cyclegan_semantic_mask.md

Cyclegan with mask labels

All the options and parameters available for cyclegan can be tuned here too.

Training parameters

Parameter	Default	Values	Role
lambda_out_mask	10.0	Positive float	weight for out mask loss
loss_out_mask	L1	L1/MSE/Charbonnier	mask loss
lr_f_s	0.0002	Positive float	f_s learning rate
D_noise	0.0	Positive float	whether to add instance noise to discriminator inputs
rec_noise	0.0	Positive float	whether to add noise to reconstruction
nb_attn	10	Postive integer	number of attention masks (for attention generator)
nb_mask_input	0	Postive integer <= nb_attn	number of attention masks applied on input image (for attention generator)

Training options

Option	Role
out_mask	activate the use of pixel out mask loss
disc_in_mask	activate the use of in-mask discriminator
train_f_s-B	train f_s on domain B (besides domain A)
fs_light	use a lighter network architecture for f_s
D_label_smooth	whether to use one-sided label smoothing with discriminator