Programming assignments and quizzes from all courses in the Coursera Generative adversarial specialization offered by deeplearning.ai
.
-
Generative model vs Discrimentative model
-
Activation and Common activations
-
Batch, padding, pooling, upsampling, convolution and transposed convolution operations.
-
Collapse mode, the generative model will produce sample of only one class.
-
To solve the collapse mode problem new loss functions has been announces and we change BCE loss function because of vanishing gradient descent problems
-
User wasserstein loss, aim to minimize the distance between generated and real AI.
-
Critic vs discriminators, in critics we have different values and there are some limits on critis.
-
The first limit is that the gradients be at most 1 : wieght clipping, gradient penalty
-
Conditional generation : classes we want + input matrix + noise
-
Controabble : features we want + input matrix + noise
-
Challenges: Entanglement and correlation between features.
-
To solve the challenges 1)classifier gradient: to use a classifier module and 2)disentaglement.
-
Evaluate GANs because discriminators are overfitting to your generative models and we don't have a universal discriminator.
-
Fidelity : quality or realism of image vs Diversity : variety of images
-
Comparing image : Pixel distance (isnt reliable), Feature distance (less sensetive to shifts)
-
FID : A lower FID score indicates that the generated images are closer to the real images
-
Inception score
-
Sampling and Truncation
-
Precision and Recall
-
Disadvantages : lack of intrinsic evaluation metric
-
Alternatives: VAE, Autoregressive Models, Flow Models
-
Bias and Fairness in Model
-
Adversial loss for solving bias
-
StyleGAN : progressive growing (avoid mode collapse and create high resolution images), noise mapping network (variation and randomness), adaptive instance normalization (control the style of generated images according to desired style).
-
Gan usage in Data augmentation, image to image translation, medicine and climate change, text to image, adverserial area.
-
Paried image to image translation using Pix2Pix which transfer style of one image to another image.
-
PatchGAN and U-Net architecture and pixel distance loss term.
-
CycleGAN and Cycle consistency
-
Least square and identity loss (input and output are same) which helps to preserve the color.