Download pre-trained models, in case you're in a hurry¶
In [1]:
#download pretrained models
!git clone https://github.com/Mainakdeb/deceptive-digits/
Cloning into 'deceptive-digits'... remote: Enumerating objects: 125, done. remote: Counting objects: 100% (125/125), done. remote: Compressing objects: 100% (122/122), done. remote: Total 125 (delta 40), reused 12 (delta 2), pack-reused 0 Receiving objects: 100% (125/125), 38.39 MiB | 27.69 MiB/s, done. Resolving deltas: 100% (40/40), done.
Imports¶
In [2]:
import torch
import torchvision.transforms as transforms
import torchvision.utils as vutils
import torchvision.datasets as datasets
from torch.utils.data import DataLoader
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from torch.autograd import Variable
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.image as mimg
Misc:¶
In [ ]:
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
workers = 2
batch_size = 64
image_size = 64
nc = 1 # Number of channels in the training images
nz = 100 # Size of z latent vector
ngf = 64 # Size of feature maps in generator
ndf = 64 # Size of feature maps in discriminator
lr = 2e-4 # Learning rate for optimizers
beta1 = 0.5 # eBta1 hyperparam for Adam optimizers
ngpu = 1
transforms = transforms.Compose(
[
transforms.Resize(image_size),
transforms.RandomRotation(25),
transforms.ToTensor(),
transforms.Normalize(
[0.5 for _ in range(nc)],
[0.5 for _ in range(nc)]
)
]
)
#dataset and dataloader
dataset = datasets.MNIST(root="/dataset/", train=True, transform=transforms, download=True)
dataloader = DataLoader(dataset, batch_size = batch_size, shuffle=True, drop_last=True)
Test if GPU is available¶
In [4]:
device
Out[4]:
device(type='cuda')
Plot some samples from the trainloader¶
In [5]:
real_batch = next(iter(dataloader))
plt.figure(figsize=(8,8))
plt.axis("off")
plt.title("Training Images")
plt.imshow(np.transpose(vutils.make_grid(real_batch[0].to(device)[:64], padding=2, normalize=True).cpu(),(1,2,0)))
Out[5]:
<matplotlib.image.AxesImage at 0x7f6ad6e5bed0>
The Generator network:¶
I concatenated a random noise vector of length 100 with another vector that represents a label of length 10 and passed the resulting tensor through the generator net. Notice that the input tensor has a length of 110. This input vector is passed through transpose-convolution layers to generate a 1x64x64 image.
In [6]:
class Generator(nn.Module):
def __init__(self, params):
super().__init__()
self.label_emb = nn.Embedding(10, 10)
self.tconv1 = nn.ConvTranspose2d(nz + 10, ngf * 8, kernel_size=4, stride=1, padding=0, bias=False)
self.bn1 = nn.BatchNorm2d(ngf * 8)
self.tconv2 = nn.ConvTranspose2d(ngf * 8, ngf * 4, 4, 2, 1, bias=False)
self.bn2 = nn.BatchNorm2d(ngf * 4)
self.tconv3 = nn.ConvTranspose2d(ngf * 4, ngf * 2, 4, 2, 1, bias=False)
self.bn3 = nn.BatchNorm2d(ngf * 2)
self.tconv4 = nn.ConvTranspose2d(ngf * 2, ngf, 4, 2, 1, bias=False)
self.bn4 = nn.BatchNorm2d(ngf)
self.tconv6 = nn.ConvTranspose2d(ngf, nc, 4, 2, 1, bias=False)
def forward(self, x, labels):
c = self.label_emb(labels)
c = c.unsqueeze(2).unsqueeze(3)
# print(c.size())
# print(x.size())
x = torch.cat([x, c], 1)
x = F.relu(self.bn1(self.tconv1(x)))
x = F.relu(self.bn2(self.tconv2(x)))
x = F.relu(self.bn3(self.tconv3(x)))
x = F.relu(self.bn4(self.tconv4(x)))
# x = F.relu(self.bn5(self.tconv5(x)))
x = torch.tanh(self.tconv6(x))
return x
netG = Generator(ngpu).to(device)
print(netG)
Generator( (label_emb): Embedding(10, 10) (tconv1): ConvTranspose2d(110, 512, kernel_size=(4, 4), stride=(1, 1), bias=False) (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (tconv2): ConvTranspose2d(512, 256, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), bias=False) (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (tconv3): ConvTranspose2d(256, 128, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), bias=False) (bn3): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (tconv4): ConvTranspose2d(128, 64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), bias=False) (bn4): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (tconv6): ConvTranspose2d(64, 1, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), bias=False) )
The Discriminator network:¶
A 1x64x64 image is passed through convolution layers and the resulting tensor of length 10 is concatenated with a label embedding of length 10 and the resulting tensor is passed through the linear layers. Notice that the first linear layer takes an input of length 20. The final output is a tensor of length 1, which represents the probability of the sample being real or fake.
In [7]:
class Discriminator(nn.Module):
def __init__(self, params):
super().__init__()
# meta data (label)
self.label_emb = nn.Embedding(10, 10)
self.conv1 = nn.Conv2d(nc, ndf, 4, 2, 1, bias=False)
self.conv3 = nn.Conv2d(ndf, ndf * 2, 4, 2, 1, bias=False)
self.bn3 = nn.BatchNorm2d(ndf * 2)
self.conv4 = nn.Conv2d(ndf * 2, ndf * 4, 4, 2, 1, bias=False)
self.bn4 = nn.BatchNorm2d(ndf * 4)
self.conv5 = nn.Conv2d(ndf * 4, ndf * 8, 4, 2, 1, bias=False)
self.bn5 = nn.BatchNorm2d(ndf * 8)
self.conv6 = nn.Conv2d(ndf * 8, 10, 4, 1, 0, bias=False)
self.fc1 = nn.Linear(20, 15)
self.fc2 = nn.Linear(15, 1)
def forward(self, x, labels):
x = F.leaky_relu(self.conv1(x), 0.2, True)
# print(x.size())
# x = F.leaky_relu(self.bn2(self.conv2(x)), 0.2, True)
# print(x.size())
x = F.leaky_relu(self.bn3(self.conv3(x)), 0.2, True)
x = F.leaky_relu(self.bn4(self.conv4(x)), 0.2, True)
x = F.leaky_relu(self.bn5(self.conv5(x)), 0.2, True)
x = F.leaky_relu(self.conv6(x))
x = torch.flatten(x, 1)
c = self.label_emb(labels)
# print(x.size())
# print(c.size())
x = torch.cat([x, c], 1)
# print(x.size())
# print(c.size())
x = F.leaky_relu(self.fc1(x))
x = F.sigmoid(self.fc2(x))
# print(x.size())
return x
netD = Discriminator(ngpu).to(device)
print(netD)
Discriminator( (label_emb): Embedding(10, 10) (conv1): Conv2d(1, 64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), bias=False) (conv3): Conv2d(64, 128, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), bias=False) (bn3): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (conv4): Conv2d(128, 256, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), bias=False) (bn4): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (conv5): Conv2d(256, 512, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), bias=False) (bn5): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (conv6): Conv2d(512, 10, kernel_size=(4, 4), stride=(1, 1), bias=False) (fc1): Linear(in_features=20, out_features=15, bias=True) (fc2): Linear(in_features=15, out_features=1, bias=True) )
Initiate Weights¶
In the DCGAN paper, the authors specify that all model weights shall be randomly initialized from a Normal distribution with mean=0, stdev=0.02. The weights_init function takes an initialized model as input and reinitializes all convolutional, convolutional-transpose, and batch normalization layers to meet this criteria.
In [8]:
def weights_init(m):
classname = m.__class__.__name__
if classname.find('Conv') != -1:
nn.init.normal_(m.weight.data, 0.0, 0.02)
elif classname.find('BatchNorm') != -1:
nn.init.normal_(m.weight.data, 1.0, 0.02)
nn.init.constant_(m.bias.data, 0)
netD.apply(weights_init)
netG.apply(weights_init)
Out[8]:
Generator( (label_emb): Embedding(10, 10) (tconv1): ConvTranspose2d(110, 512, kernel_size=(4, 4), stride=(1, 1), bias=False) (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (tconv2): ConvTranspose2d(512, 256, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), bias=False) (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (tconv3): ConvTranspose2d(256, 128, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), bias=False) (bn3): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (tconv4): ConvTranspose2d(128, 64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), bias=False) (bn4): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (tconv6): ConvTranspose2d(64, 1, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), bias=False) )
Load pretrained weights, in case you don't want to train.¶
In [9]:
# load
netG.load_state_dict(torch.load("/content/deceptive-digits/models/generator_w.pth"))
netD.load_state_dict(torch.load("/content/deceptive-digits/models/discriminator_w.pth"))
Out[9]:
<All keys matched successfully>
In [10]:
fixed_noise = torch.randn(batch_size, nz, 1, 1, device=device)
real_label = 1.
fake_label = 0.
criterion = nn.BCELoss()
optimizerD = optim.Adam(netD.parameters(), lr=lr, betas=(beta1, 0.999))
optimizerG = optim.Adam(netG.parameters(), lr=lr, betas=(beta1, 0.999))
img_list = []
G_losses = []
D_losses = []
iters = 0
Training Loop¶
In [ ]:
num_epochs = 6
for epoch in range(num_epochs):
for i, data in enumerate(dataloader, 0):
# (1) Update D network: maximize log(D(x)) + log(1 - D(G(z)))
#print((data[1]))
netD.zero_grad()
# Format batch
real_cpu = data[0].to(device)
real_labels = data[1].to(device)
b_size = real_cpu.size(0)
label = torch.full((b_size,), real_label, dtype=torch.float, device=device)
# Forward pass real batch through D
output = netD(real_cpu, real_labels).view(-1)
# Calculate loss on all-real batch
# print(output[0])
# print(label[0])
errD_real = criterion(output, label)
# Calculate gradients for D in backward pass
errD_real.backward()
D_x = output.mean().item()
## Train with all-fake batch
# Generate batch of latent vectors
noise = torch.randn(b_size, nz, 1, 1, device=device)
fake_labels = Variable(torch.LongTensor(np.random.randint(0, 10, batch_size))).to(device)
# Generate fake image batch with G
fake = netG(noise, fake_labels)
label.fill_(fake_label)
# Classify all fake batch with D
output = netD(fake.detach(), fake_labels).view(-1)
# Calculate D's loss on the all-fake batch
errD_fake = criterion(output, label)
# Calculate the gradients for this batch
errD_fake.backward()
D_G_z1 = output.mean().item()
# Add the gradients from the all-real and all-fake batches
errD = errD_real + errD_fake
# Update D
optimizerD.step()
############################
# (2) Update G network: maximize log(D(G(z)))
###########################
netG.zero_grad()
label.fill_(real_label) # fake labels are real for generator cost
# Since we just updated D, perform another forward pass of all-fake batch through D
output = netD(fake, fake_labels).view(-1)
# Calculate G's loss based on this output
errG = criterion(output, label)
# Calculate gradients for G
errG.backward()
D_G_z2 = output.mean().item()
# Update G
optimizerG.step()
# print training stats
if i % 50 == 0:
print('[%d/%d][%d/%d]\tLoss_D: %.4f\tLoss_G: %.4f\tD(x): %.4f\tD(G(z)): %.4f / %.4f'
% (epoch, num_epochs, i, len(dataloader),
errD.item(), errG.item(), D_x, D_G_z1, D_G_z2))
G_losses.append(errG.item())
D_losses.append(errD.item())
# Check how the generator is doing by saving G's output on fixed_noise
if (iters % 500 == 0) or ((epoch == num_epochs-1) and (i == len(dataloader)-1)):
with torch.no_grad():
fake = netG(fixed_noise, fake_labels).detach().cpu()
img_list.append(vutils.make_grid(fake, padding=2, normalize=True))
iters += 1
[0/6][0/937] Loss_D: 1.3971 Loss_G: 0.8272 D(x): 0.4695 D(G(z)): 0.4724 / 0.4382
/usr/local/lib/python3.6/dist-packages/torch/nn/functional.py:1639: UserWarning: nn.functional.sigmoid is deprecated. Use torch.sigmoid instead.
warnings.warn("nn.functional.sigmoid is deprecated. Use torch.sigmoid instead.")
/usr/local/lib/python3.6/dist-packages/torch/nn/functional.py:1628: UserWarning: nn.functional.tanh is deprecated. Use torch.tanh instead.
warnings.warn("nn.functional.tanh is deprecated. Use torch.tanh instead.")
[0/6][50/937] Loss_D: 0.0842 Loss_G: 4.3825 D(x): 0.9752 D(G(z)): 0.0569 / 0.0129 [0/6][100/937] Loss_D: 0.2943 Loss_G: 4.7062 D(x): 0.9756 D(G(z)): 0.2324 / 0.0096 [0/6][150/937] Loss_D: 0.3116 Loss_G: 3.8952 D(x): 0.9245 D(G(z)): 0.1988 / 0.0226 [0/6][200/937] Loss_D: 0.7395 Loss_G: 3.4416 D(x): 0.8753 D(G(z)): 0.4216 / 0.0349 [0/6][250/937] Loss_D: 0.2182 Loss_G: 2.9372 D(x): 0.9438 D(G(z)): 0.1449 / 0.0570 [0/6][300/937] Loss_D: 0.4128 Loss_G: 2.7404 D(x): 0.8790 D(G(z)): 0.2350 / 0.0693 [0/6][350/937] Loss_D: 2.2313 Loss_G: 3.4149 D(x): 0.1429 D(G(z)): 0.0099 / 0.0363 [0/6][400/937] Loss_D: 4.1659 Loss_G: 0.0647 D(x): 0.9923 D(G(z)): 0.9831 / 0.9380 [0/6][450/937] Loss_D: 0.6280 Loss_G: 3.0119 D(x): 0.8818 D(G(z)): 0.3645 / 0.0545 [0/6][500/937] Loss_D: 0.9133 Loss_G: 1.4494 D(x): 0.6444 D(G(z)): 0.3306 / 0.2701 [0/6][550/937] Loss_D: 0.8559 Loss_G: 1.5046 D(x): 0.6863 D(G(z)): 0.3354 / 0.2476 [0/6][600/937] Loss_D: 0.6936 Loss_G: 3.5186 D(x): 0.9078 D(G(z)): 0.4189 / 0.0313 [0/6][650/937] Loss_D: 0.8711 Loss_G: 2.0018 D(x): 0.8461 D(G(z)): 0.4568 / 0.1588 [0/6][700/937] Loss_D: 1.0857 Loss_G: 2.5062 D(x): 0.8720 D(G(z)): 0.5794 / 0.0879 [0/6][750/937] Loss_D: 0.9209 Loss_G: 1.3198 D(x): 0.4980 D(G(z)): 0.0986 / 0.3052 [0/6][800/937] Loss_D: 0.5540 Loss_G: 1.5304 D(x): 0.7135 D(G(z)): 0.1686 / 0.2404 [0/6][850/937] Loss_D: 0.7702 Loss_G: 1.0274 D(x): 0.6089 D(G(z)): 0.1568 / 0.3869 [0/6][900/937] Loss_D: 0.5993 Loss_G: 2.4905 D(x): 0.7752 D(G(z)): 0.2608 / 0.0944 [1/6][0/937] Loss_D: 0.7757 Loss_G: 1.3793 D(x): 0.6744 D(G(z)): 0.2614 / 0.2819 [1/6][50/937] Loss_D: 0.7228 Loss_G: 1.4023 D(x): 0.6700 D(G(z)): 0.2307 / 0.2931 [1/6][100/937] Loss_D: 0.9148 Loss_G: 2.2850 D(x): 0.8692 D(G(z)): 0.4818 / 0.1275 [1/6][150/937] Loss_D: 0.5821 Loss_G: 1.8017 D(x): 0.7699 D(G(z)): 0.2365 / 0.1910 [1/6][200/937] Loss_D: 0.6411 Loss_G: 2.5631 D(x): 0.8494 D(G(z)): 0.3406 / 0.0970 [1/6][250/937] Loss_D: 0.7635 Loss_G: 1.5612 D(x): 0.6437 D(G(z)): 0.2274 / 0.2448 [1/6][300/937] Loss_D: 0.8013 Loss_G: 1.5029 D(x): 0.5807 D(G(z)): 0.1512 / 0.2600 [1/6][350/937] Loss_D: 0.9978 Loss_G: 2.7179 D(x): 0.9118 D(G(z)): 0.5325 / 0.0883 [1/6][400/937] Loss_D: 0.8611 Loss_G: 1.2808 D(x): 0.5788 D(G(z)): 0.1938 / 0.3281 [1/6][450/937] Loss_D: 0.6313 Loss_G: 1.9585 D(x): 0.7765 D(G(z)): 0.2858 / 0.1746 [1/6][500/937] Loss_D: 0.7391 Loss_G: 2.4554 D(x): 0.7699 D(G(z)): 0.3353 / 0.1029 [1/6][550/937] Loss_D: 0.9106 Loss_G: 1.9302 D(x): 0.7891 D(G(z)): 0.4410 / 0.1772 [1/6][600/937] Loss_D: 0.6591 Loss_G: 2.1399 D(x): 0.7958 D(G(z)): 0.3036 / 0.1408 [1/6][650/937] Loss_D: 0.7588 Loss_G: 2.7239 D(x): 0.8722 D(G(z)): 0.4266 / 0.0802 [1/6][700/937] Loss_D: 0.6461 Loss_G: 2.1614 D(x): 0.8251 D(G(z)): 0.3364 / 0.1338 [1/6][750/937] Loss_D: 0.8830 Loss_G: 2.7413 D(x): 0.9305 D(G(z)): 0.5074 / 0.0893 [1/6][800/937] Loss_D: 0.6280 Loss_G: 1.9033 D(x): 0.7236 D(G(z)): 0.2164 / 0.1773 [1/6][850/937] Loss_D: 0.6661 Loss_G: 1.9002 D(x): 0.6786 D(G(z)): 0.1895 / 0.1801 [1/6][900/937] Loss_D: 0.6538 Loss_G: 2.6403 D(x): 0.8044 D(G(z)): 0.2855 / 0.0967 [2/6][0/937] Loss_D: 0.5194 Loss_G: 1.9279 D(x): 0.7745 D(G(z)): 0.2098 / 0.1654 [2/6][50/937] Loss_D: 0.5617 Loss_G: 1.9816 D(x): 0.7100 D(G(z)): 0.1482 / 0.1708 [2/6][100/937] Loss_D: 0.3800 Loss_G: 2.8294 D(x): 0.8495 D(G(z)): 0.1699 / 0.0781 [2/6][150/937] Loss_D: 0.6115 Loss_G: 1.9527 D(x): 0.7552 D(G(z)): 0.2306 / 0.1707 [2/6][200/937] Loss_D: 0.5846 Loss_G: 1.9586 D(x): 0.7123 D(G(z)): 0.1596 / 0.1753 [2/6][250/937] Loss_D: 0.4682 Loss_G: 2.1909 D(x): 0.7581 D(G(z)): 0.1341 / 0.1476 [2/6][300/937] Loss_D: 0.6942 Loss_G: 2.1204 D(x): 0.7991 D(G(z)): 0.3167 / 0.1555 [2/6][350/937] Loss_D: 0.5770 Loss_G: 2.1048 D(x): 0.7736 D(G(z)): 0.2133 / 0.1654 [2/6][400/937] Loss_D: 0.4831 Loss_G: 2.8035 D(x): 0.8641 D(G(z)): 0.2354 / 0.0837 [2/6][450/937] Loss_D: 0.5061 Loss_G: 1.8743 D(x): 0.7713 D(G(z)): 0.1754 / 0.1784 [2/6][500/937] Loss_D: 0.6322 Loss_G: 2.3005 D(x): 0.7744 D(G(z)): 0.2387 / 0.1428 [2/6][550/937] Loss_D: 0.6574 Loss_G: 2.2699 D(x): 0.7679 D(G(z)): 0.2409 / 0.1575 [2/6][600/937] Loss_D: 0.6311 Loss_G: 2.7589 D(x): 0.7257 D(G(z)): 0.1886 / 0.0993 [2/6][650/937] Loss_D: 0.6443 Loss_G: 2.1509 D(x): 0.8020 D(G(z)): 0.2779 / 0.1676 [2/6][700/937] Loss_D: 0.6521 Loss_G: 2.6712 D(x): 0.8417 D(G(z)): 0.3076 / 0.1074 [2/6][750/937] Loss_D: 0.6473 Loss_G: 2.1399 D(x): 0.8063 D(G(z)): 0.2996 / 0.1602 [2/6][800/937] Loss_D: 0.8931 Loss_G: 1.3559 D(x): 0.5304 D(G(z)): 0.0916 / 0.3730 [2/6][850/937] Loss_D: 0.6838 Loss_G: 2.6148 D(x): 0.7160 D(G(z)): 0.2071 / 0.1167 [2/6][900/937] Loss_D: 0.9255 Loss_G: 1.2604 D(x): 0.5124 D(G(z)): 0.1315 / 0.3517 [3/6][0/937] Loss_D: 0.7366 Loss_G: 2.1959 D(x): 0.7565 D(G(z)): 0.2887 / 0.1741 [3/6][50/937] Loss_D: 1.0536 Loss_G: 3.2784 D(x): 0.9458 D(G(z)): 0.5445 / 0.0610 [3/6][100/937] Loss_D: 0.7529 Loss_G: 1.6781 D(x): 0.6530 D(G(z)): 0.2017 / 0.2429 [3/6][150/937] Loss_D: 1.0734 Loss_G: 1.2007 D(x): 0.4430 D(G(z)): 0.0949 / 0.3909 [3/6][200/937] Loss_D: 0.7603 Loss_G: 1.6739 D(x): 0.6697 D(G(z)): 0.2230 / 0.2261 [3/6][250/937] Loss_D: 0.7354 Loss_G: 1.4602 D(x): 0.6352 D(G(z)): 0.1611 / 0.3046 [3/6][300/937] Loss_D: 1.3889 Loss_G: 1.7320 D(x): 0.6987 D(G(z)): 0.5518 / 0.2282 [3/6][350/937] Loss_D: 0.9607 Loss_G: 1.1747 D(x): 0.5583 D(G(z)): 0.2123 / 0.3771 [3/6][400/937] Loss_D: 0.7790 Loss_G: 2.2825 D(x): 0.7496 D(G(z)): 0.3221 / 0.1313 [3/6][450/937] Loss_D: 0.8229 Loss_G: 1.6721 D(x): 0.7320 D(G(z)): 0.3311 / 0.2460 [3/6][500/937] Loss_D: 0.8470 Loss_G: 1.4376 D(x): 0.5829 D(G(z)): 0.1670 / 0.2826 [3/6][550/937] Loss_D: 0.5949 Loss_G: 1.7149 D(x): 0.7114 D(G(z)): 0.1652 / 0.2227 [3/6][600/937] Loss_D: 1.1493 Loss_G: 3.6192 D(x): 0.9431 D(G(z)): 0.5955 / 0.0383 [3/6][650/937] Loss_D: 1.4633 Loss_G: 1.1568 D(x): 0.7382 D(G(z)): 0.6039 / 0.3708 [3/6][700/937] Loss_D: 0.7688 Loss_G: 1.5833 D(x): 0.7419 D(G(z)): 0.3252 / 0.2557 [3/6][750/937] Loss_D: 0.8297 Loss_G: 1.6556 D(x): 0.7387 D(G(z)): 0.3431 / 0.2468 [3/6][800/937] Loss_D: 1.1064 Loss_G: 1.9635 D(x): 0.7851 D(G(z)): 0.5087 / 0.1928 [3/6][850/937] Loss_D: 0.9958 Loss_G: 1.7125 D(x): 0.7542 D(G(z)): 0.4460 / 0.2246 [3/6][900/937] Loss_D: 0.8332 Loss_G: 1.5794 D(x): 0.6712 D(G(z)): 0.2693 / 0.2579 [4/6][0/937] Loss_D: 0.7527 Loss_G: 2.5498 D(x): 0.6912 D(G(z)): 0.2502 / 0.1059 [4/6][50/937] Loss_D: 1.0783 Loss_G: 1.9825 D(x): 0.7103 D(G(z)): 0.4452 / 0.1775 [4/6][100/937] Loss_D: 1.3310 Loss_G: 0.7294 D(x): 0.3785 D(G(z)): 0.1405 / 0.5262 [4/6][150/937] Loss_D: 0.8747 Loss_G: 2.3976 D(x): 0.8541 D(G(z)): 0.4427 / 0.1198 [4/6][200/937] Loss_D: 0.7489 Loss_G: 1.4591 D(x): 0.6294 D(G(z)): 0.1773 / 0.2868 [4/6][250/937] Loss_D: 0.9753 Loss_G: 1.8635 D(x): 0.7895 D(G(z)): 0.4552 / 0.2029 [4/6][300/937] Loss_D: 1.0311 Loss_G: 1.8205 D(x): 0.5503 D(G(z)): 0.2423 / 0.2114 [4/6][350/937] Loss_D: 1.0432 Loss_G: 0.9367 D(x): 0.5040 D(G(z)): 0.1967 / 0.4204 [4/6][400/937] Loss_D: 1.4200 Loss_G: 2.3837 D(x): 0.8474 D(G(z)): 0.6336 / 0.1370 [4/6][450/937] Loss_D: 0.9331 Loss_G: 1.5836 D(x): 0.6783 D(G(z)): 0.3480 / 0.2651 [4/6][500/937] Loss_D: 1.0904 Loss_G: 1.8768 D(x): 0.6608 D(G(z)): 0.3935 / 0.2006 [4/6][550/937] Loss_D: 1.1448 Loss_G: 1.3711 D(x): 0.5742 D(G(z)): 0.3842 / 0.2874 [4/6][600/937] Loss_D: 1.3558 Loss_G: 3.0157 D(x): 0.7660 D(G(z)): 0.6017 / 0.0650 [4/6][650/937] Loss_D: 1.2633 Loss_G: 2.1658 D(x): 0.7008 D(G(z)): 0.5326 / 0.1452 [4/6][700/937] Loss_D: 1.2083 Loss_G: 1.9865 D(x): 0.6481 D(G(z)): 0.4676 / 0.1585 [4/6][750/937] Loss_D: 1.1847 Loss_G: 1.3346 D(x): 0.5911 D(G(z)): 0.3854 / 0.3054 [4/6][800/937] Loss_D: 1.4082 Loss_G: 2.1748 D(x): 0.7900 D(G(z)): 0.6284 / 0.1572 [4/6][850/937] Loss_D: 1.1957 Loss_G: 1.3225 D(x): 0.6740 D(G(z)): 0.4614 / 0.3088 [4/6][900/937] Loss_D: 1.1786 Loss_G: 1.8380 D(x): 0.7154 D(G(z)): 0.4976 / 0.1870 [5/6][0/937] Loss_D: 1.1507 Loss_G: 1.4684 D(x): 0.6255 D(G(z)): 0.4345 / 0.2623 [5/6][50/937] Loss_D: 1.0926 Loss_G: 1.2723 D(x): 0.5991 D(G(z)): 0.3765 / 0.3213 [5/6][100/937] Loss_D: 1.0768 Loss_G: 1.5197 D(x): 0.7199 D(G(z)): 0.4753 / 0.2500 [5/6][150/937] Loss_D: 1.1411 Loss_G: 1.2212 D(x): 0.5903 D(G(z)): 0.3983 / 0.3231 [5/6][200/937] Loss_D: 1.1095 Loss_G: 1.2435 D(x): 0.5747 D(G(z)): 0.3661 / 0.3221 [5/6][250/937] Loss_D: 1.9720 Loss_G: 3.9835 D(x): 0.8853 D(G(z)): 0.7631 / 0.0296 [5/6][300/937] Loss_D: 1.0578 Loss_G: 1.7481 D(x): 0.6757 D(G(z)): 0.4372 / 0.2027 [5/6][350/937] Loss_D: 0.7972 Loss_G: 1.4888 D(x): 0.6977 D(G(z)): 0.3161 / 0.2637 [5/6][400/937] Loss_D: 0.8421 Loss_G: 2.0829 D(x): 0.7394 D(G(z)): 0.3842 / 0.1598 [5/6][450/937] Loss_D: 2.5561 Loss_G: 0.3430 D(x): 0.1122 D(G(z)): 0.0759 / 0.7270 [5/6][500/937] Loss_D: 1.0530 Loss_G: 1.2401 D(x): 0.5967 D(G(z)): 0.3583 / 0.3263 [5/6][550/937] Loss_D: 1.2677 Loss_G: 1.0243 D(x): 0.5966 D(G(z)): 0.4830 / 0.3926 [5/6][600/937] Loss_D: 1.4984 Loss_G: 0.9544 D(x): 0.3077 D(G(z)): 0.0869 / 0.4291 [5/6][650/937] Loss_D: 0.9258 Loss_G: 1.5528 D(x): 0.6910 D(G(z)): 0.3851 / 0.2311 [5/6][700/937] Loss_D: 0.6968 Loss_G: 3.2262 D(x): 0.8535 D(G(z)): 0.3587 / 0.0478 [5/6][750/937] Loss_D: 1.1165 Loss_G: 1.5030 D(x): 0.6897 D(G(z)): 0.4750 / 0.2569 [5/6][800/937] Loss_D: 0.8155 Loss_G: 2.5515 D(x): 0.8293 D(G(z)): 0.4170 / 0.1025 [5/6][850/937] Loss_D: 0.8408 Loss_G: 4.6982 D(x): 0.8395 D(G(z)): 0.4356 / 0.0149 [5/6][900/937] Loss_D: 0.8389 Loss_G: 1.5198 D(x): 0.6083 D(G(z)): 0.1918 / 0.2525
Visualize training metrics¶
In [ ]:
plt.figure(figsize=(10,5))
plt.title("Generator and Discriminator Loss During Training")
plt.plot(G_losses,label="Gen")
plt.plot(D_losses,label="Disc")
plt.xlabel("iterations")
plt.ylabel("Loss")
plt.legend()
plt.show()
Generate one digit:¶
In [37]:
def generate_digit_from_label(label, seed):
fake_label = torch.tensor([label]).cuda()
# print("sn", single_noise.shape)
# print("label", fake_label.shape)
with torch.no_grad():
fake_ = netG(seed, fake_label).detach().cpu()
# print(single_noise.shape, fake_label.shape)
# print(type(single_noise), type(fake_label))
return(fake_.squeeze())
seed = torch.randn(1, nz, 1, 1, device=device)
plt.imshow(generate_digit_from_label(0, seed))
Out[37]:
<matplotlib.image.AxesImage at 0x7f6ad21d27d0>
Generate images of all classes¶
In [38]:
fig, axes = plt.subplots(1,10, figsize = (20,8))
axes[0].set_title('label 0')
axes[0].imshow(generate_digit_from_label(0, seed), cmap="gray")
axes[1].set_title('label 1')
axes[1].imshow(generate_digit_from_label(1, seed), cmap="gray")
axes[2].set_title('label 2')
axes[2].imshow(generate_digit_from_label(2, seed), cmap="gray")
axes[3].set_title('label 3')
axes[3].imshow(generate_digit_from_label(3, seed), cmap="gray")
axes[4].set_title('label 4')
axes[4].imshow(generate_digit_from_label(4, seed), cmap="gray")
axes[5].set_title('label 5')
axes[5].imshow(generate_digit_from_label(5, seed), cmap="gray")
axes[6].set_title('label 6')
axes[6].imshow(generate_digit_from_label(6, seed), cmap="gray")
axes[7].set_title('label 7')
axes[7].imshow(generate_digit_from_label(7, seed), cmap="gray")
axes[8].set_title('label 8')
axes[8].imshow(generate_digit_from_label(8, seed), cmap="gray")
axes[9].set_title('label 9')
axes[9].imshow(generate_digit_from_label(9, seed), cmap="gray")
Out[38]:
<matplotlib.image.AxesImage at 0x7f6ad1c5be50>
Exploring the Latent Space:¶
The generator net (here) accepts a latent vector of length 100 and a label embeddding of length 10. While the network trains, it learns to map these latent points to generated images. Every single latent vector is a point in an n-dimensional space where n is the length of the latent vector, which is 100 in this case.
What if you take 2 points in this 100-dimensional space and generate samples by interpolating between them? Every adjacent point leads to the generation of a slightly different image.
The folllowing cell generates images with latent vectors that are interpolated b/w 2 random points.
In [39]:
def generate_latent_points(latent_dim, n_samples, n_classes=10):
# generate points in the latent space
x_input = np.random.randn(latent_dim * n_samples)
# reshape into a batch of inputs for the network
z_input = x_input.reshape(n_samples, latent_dim)
return z_input
# uniform interpolation between two points in latent space
def interpolate_points(p1, p2, n_steps=10):
# interpolate ratios between the points
ratios = np.linspace(0, 1, num=n_steps)
# linear interpolate vectors
vectors = list()
for ratio in ratios:
v = (1.0 - ratio) * p1 + ratio * p2
vectors.append(v)
return np.asarray(vectors)
n=20
pts = generate_latent_points(100, n)
interps = []
for i in range(0, n, 2):
interpolated = interpolate_points(pts[i], pts[i+1], n_steps = 10)
z = torch.tensor(interpolated).float().cuda()
z=z.unsqueeze(2).unsqueeze(3)
#print(z.shape)
labels_ = torch.LongTensor(np.ones(10, dtype=int)*i//2).cuda()
#print(labels_.shape)
images = netG(z,labels_).detach().cpu().numpy()
# for x in images:
# plt.imshow(x[0])
# plt.show()
interps.append(images)
interps=np.array(interps)
Visualize images with neighbouring latent points¶
Notice how each frame is slightly different from the previous. Lets visualize some interpolations side by side, with a different set of interpolated points:
In [40]:
#Big Plot
fig, axes = plt.subplots(10,10, figsize = (20,20))
for i in range(0, 10):
for j in range(0,10):
axes[i,j].set_title('label '+str(i))
axes[i,j].imshow(interps[i][j][0], cmap='gray')
plt.tight_layout()
plt.show()
The following gif showcases generated images with latent vectors interpolated between 2 points, looping back and forth between the 2 extremes.
Save interpolated images¶
The following cell was used to save frames, using which the above shown gif was made.
In [ ]:
#save interpolations
fig, axes = plt.subplots(1,10, figsize = (70,10))
for j in range(0,10):
for i in range(0, 10):
axes[i].set_title('label '+str(i), FONTSIZE='3')
axes[i].imshow(interps[i][j][0], cmap="gray")
fig.savefig("interpolation_"+str(j)+".png") #create gif
for i in range(0,10):
im = plt.imread("interpolation_"+str(i)+".png")
plt.imshow(im)
plt.show()
How Far Can We Go?¶
what if you wanted to generate images of numbers with multiple digits? I've used numpy to crop, invert and stich generated images horizontally, check out the generated image below.
In [41]:
def generate_and_save_image_with_multiple_digits(digits, seed):
gen_list=[]
for dig in str(digits):
gen_list.append(generate_digit_from_label(int(dig), seed)[4:60, 10:58])
vis = np.concatenate((gen_list), axis=1)
vis = 255-vis
mimg.imsave("generated"+str(digits)+".png", vis, cmap='gray')
return(vis)
In [42]:
plt.imshow(generate_and_save_image_with_multiple_digits(12110, seed), cmap='gray')
plt.axis("off")
plt.show()
plt.imshow(generate_and_save_image_with_multiple_digits(410, seed), cmap='gray')
plt.axis("off")
plt.show()
plt.imshow(generate_and_save_image_with_multiple_digits(6789, seed), cmap='gray')
plt.axis("off")
plt.show()
So, can I make the model write phone numbers for me?¶
The results speak for themselves.
In [45]:
d=9830976450
plt.imshow(generate_and_save_image_with_multiple_digits(d, seed), cmap='gray')
plt.title("input: "+str(d))
plt.show()
Save model weights:¶
In [44]:
torch.save(netG.state_dict(), "generator_w.pth")
torch.save(netD.state_dict(), "discriminator_w.pth")
In [ ]: