## VAE generates bad images. due to unbalanced loss functions?

3

I'm training a variational autoencoder on CelebA dataset using TensorFlow.keras

The problem I'm facing is that the generated images are not diverse enough and look kinda bad.

(new) Example:

What I think:

• it's bad because the reconstruction and KL loss are unbalanced.
• I read this question and followed its solution - read about KL annealing and tried to implement it myself but didn't work.

Note:

• It's my first time working with autoencoders so maybe I missed something obvious.

• it would be super appreciated if you could give a programmatic/technical solution and not a theoretical one with equations and complicated math

The loss function:

def r_loss(self, y_true, y_pred):
return K.mean(K.square(y_true - y_pred), axis=[1, 2, 3])

def kl_loss(self, y_true, y_pred):
return  -0.5 * K.sum(1 + self.sd_layer - K.square(self.mean_layer) - K.exp(self.sd_layer), axis=1)

def total_loss(self, y_true, y_pred):
return K.mean(self.r_loss(y_true, y_pred) + self.kl_loss(y_true, y_pred))


The encoder:

    def build_encoder(self):

conv_filters = [32, 64, 64, 64]
conv_kernel_size = [3, 3, 3, 3]
conv_strides = [2, 2, 2, 2]

# Number of Conv layers
n_layers = len(conv_filters)

# Define model input
x = self.encoder_input

# Add convolutional layers
for i in range(n_layers):
x = Conv2D(filters=conv_filters[i],
kernel_size=conv_kernel_size[i],
strides=conv_strides[i],
name='encoder_conv_' + str(i)
)(x)
if self.use_batch_norm: # True
x = BatchNormalization()(x)

x = LeakyReLU()(x)

if self.use_dropout: # False
x = Dropout(rate=0.25)(x)

# Required for reshaping latent vector while building Decoder
self.shape_before_flattening = K.int_shape(x)[1:]

x = Flatten()(x)

self.mean_layer = Dense(self.encoder_output_dim, name='mu')(x)
self.sd_layer = Dense(self.encoder_output_dim, name='log_var')(x)

# Defining a function for sampling
def sampling(args):
mean_mu, log_var = args
epsilon = K.random_normal(shape=K.shape(mean_mu), mean=0., stddev=1.)
return mean_mu + K.exp(log_var / 2) * epsilon

# Using a Keras Lambda Layer to include the sampling function as a layer

# in the model
encoder_output = Lambda(sampling, name='encoder_output')([self.mean_layer, self.sd_layer])

return Model(self.encoder_input, encoder_output, name="VAE_Encoder")


The decoder:

def build_decoder(self):
conv_filters = [64, 64, 32, 3]
conv_kernel_size = [3, 3, 3, 3]
conv_strides = [2, 2, 2, 2]

n_layers = len(conv_filters)

# Define model input
decoder_input = self.decoder_input

# To get an exact mirror image of the encoder
x = Dense(np.prod(self.shape_before_flattening))(decoder_input)
x = Reshape(self.shape_before_flattening)(x)

# Add convolutional layers
for i in range(n_layers):
x = Conv2DTranspose(filters=conv_filters[i],
kernel_size=conv_kernel_size[i],
strides=conv_strides[i],
name='decoder_conv_' + str(i)
)(x)

# Adding a sigmoid layer at the end to restrict the outputs
# between 0 and 1
if i < n_layers - 1:
x = LeakyReLU()(x)
else:
x = Activation('sigmoid')(x)

# Define model output
self.decoder_output = x

return Model(decoder_input, self.decoder_output, name="VAE_Decoder")


The combined model:

def build_autoencoder(self):
self.encoder = self.build_encoder()
self.decoder = self.build_decoder()

# Input to the combined model will be the input to the encoder.
# Output of the combined model will be the output of the decoder.
self.autoencoder = Model(self.encoder_input, self.decoder(self.encoder(self.encoder_input)),
name="Variational_Auto_Encoder")

metrics=[self.total_loss],
experimental_run_tf_function=False)
self.autoencoder.summary()


EDIT:

the latent size is 256 and the sample method is as follows;

def generate(self, image=None):
if not os.path.exists(self.sample_dir):
os.makedirs(self.sample_dir)
if image is None:
img = np.random.normal(size=(9, self.encoder_output_dim))

prediction = self.decoder.predict(img)

op = np.vstack((np.hstack((prediction[0], prediction[1], prediction[2])),
np.hstack((prediction[3], prediction[4], prediction[5])),
np.hstack((prediction[6], prediction[7], prediction[8]))))
print(op.shape)
op = cv2.resize(op, (self.input_size * 9, self.input_size * 9), interpolation=cv2.INTER_AREA)
op = cv2.cvtColor(op, cv2.COLOR_BGR2RGB)
cv2.imshow("generated", op)
cv2.imwrite(self.sample_dir + "generated" + str(r(0, 9999)) + ".jpg", (op * 255).astype("uint8"))

else:
img = cv2.resize(img, (self.input_size, self.input_size), interpolation=cv2.INTER_AREA)
img = img.astype("float32")
img = img / 255

prediction = self.autoencoder.predict(img.reshape(1, self.input_size, self.input_size, 3))
img = cv2.resize(prediction[0][:, :, ::-1], (960, 960), interpolation=cv2.INTER_AREA)

cv2.imshow("prediction", img)

cv2.imwrite(self.sample_dir + "generated" + str(r(0, 9999)) + ".jpg", (img * 255).astype("uint8"))


Could you maybe also give the dimension of the latent space and how you sample from it? – matthiaw91 – 2020-06-02T11:22:41.383

@matthiaw91 edited the question. – Lidor shimoni – 2020-06-02T12:52:15.053

1

The issue is in your sampling procedure. The purpose of a VAE is to train a neural network, the decoder, that takes samples $$z$$ from a normal distribution $$p(z)$$ and maps them to images $$x$$ such that the images follow the original image distribution $$p(x)$$. The encoder's job is essentially to facillitate the training of the decoder, but for sampling it is not needed.

What you do is that you sample an image with random pixel values, which has nothing to do with the original image distribution $$p(x)$$, and map it to the latent space. The encoder is trained to map images to the latent space, not noise, hence the encoding is way off.

Since the images with the normaly distributed values in the pixels are probably all similarly "wrong" compared to $$p(x)$$, they get mapped to a similar domain in the latent space and hence produce similar outputs.

For generation of new samples you only need the decoder, so instead of sampling images with normally distributed pixel values, sample normally distributed vectors in 256 dimensions and pass those through the decoder only.

Side note: it seems a bit odd to me that you do not use fully-connected layers with non-linearities in the end of the encoder / beginning of decoder. If it works with only a linear mapping from the last feature map to the latent space, then it's fine, but intuitively I would have assumed that there should be at least one fully-connected layer with non-linear activation. But again, if it works then don't worry.

thanks a lot for your help! I didn't notice the silly mistake in the sampling method and now it's fixed. but even after the fix, I got the same bad generated samples, updated the question with the new samples, and added the full code of the sampling method, I'm not only sampling new images using only the decoder I also tried reconstructing a given image and the results were the same. your help and patience are really appreciated! – Lidor shimoni – 2020-06-02T17:35:10.407

Puh, that's odd. If the reconstructions don't work, it might be what I wrote in the side note, though. Otherwise I have to give it more thought – matthiaw91 – 2020-06-02T17:54:25.587