:
    zip_ref.extractall("data_faces/")
    
from torch.utils.data import DataLoader

transform = transforms.Compose([

udithhaputhanthri

class conv_trans_block(nn.Module):
  def __init__(self,in_channels,out_channels,kernal_size=4,stride=2,padding=1):
    super(conv_trans_block,self).__init__()
    self.block=nn.Sequential(
        nn.ConvTranspose2d(in_channels,out_channels,kernal_size,stride,padding),
        nn.BatchNorm2d(out_channels),
        nn.ReLU())
  def forward(self,x):
    return self.block(x)

udithhaputhanthri

class Generator(nn.Module):
  def __init__(self,noise_channels,img_channels,hidden_G):
    super(Generator,self).__init__()
    self.G=nn.Sequential(
        conv_trans_block(noise_channels,hidden_G*16,kernal_size=4,stride=1,padding=0),
        conv_trans_block(hidden_G*16,hidden_G*8),
        conv_trans_block(hidden_G*8,hidden_G*4),
        conv_trans_block(hidden_G*4,hidden_G*2),
        nn.ConvTranspose2d(hidden_G*2,img_channels,kernel_size=4,stride=2,padding=1),
        nn.Tanh()

udithhaputhanthri

def get_gradient(crit, real_imgs, fake_imgs, epsilon):

  mixed_imgs= real_imgs* epsilon + fake_imgs*(1- epsilon)
  mixed_scores= crit(mixed_imgs)

  gradient= torch.autograd.grad(outputs= mixed_scores,
                                inputs= mixed_imgs,
                                grad_outputs= torch.ones_like(mixed_scores),
                                create_graph=True,
                                retain_graph=True)[0]

udithhaputhanthri

def get_gen_loss(crit_fake_pred):
  gen_loss= -torch.mean(crit_fake_pred)
  return gen_loss

def get_crit_loss(crit_fake_pred, crit_real_pred, gradient_penalty, c_lambda):
  crit_loss= torch.mean(crit_fake_pred)- torch.mean(crit_real_pred)+ c_lambda* gradient_penalty
  return crit_loss

udithhaputhanthri

C=Critic(img_channels,hidden_C).to(device)
G=Generator(noise_channels,img_channels,hidden_G).to(device)

#C=C.apply(init_weights)
#G=G.apply(init_weights)

wandb.watch(G, log='all', log_freq=10)
wandb.watch(C, log='all', log_freq=10)

opt_C=torch.optim.Adam(C.parameters(),lr=lr, betas=(0.5,0.999))