if args.pre_model_path is not None:
load_params = tl.files.load_npz(path=args.pre_model_path,
name=args.pre_model_name)
tl.files.assign_params(sess, load_params, network)
print 'set done'
nf.customfit(sess=sess,
network=network,
cost=cost,
train_op=train_op,
tra_provider=tra_provider,
x=x,
y_=y_,
acc=None,
n_epoch=args.epoch,
print_freq=1,
val_provider=val_provider,
save_model=1,
tra_kwag=tra_kwag,
val_kwag=val_kwag,
save_path=args.model_path + args.model_name,
epoch_identifier=None)
if args.ouroboros_e > 0:
nf.Ouroborosfit(sess=sess,
network=network,
cost=cost,
train_op=train_op,
x=x,
y_=y_,
epsilon=1e-08,
use_locking=False).minimize(
cost, var_list=train_params)
# initialize all variables
sess.run(tf.initialize_all_variables())
if args.pre_model_path is not None:
load_params = tl.files.load_npz(path=args.pre_model_path,
name=args.pre_model_name)
tl.files.assign_params(sess, load_params, network)
print 'set done'
nf.customfit(sess=sess,
network=network,
cost=cost,
train_op=train_op,
tra_provider=tra_provider,
x=x,
y_=y_,
acc=None,
n_epoch=args.epoch,
print_freq=1,
val_provider=val_provider,
save_model=1,
tra_kwag=tra_kwag,
val_kwag=val_kwag,
save_path=args.model_path + args.model_name,
epoch_identifier=None)
test_val = {
'inputs':test_set_y,
'targets':test_set_y,
'special': True if downscale == 'mix' else False,
'keepdims':True}
if args.downscale == 2:
from up2 import zipper
if args.downscale == 4 or 'mix':
from up4 import zipper
if args.downscale == 10:
from up10 import zipper
x = tf.placeholder(tf.float32, shape=[None, observations, args.input_x, args.input_y], name='x')
network = zipper(x, downscale, args.input_x, args.input_y, is_train=True, observation=args.observations)
if args.pre_model_path is not None:
load_params = tl.files.load_npz(path=args.pre_model_path, name=args.pre_model_name)
tl.files.assign_params(sess, load_params, network)
print 'set done'
prediction = nf.custompredict(sess=sess, network=network, output_provider=test_provider, x=x, fragment_size=1,
output_length=6400, y_op=None, out_kwag=test_val)
prediction = prediction[0]*args.std+args.mean
np.save(args.prediction_file, prediction)
for e in range(joint_epoch):
print 'Joint epoch', e + 1
if e == 0:
d_epoch = args.d_pretrain
else:
d_epoch = args.d_step
nf.customfit(sess=sess,
network=D_generate,
cost=D_lost,
train_op=D_op,
tra_provider=tra_provider,
x=g_x,
y_=d_x,
acc=None,
n_epoch=d_epoch,
print_freq=1,
val_provider=None,
save_model=args.save_d,
tra_kwag=tra_kwag,
val_kwag=None,
save_path=args.save_model_path + args.save_d_name,
epoch_identifier=None)
nf.customfit(sess=sess,
network=G,
cost=G_lost,
train_op=G_op,
tra_provider=tra_provider,
x=g_x,
y_=g_y,
'inputs': test_set[:, :,
args.start_t - args.observations - 1:args.start_t],
'framebatch': framebatch,
'mean': args.mean,
'std': args.std,
'norm_tar': True
}
dstn_predictions = nf.out_futurepredictor(sess=sess,
network=network,
network2=network_ots,
timestamp=args.period_t,
season=history,
output_provider=test_provider,
x=x,
mean=args.mean,
std=args.std,
fragment_size=args.fragment_size,
output_length=1,
y_op=None,
out_kwag=test_kwag,
frameshape=frameshape,
future=args.step,
w=args.w,
bias=args.b,
low_weight=args.delta,
weight_decay=1.0 / args.observations)
prediction = dstn_predictions[0].reshape(-1, frameshape[0],
frameshape[1]) * args.std + args.mean
np.save(args.save_file, prediction)