def get_config():
config = get_default_configs()
# training
training = config.training
training.sde = 'vpsde'
training.continuous = True
training.reduce_mean = True
# sampling
sampling = config.sampling
sampling.method = 'pc'
sampling.predictor = 'euler_maruyama'
sampling.corrector = 'none'
# data
data = config.data
data.centered = True
# model
model = config.model
model.name = 'ddpm'
model.scale_by_sigma = False
model.ema_rate = 0.9999
model.normalization = 'GroupNorm'
model.nonlinearity = 'swish'
model.nf = 128
model.ch_mult = (1, 2, 2, 2)
model.num_res_blocks = 2
model.attn_resolutions = (16, )
model.resamp_with_conv = True
model.conditional = True
return config
def get_config():
config = get_default_configs()
# training
training = config.training
training.sde = 'vesde'
training.continuous = False
# sampling
sampling = config.sampling
sampling.method = 'pc'
sampling.predictor = 'reverse_diffusion'
sampling.corrector = 'langevin'
# model
model = config.model
model.name = 'ddpm'
model.scale_by_sigma = True
model.ema_rate = 0.999
model.normalization = 'GroupNorm'
model.nonlinearity = 'swish'
model.nf = 128
model.ch_mult = (1, 2, 2, 2)
model.num_res_blocks = 2
model.attn_resolutions = (16, )
model.resamp_with_conv = True
model.conditional = True
model.conv_size = 3
return config
def get_config(): config = get_default_configs() # training training = config.training training.sde = 'vesde' training.continuous = False # sampling sampling = config.sampling sampling.method = 'pc' sampling.predictor = 'none' sampling.corrector = 'ald' sampling.n_steps_each = 5 sampling.snr = 0.176 # model model = config.model model.name = 'ncsn' model.scale_by_sigma = False model.num_scales = 232 model.ema_rate = 0. model.normalization = 'InstanceNorm++' model.nonlinearity = 'elu' model.nf = 128 model.interpolation = 'bilinear' # optim optim = config.optim optim.weight_decay = 0 optim.optimizer = 'Adam' optim.lr = 1e-3 optim.beta1 = 0.9 optim.amsgrad = False optim.eps = 1e-8 optim.warmup = 0 optim.grad_clip = -1. return config
def get_config():
config = get_default_configs()
# training
training = config.training
training.sde = 'subvpsde'
training.continuous = True
training.n_iters = 950001
training.reduce_mean = True
# sampling
sampling = config.sampling
sampling.method = 'pc'
sampling.predictor = 'euler_maruyama'
sampling.corrector = 'none'
# data
data = config.data
data.centered = True
# model
model = config.model
model.name = 'ncsnpp'
model.fourier_scale = 16
model.scale_by_sigma = False
model.ema_rate = 0.9999
model.normalization = 'GroupNorm'
model.nonlinearity = 'swish'
model.nf = 128
model.ch_mult = (1, 2, 2, 2)
model.num_res_blocks = 8
model.attn_resolutions = (16, )
model.resamp_with_conv = True
model.conditional = True
model.fir = True
model.fir_kernel = [1, 3, 3, 1]
model.skip_rescale = True
model.resblock_type = 'biggan'
model.progressive = 'none'
model.progressive_input = 'residual'
model.progressive_combine = 'sum'
model.attention_type = 'ddpm'
model.embedding_type = 'positional'
model.init_scale = 0.0
model.conv_size = 3
return config
def get_config():
config = get_default_configs()
# training
training = config.training
training.sde = 'vesde'
training.continuous = True
# sampling
sampling = config.sampling
sampling.method = 'pc'
sampling.predictor = 'reverse_diffusion'
sampling.corrector = 'langevin'
# model
model = config.model
model.name = 'ncsnpp'
model.scale_by_sigma = True
model.ema_rate = 0.999
model.normalization = 'GroupNorm'
model.nonlinearity = 'swish'
model.nf = 128
model.ch_mult = (1, 2, 2, 2)
model.num_res_blocks = 4
model.attn_resolutions = (16, )
model.resamp_with_conv = True
model.conditional = True
model.fir = True
model.fir_kernel = [1, 3, 3, 1]
model.skip_rescale = True
model.resblock_type = 'biggan'
model.progressive = 'none'
model.progressive_input = 'residual'
model.progressive_combine = 'sum'
model.attention_type = 'ddpm'
model.init_scale = 0.
model.fourier_scale = 16
model.conv_size = 3
return config
