elif args.dataset == 'binary_mnist':
d = dataset.MNIST(binary=True)
elif args.dataset == 'mnist':
d = dataset.MNIST()
elif args.dataset in ['omniglot', 'omni']:
d = dataset.Omniglot()
elif args.dataset == 'dsprites':
d = dataset.DSprites()
elif args.dataset == "cifar10" or args.dataset == 'cifar':
d = dataset.Cifar10()
if args.per_label is not None:
d.shrink_supervised(args.per_label)
m = model.NoiseModel(d,
config=config,
filename=args.filename,
verbose=args.verbose)
m.lagrangian_fit = args.fit_tf
if args.noise is not None:
if args.noise == 'multiplicative':
m.layers[0]['layer_kwargs']['multiplicative'] = True
if args.noise == 'additive':
m.layers[0]['layer_kwargs']['multiplicative'] = False
if args.beta is not None:
for loss in m.losses:
if not isinstance(
loss,
dict): # reconstruction is a Loss object, so will be skipped
continue
if 'mmd' in loss['type']:
import importlib
import model
import dataset
import layers
import loss_args
import losses
from keras.layers import Dense, Input
import argparse
import tensorflow as tf
parser = argparse.ArgumentParser()
parser.add_argument('--config', type=str, default='test_config.json')
d = dataset.MNIST(binary=True)
#i = Input(shape = (d.x_train.shape[-1],))
#l = Dense(10)(i)
#print(l, callable(l))
#l = Dense(10)
#l = l(i)
#print(l, callable(l))
mi_inits = [1, 2]
for i in range(len(mi_inits)):
mi_init = mi_inits[i]
with tf.name_scope(str("run_" + str(i))):
m = model.NoiseModel(d,
config='test_config.json',
filename=str('mi_' + str(round(mi_init, 2))))
m.layers[0]['layer_kwargs']['init'] = mi_init
m.fit(d.x_train)
else:
args.filename = args.filename + (str(
args.count) if args.count is not None else '') + '_beta_' + str(
args.beta if args.beta is not None else 1) + '_tc_' + str(
args.tc_weight if args.tc_weight is not None else 0) + (
'_remove_' + str(args.remove_prob)
if 'corr' in args.dataset else 'full')
args.filename = os.path.join(args.folder, args.filename)
print("FILENAME ", args.filename)
dd = copy.copy(d)
m = model.NoiseModel(dd,
config=config,
filename=args.filename,
verbose=args.verbose,
seed=args.seed)
ep = m.epochs if args.epochs is None else args.epochs
try:
ep = ep * epoch_mult
except:
pass
m.epochs = 0
m.fit(dd.x_train,
verbose=args.verbose,
validate=args.validate,
fit_gen=args.fit_gen,
epochs=0)