Hello there,
I am trying to create a Deep Fake using an autoencoder. I use one encoder and two decoders: one for the target image, and another for the source image (the target-face is the face I want to “paste” on the sources head). So first I am trying to train the encoder and both decoders to reconstruct the input faces (300, 300, 3). But for both decoders, the output is not a colorful image but its gray, because for each pixel the Red, Green and Blue value are almost the same. And additionally to that, there is a weird 3x3 grid on the output images:
I am using a batch size of 1 because I dont know how to do it with minibatches in that case (but thats another problem). Im also using residual connections, which improved the quality. The last layer has a sigmoid activation (which could be wrong). My loss is Binary-Cross-Entropy and optimizer is Adam. Learning rate is 0.001 (I also tried 0.0001, and 0.00075).
Here is my model:
import matplotlib.pyplot as plt
import torchvision
import torch
import torch.nn as nn
import torch.nn.functional as F
class Model(nn.Module):
def __init__(self):
super(Model, self).__init__()
""" encoder """
self.conv1 = nn.Conv2d(3, 32, kernel_size=(4, 4))
self.batchnorm1 = nn.BatchNorm2d(32)
self.conv2 = nn.Conv2d(32, 64, kernel_size=(4, 4))
self.batchnorm2 = nn.BatchNorm2d(64)
self.conv3 = nn.Conv2d(64, 128, kernel_size=(3, 3))
self.batchnorm3 = nn.BatchNorm2d(128)
self.conv4 = nn.Conv2d(128, 256, kernel_size=(4, 4))
self.batchnorm4 = nn.BatchNorm2d(256)
self.maxpool3x3 = nn.MaxPool2d(3)
self.maxpool2x2 = nn.MaxPool2d(2)
""" target-decoder """
self.targetDeconv1 = nn.ConvTranspose2d(256, 128, kernel_size=(4, 4))
self.targetBatchnorm1 = nn.BatchNorm2d(128)
self.targetDeconv2 = nn.ConvTranspose2d(128, 64, kernel_size=(3, 3))
self.targetBatchnorm2 = nn.BatchNorm2d(64)
self.targetDeconv3 = nn.ConvTranspose2d(64, 32, kernel_size=(4, 4))
self.targetBatchnorm3 = nn.BatchNorm2d(32)
self.targetDeconv4 = nn.ConvTranspose2d(32, 3, kernel_size=(4, 4))
self.upsample3x3 = nn.Upsample(scale_factor=3)
self.upsample2x2 = nn.Upsample(scale_factor=2)
""" source-decoder """
self.sourceDeconv1 = nn.ConvTranspose2d(256, 128, kernel_size=(4, 4))
self.sourceBatchnorm1 = nn.BatchNorm2d(128)
self.sourceDeconv2 = nn.ConvTranspose2d(128, 64, kernel_size=(3, 3))
self.sourceBatchnorm2 = nn.BatchNorm2d(64)
self.sourceDeconv3 = nn.ConvTranspose2d(64, 32, kernel_size=(4, 4))
self.sourceBatchnorm3 = nn.BatchNorm2d(32)
self.sourceDeconv4 = nn.ConvTranspose2d(32, 3, kernel_size=(4, 4))
self.upsample3x3 = nn.Upsample(scale_factor=3)
self.upsample2x2 = nn.Upsample(scale_factor=2)
def _visualize_features(self, feature_maps, dim: tuple=(), title: str=""):
try:
x, y = dim
fig, axs = plt.subplots(x, y)
c = 0
for i in range(x):
for j in range(y):
axs[i][j].matshow(feature_maps.detach().cpu().numpy()[0][c])
c += 1
fig.suptitle(title)
plt.show()
except:
pass
def forward(self, x, label: str="0", visualize: bool=False):
""" encoder """
x = self.conv1(x)
x = self.batchnorm1(x)
x = F.relu(x)
x_1 = self.maxpool3x3(x)
if visualize: print(x_1.shape); self._visualize_features(x_1, dim=(4, 4))
x = self.conv2(x_1)
x = self.batchnorm2(x)
x = F.relu(x)
x_2 = self.maxpool3x3(x)
if visualize: print(x_2.shape); self._visualize_features(x_2, dim=(4, 4))
x = self.conv3(x_2)
x = self.batchnorm3(x)
x = F.relu(x)
x_3 = self.maxpool2x2(x)
if visualize: print(x_3.shape); self._visualize_features(x_3, dim=(4, 4))
x = self.conv4(x_3)
x = self.batchnorm4(x)
x = F.relu(x)
x = self.maxpool2x2(x)
if visualize: print(x.shape); self._visualize_features(x, dim=(4, 4))
""" target-decoder """
if label == "0":
x = self.upsample2x2(x)
x = self.targetDeconv1(x)
x += x_3
x = self.targetBatchnorm1(x)
x = F.relu(x)
if visualize: print(x.shape); self._visualize_features(x, dim=(4, 4))
x = self.upsample2x2(x)
x = self.targetDeconv2(x)
x += x_2
x = self.targetBatchnorm2(x)
x = F.relu(x)
if visualize: print(x.shape); self._visualize_features(x, dim=(4, 4))
x = self.upsample3x3(x)
x = self.targetDeconv3(x)
x += x_1
x = self.targetBatchnorm3(x)
x = F.relu(x)
if visualize: print(x.shape); self._visualize_features(x, dim=(4, 4))
x = self.upsample3x3(x)
x = self.targetDeconv4(x)
x = torch.sigmoid(x)
if visualize: print(x.shape); self._visualize_features(x, dim=(3, 1))
return x
""" source-decoder """
if label == "1":
x = self.upsample2x2(x)
x = self.sourceDeconv1(x)
x += x_3
x = self.sourceBatchnorm1(x)
x = F.relu(x)
if visualize: print(x.shape); self._visualize_features(x, dim=(4, 4))
x = self.upsample2x2(x)
x = self.sourceDeconv2(x)
x += x_2
x = self.sourceBatchnorm2(x)
x = F.relu(x)
if visualize: print(x.shape); self._visualize_features(x, dim=(4, 4))
x = self.upsample3x3(x)
x = self.sourceDeconv3(x)
x += x_1
x = self.sourceBatchnorm3(x)
x = F.relu(x)
if visualize: print(x.shape); self._visualize_features(x, dim=(4, 4))
x = self.upsample3x3(x)
x = self.sourceDeconv4(x)
x = torch.sigmoid(x)
if visualize: print(x.shape); self._visualize_features(x, dim=(3, 1))
return x
So my problems are: Why is the output image gray? What is that grid on my output image? Are both problems related? I checked if it has anything to do with rgb vs. bgr but it doesnt seem like it.
I hope anyone can fix my problem, thanks in advance 