Table 1 Generator network architecture: layer types, activations, output shapes (channels × height × width) and number of trainable parameters for each layer. TransposedConv have strides =2

From: CosmoGAN: creating high-fidelity weak lensing convergence maps using Generative Adversarial Networks

	Activ.	Output shape	Params.
Latent	–	64	–
Dense	–	512 × 16 × 16	8.5M
BatchNorm	ReLU	512 × 16 × 16	1024
TransposedConv 5 × 5	–	256 × 32 × 32	3.3M
BatchNorm	ReLU	256 × 32 × 32	512
TransposedConv 5 × 5	–	128 × 64 × 64	819K
BatchNorm	ReLU	128 × 64 × 64	256
TransposedConv 5 × 5	–	64 × 128 × 128	205K
BatchNorm	ReLU	64 × 128 × 128	128
TransposedConv 5 × 5	Tanh	1 × 256 × 256	1601
Total trainable parameters			12.3M

Back to article page