def predict_sino8v2(load_step=None, filenames=[], **kwargs):
net = SRSino8(filenames=filenames, **kwargs)
net.build()
with Sinograms2(filenames=filenames) as dataset_train:
with Sinograms2(filenames=filenames, mode='test') as dataset_test:
pt = ProgressTimer(total_step)
for i in range(total_step):
ss = next(dataset_train)
loss_v, _ = net.sess.run([net.loss2x, net.train_2x],
feed_dict={
net.ip: ss[0],
net.ll: ss[1][0],
net.lr: ss[1][1]
})
pt.event(i, msg='loss %f.' % loss_v)
if i % 100 == 0:
summ = net.sess.run(net.summary_op,
feed_dict={
net.ip: ss[0],
net.ll: ss[1][0],
net.lr: ss[1][1]
})
net.sw.add_summary(summ, net.sess.run(net.global_step))
if i % 1000 == 0:
net.save()
net.save()
def train_sino8v3_pet(load_step=None, total_step=None, filenames=[], **kwargs):
print("TRAINGING v3 net on PET data.")
net = SRSino8v3(filenames=filenames, **kwargs)
net.build()
if load_step is not None:
net.load(load_step=load_step)
pre_sum = time.time()
pre_save = time.time()
with SinogramsPETRebin(filenames=filenames) as dataset_train:
with SinogramsPETRebin(filenames=filenames,
mode='test') as dataset_test:
pt = ProgressTimer(total_step)
for i in range(total_step):
ss = next(dataset_train)
loss_v, _ = net.train(ss)
pt.event(i, msg='loss %e.' % loss_v)
now = time.time()
if now - pre_sum > 120:
ss = next(dataset_train)
net.summary(ss, True)
ss = next(dataset_test)
net.summary(ss, False)
pre_sum = now
if now - pre_save > 600:
net.save()
pre_save = now
net.save()
def train(filenames=None, settings=None, **kwargs):
dataset = MNIST(filenames=filenames)
print("=" * 30)
print("DATASET SETTINGS:")
print(dataset.pretty_settings())
print("=" * 30)
net = AAE1D(filenames=filenames)
net.define_net()
print("=" * 30)
print("NETWORK SETTINGS:")
print(net.pretty_settings())
print("=" * 30)
nb_batches = settings['nb_batches']
net.load()
ptp = ProgressTimer(nb_batches)
# for i in range(nb_batches // 3 * 2):
# loss_ae = train_ae(net, dataset)
# msg = 'T:AuE, loss=%05f' % (loss_ae)
# ptp.event(net.step, msg)
# net.lr_decay()
# for i in range(nb_batches // 3):
# loss_ae = train_ae(net, dataset)
# msg = 'T:AuE, loss=%05f' % (loss_ae)
# ptp.event(net.step, msg)
for i in range(nb_batches // 3):
loss_cri = train_cri(net, dataset)
msg = '|T:Cri, loss=%05f|' % (loss_cri)
loss_gen = train_gen(net, dataset)
msg += '|T:Gen, loss=%05f|' % (loss_gen)
loss_ae = train_ae(net, dataset)
msg = '|T:AuE, loss=%05f|' % (loss_ae)
ptp.event(net.step, msg)
ptp = ProgressTimer(nb_batches)
# for i in range(nb_batches // 2):
# loss_ae = train_ae(net, dataset)
# msg = 'step #%5d, AuE, loss=%05f' % (net.step, loss_ae)
# ptp.event(net.step, msg)
# net.lr_decay()
# for i in range(nb_batches // 2):
# loss_ae = train_ae(net, dataset)
# msg = 'step #%5d, AuE, loss=%05f' % (net.step, loss_ae)
# ptp.event(net.step, msg)
# for i in range(nb_batches):
# loss_cri = train_cri(net, dataset)
# msg = 'Cri, loss=%05f ' % (loss_cri)
# loss_gen = train_gen(net, dataset)
# msg += 'Gen, loss=%05f ' % (loss_gen)
# ptp.event(net.step, msg)
net.save('net', is_print=True)
def train2(cfg='train.json', **kwargs):
with open(cfg, 'r') as fin:
cfgs = json.load(fin)
pp_json('TRAIN 2 CONFIGS:', cfgs)
dataset_class = getattr(datasets, cfgs['dataset'])
net_class = getattr(nets2, cfgs['net'])
filenames = cfgs.get('filenames')
load_step = cfgs.get('load_step')
net = net_class(filenames=filenames, **kwargs)
net.build()
if load_step is not None:
net.load(load_step=load_step)
pre_sum = time.time()
pre_save = time.time()
lrs = cfgs['lrs']
total_step = cfgs['total_step']
summary_freq = cfgs['summary_freq']
save_freq = cfgs['save_freq']
with dataset_class(filenames=filenames) as dataset_train:
with dataset_class(filenames=filenames, mode='test') as dataset_test:
pt = ProgressTimer(total_step * len(lrs))
cstep = 0
for lrv in lrs:
net.learning_rate_value = lrv
for i in range(total_step):
ss = next(dataset_train)
loss_v, _ = net.train(ss)
pt.event(cstep, msg='loss %e.' % loss_v)
cstep += 1
now = time.time()
if now - pre_sum > summary_freq:
ss = next(dataset_train)
net.summary(ss, True)
ss = next(dataset_test)
net.summary(ss, False)
pre_sum = now
if now - pre_save > save_freq:
net.save()
pre_save = now
net.save()
def train_sino8v2(load_step=-1,
total_step=None,
step8=1,
step4=1,
step2=1,
sumf=5,
savef=100,
filenames=[],
**kwargs):
click.echo("START TRAINING!!!!")
net = SRSino8v2(filenames='srsino8v2.json', **kwargs)
net.build()
if load_step > 0:
net.load(load_step=load_step)
ds8x_tr = Sinograms2(filenames='sino2_shep8x.json', mode='train')
ds4x_tr = Sinograms2(filenames='sino2_shep4x.json', mode='train')
ds2x_tr = Sinograms2(filenames='sino2_shep2x.json', mode='train')
ds8x_te = Sinograms2(filenames='sino2_shep8x.json', mode='test')
ds4x_te = Sinograms2(filenames='sino2_shep4x.json', mode='test')
ds2x_te = Sinograms2(filenames='sino2_shep2x.json', mode='test')
datasets = [ds8x_tr, ds4x_tr, ds2x_tr, ds8x_te, ds4x_te, ds2x_te]
for ds in datasets:
ds.init()
pt = ProgressTimer(total_step * 3)
cstp = 0
for i in range(total_step // (step8 + step4 + step2)):
for _ in range(step8):
ss = next(ds8x_tr)
loss_v, _ = net.train('net_8x', ss)
pt.event(cstp, msg='train net_8x, loss %e.' % loss_v)
cstp += 1
for _ in range(step4):
ss = next(ds4x_tr)
loss_v, _ = net.train('net_4x', ss)
pt.event(cstp, msg='train net_4x, loss %e.' % loss_v)
cstp += 1
for _ in range(step2):
ss = next(ds2x_tr)
loss_v, _ = net.train('net_2x', ss)
pt.event(cstp, msg='train net_2x, loss %e.' % loss_v)
cstp += 1
if i % sumf == 0:
ss = next(ds8x_tr)
net.summary('net_8x', ss, True)
ss = next(ds4x_tr)
net.summary('net_4x', ss, True)
ss = next(ds2x_tr)
net.summary('net_2x', ss, True)
ss = next(ds8x_te)
net.summary('net_8x', ss, False)
ss = next(ds4x_te)
net.summary('net_4x', ss, False)
ss = next(ds2x_te)
net.summary('net_2x', ss, False)
if i % savef == 0:
net.save()
net.save()
for ds in datasets:
ds.close()
def train_sr_d(dataset_name,
net_name,
epochs,
steps_per_epoch,
load_step=-1,
filenames=[],
**kwargs):
""" train super resolution net
"""
dsc = getattr(xlearn.datasets, dataset_name)
netc = getattr(xlearn.nets, net_name)
print_pretty_args(train_sr_d, locals())
with dsc(filenames=filenames) as dataset:
net_settings = {'filenames': filenames}
if load_step is not None:
net_settings.update({'init_step': load_step})
net = netc(**net_settings)
net.define_net()
cpx, cpy = net.crop_size
if load_step is not None:
if load_step > 0:
click.secho(
'=' * 5 +
'LOAD PRE TRAIN WEIGHTS OF {0:7d} STEPS.'.format(load_step)
+ '=' * 5,
fg='yellow')
net.load(step=load_step, is_force=True)
net.global_step = load_step
click.secho(net.pretty_settings())
pt = ProgressTimer(epochs * steps_per_epoch)
for _ in range(epochs):
for _ in range(steps_per_epoch):
s = next(dataset)
loss = net.train_on_batch(inputs=s[0], outputs=s[1])
msg = "model:{0:5s}, loss={1:10e}, gs={2:7d}.".format(
net._scadule_model(), loss, net.global_step)
pt.event(msg=msg)
net.save(step=net.global_step)
net.dump_loss()
def train(nb_batches, cfs):
dataset = MNIST(filenames=cfs)
print("=" * 30)
print("DATASET SETTINGS:")
print(dataset.pretty_settings())
print("=" * 30)
net = LSGAN(filenames=cfs)
net.define_net()
print("=" * 30)
print("NETWORK SETTINGS:")
print(net.pretty_settings())
print("=" * 30)
ptp = ProgressTimer(net.pre_train)
for i in range(net.pre_train):
s = next(dataset)
z = net.gen_latent()
loss_c = net.train_on_batch('Cri', [s[0], z], [])
msg = 'loss_c= %f' % loss_c
ptp.event(i, msg)
pt = ProgressTimer(nb_batches)
loss_c = np.nan
loss_g = np.nan
for i in range(nb_batches):
s = next(dataset)
z = net.gen_latent()
if i % net.gen_freq > 0:
loss_c = net.train_on_batch('Cri', [s[0], z], [])
msg = 'c_step, loss_c= %f; loss_g= %f' % (loss_c, loss_g)
pt.event(i, msg)
else:
loss_g = net.train_on_batch('WGan', [s[0], z], [])
msg = 'g_step, loss_c= %f; loss_g= %f' % (loss_c, loss_g)
pt.event(i, msg)
if i % 1000 == 0:
net.save('AutoEncoder')
net.save('AutoEncoder', is_print=True)