Table 3 The discriminator and style encoder architectures implemented by IMPA

From: Predicting cell morphological responses to perturbations using generative modeling

Module

Layer

Dimension discr.

Dimension style enc.

Discriminator

3 × 3 2D Conv.

64 × 96 × 96

64 × 96 × 96

 

Down. ResBlock

128 × 48 × 48

128 × 48 × 48

 

Down. ResBlock

256 × 24 × 24

256 × 24 × 24

 

Down. ResBlock

512 × 12 × 12

512 × 12 × 12

 

Down. ResBlock

512 × 6 × 6

512 × 6 × 6

 

Down. ResBlock

512 × 3 × 3

512 × 3 × 3

 

3 × 3 2D Conv.

p × 1 × 1

s × 1 × 1

 

Reshape

p

s

 

Sigmoid

p

-

  1. With p, we indicate the number of perturbation categories, whereas s is the dimensionality of the style space. In the table we assume input spatial dimension of 96, but the same architecture works for larger images as well. The Dimension column refers to the dimensionality of the output of each layer.
Back to article page