'weight_decay': 0,
'amsgrad': False
}
loss_params = {'n_critic': 5, 'grad_lambda': 10, 'l1_lambda': (1e4) / 0.05}
paper_opts = adam_opts
paper_opts['betas'] = (0.5, 0.999)
paper_opts['lr'] = 0.0002
range_compress_transform, range_compress_inv_transform = \
create_range_compress_transforms(k_values={"dm": 4, "pressure": 4},
modes={'dm': 'x/(1+x)',
'pressure': 'log'})
transform = chain_transformations([range_compress_transform, atleast_3d])
inv_transform = chain_transformations([squeeze, range_compress_inv_transform])
def Schedule(name,
transform=transform,
inv_transform=inv_transform,
loss_params=loss_params,
paper_opts=paper_opts,
epoch_end=100,
n_test=64):
schedule = {
'type': 'translator',
'transform': transform,
'inv_transform': inv_transform,
folder = os.path.basename(os.path.dirname(__file__))
subfolder = os.path.splitext(os.path.basename(__file__))[0]
name = '/' + folder + '/' + subfolder + '/'
from src.configs.schedules.round_16.stock import Schedule
from src.configs.resnet.dim256x1 import g_structure
from src.configs.patchgan.dim256x2_70_nobn_nosig import d_structure
fc_transform, fc_transform_inv = create_fcs(k_values={
'dm': 2,
'pressure': 4
},
scale=1.75,
shift=-1)
transform = chain_transformations([fc_transform, atleast_3d])
inv_transform = chain_transformations([squeeze, fc_transform_inv])
schedule = Schedule(name)
schedule['sample_interval'] = 100
schedule['batch_size'] = 4
schedule['decay_iter'] = 10
schedule['g_optim_opts']['lr'] = 0.0002
schedule['d_optim_opts']['lr'] = 0.0002
schedule['save_summary']['iters'] = [1] + np.arange(0, 10000, 50).tolist()
schedule['transform'] = transform
schedule['inv_transform'] = inv_transform
# "pressure" : (1,0)},
# modes={"dm":"x/(1+x)",
# "pressure" : "x/(1+x)"},
eps=1e-4)
# range_compress_transform, range_compress_inv_transform = data_transforms.create_range_compress_transforms(
# k_values={"dm" : 1.5,
# "pressure" : 1},
# modes={"dm":"x/(1+x)",
# "pressure" : "1/x"})
with open(os.path.join(data_path, "train_files_info.pickle"), "rb") as f:
training_files_info = pickle.load(f)
transform = data_transforms.chain_transformations([
range_compress_transform,
data_transforms.atleast_3d,
])
inv_transform = data_transforms.chain_transformations([
data_transforms.squeeze,
range_compress_inv_transform,
])
training_dataset = datasets.BAHAMASDataset(files=training_files_info,
root_path=data_path,
redshifts=redshifts,
label_fields=label_fields,
n_stack=n_training_stack,
transform=transform,
inverse_transform=inv_transform,
n_feature_per_field=n_scale,
split_scale_transform, inv_split_scale_transform = \
data_transforms.create_split_scale_transform(n_scale=2,
step_size=8,
include_original=False,
truncate=2.0)
fc_transform, fc_transform_inv = create_fcs(k_values={
'dm': 2,
'pressure': 4
},
scale=1.75,
shift=-1)
transform = data_transforms.chain_transformations([
split_scale_transform,
fc_transform,
data_transforms.atleast_3d,
])
inv_transform = data_transforms.chain_transformations([
data_transforms.squeeze,
fc_transform_inv,
inv_split_scale_transform,
])
schedule = Schedule(name)
schedule['sample_interval'] = 100
schedule['batch_size'] = 4
schedule['decay_iter'] = 10
schedule['g_optim_opts']['lr'] = 0.0002
schedule['d_optim_opts']['lr'] = 0.0002
