$ python3 main.pyNOTE: on Colab Notebook use following command:
!git clone link-to-repo
%run main.pyusage: main.py [-h] [--root_dir ROOT_DIR] [--num_workers NUM_WORKERS] [--batch_size BATCH_SIZE]
[--nz NZ] [--num_epochs NUM_EPOCHS] [--lr LR] [--beta1 BETA1] [--outdir OUTDIR]
optional arguments:
-h, --help show this help message and exit
--root_dir ROOT_DIR path to dataset
--num_workers NUM_WORKERS
number of data loading workers
--batch_size BATCH_SIZE
input batch size
--nz NZ size of the latent z vector
--num_epochs NUM_EPOCHS
number of epochs to train for
--lr LR learning rate, default=0.0002
--beta1 BETA1 beta1 for adam. default=0.5
--outdir OUTDIR folder to output images and model checkpoints
- Title: Conditional Image Synthesis With Auxiliary Classifier GANs
- Authors: Augustus Odena, Christopher Olah, Jonathon Shlens
- Link: https://arxiv.org/abs/1610.09585
- Tags: Neural Networks, Generative Networks, GANs
- Year: 2017
Generative adversarial nets were introduced in 2014 by Goodfellow et al. as a novel way to train a generative model.
They consist of two ‘adversarial’ models: a generative model that captures the data distribution, and a discriminative model
that estimates the probability that a sample came from the training data rather than
.
GANs can produce convincing image samples on datasets with low variability and low resolution. However, they struggle to generate globally coherent, high resolution samples - particularly from datasets with high variability. To improve upon this, the authors introduced Auxiliary Classifier GANs (ACGAN).
In the ACGAN, every generated sample has a corresponding class label, in addition to the noise
.
uses both to generate images
. The discriminator gives both a probability distribution over sources and a probability distribution over the class labels,
. The objective function has two parts: the loglikelihood of the correct source,
, and the log-likelihood of the correct class,
.
is trained to maximize
while
is trained to
maximize
. AC-GANs learn a representation for
that is independent of class label.
The comparision of ACGAN with other GANs is shown below.
In this implementation of ACGAN, we use the CIFAR-10 dataset unlike the ImageNet dataset as used in the paper, due to limited computational resources.
The Generator takes in a dimensional noise vector, and generates images of shape
.
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Linear-1 [-1, 384] 42,624
ConvTranspose2d-2 [-1, 192, 4, 4] 1,179,840
BatchNorm2d-3 [-1, 192, 4, 4] 384
ReLU-4 [-1, 192, 4, 4] 0
ConvTranspose2d-5 [-1, 96, 8, 8] 295,008
BatchNorm2d-6 [-1, 96, 8, 8] 192
ReLU-7 [-1, 96, 8, 8] 0
ConvTranspose2d-8 [-1, 48, 16, 16] 73,776
BatchNorm2d-9 [-1, 48, 16, 16] 96
ReLU-10 [-1, 48, 16, 16] 0
ConvTranspose2d-11 [-1, 3, 32, 32] 2,307
Tanh-12 [-1, 3, 32, 32] 0
================================================================
Total params: 1,594,227
Trainable params: 1,594,227
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.00
Forward/backward pass size (MB): 0.54
Params size (MB): 6.08
Estimated Total Size (MB): 6.62
----------------------------------------------------------------
The Discriminator takes an image as a input and outputs the predicted source (real / fake) and the class of the input image.
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [-1, 16, 16, 16] 448
LeakyReLU-2 [-1, 16, 16, 16] 0
Conv2d-3 [-1, 32, 16, 16] 4,640
BatchNorm2d-4 [-1, 32, 16, 16] 64
LeakyReLU-5 [-1, 32, 16, 16] 0
Dropout-6 [-1, 32, 16, 16] 0
Conv2d-7 [-1, 64, 8, 8] 18,496
BatchNorm2d-8 [-1, 64, 8, 8] 128
LeakyReLU-9 [-1, 64, 8, 8] 0
Dropout-10 [-1, 64, 8, 8] 0
Conv2d-11 [-1, 128, 8, 8] 73,856
BatchNorm2d-12 [-1, 128, 8, 8] 256
LeakyReLU-13 [-1, 128, 8, 8] 0
Dropout-14 [-1, 128, 8, 8] 0
Conv2d-15 [-1, 256, 4, 4] 295,168
BatchNorm2d-16 [-1, 256, 4, 4] 512
LeakyReLU-17 [-1, 256, 4, 4] 0
Dropout-18 [-1, 256, 4, 4] 0
Conv2d-19 [-1, 512, 4, 4] 1,180,160
BatchNorm2d-20 [-1, 512, 4, 4] 1,024
LeakyReLU-21 [-1, 512, 4, 4] 0
Dropout-22 [-1, 512, 4, 4] 0
Linear-23 [-1, 1] 8,193
Sigmoid-24 [-1, 1] 0
Linear-25 [-1, 10] 81,930
================================================================
Total params: 1,664,875
Trainable params: 1,664,875
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.01
Forward/backward pass size (MB): 1.06
Params size (MB): 6.35
Estimated Total Size (MB): 7.43
----------------------------------------------------------------
We use the following loss functions :
- Binary Cross Entropy Loss as the source classification criterion.
- Categorical Cross Entropy Loss as the class classification criterion.
- Discriminator Loss
where
,
- Generator Loss
Here and
refer to source and class respectively.
We use the Adam Optimizer with default learning rate as 0.0002 and as 0.5.
NOTE: All the plots shown below have been smoothened out, the x axis corresponds to number of iterations.
Results after training for 300 epochs.
The image generated for a random noise vector :
Comparing this with the real images from the dataset :
Generator and Discriminator Loss as the training progresses :
Blue - Discriminator Loss ; Red - Generator Loss
The accuracy of classification by the discriminator :
The values ,
and
which denote the output of the discriminator on real and generated images :
Green -
; Grey -
; Orange -
The total time taken on a Tesla K80 GPU to train the model was 97 minutes.
The results may not seem very good since the CIFAR-10 images are only 32*32 px in size and this is what a generator can do at best for low resolution images. On the other hand if we train ACGANs on ImageNet/ other datasets having larger images, one can at once see the difference .