def train_network(parameters,
data,
epochs=1000,
batch_size=32,
loss='mse',
verbose=False,
seed=None,
use_min_perf=False,
callbacks=[],
model_path='',
model_storage=''):
assert (isinstance(parameters, dict)), 'Parameters should be a dictionary.'
clear_session()
set_random_seed(seed)
set_numpy_seed(seed)
t_start = datetime.datetime.now()
input_eular, input_crp, data = get_inputs(data,
rnn_type=parameters['rnn_type'])
[[x1t, x1v], [x2t, x2v], [yt, yv]] = data
model = build_model(input_eular=input_eular,
input_crp=input_crp,
parameters=parameters,
seed=seed)
if verbose: model.summary()
optimizer = get_optimizer(parameters['optimizer'],
parameters['learning_rate'])
model.compile(optimizer=optimizer, loss=loss)
hist = model.fit(x=[x1t, x2t],
y=yt,
batch_size=batch_size,
epochs=epochs,
verbose=verbose,
callbacks=callbacks,
validation_data=([x1v, x2v], yv))
min_perf = min(hist.history['val_loss'])
last_perf = hist.history['val_loss'][-1]
time = datetime.datetime.now() - t_start
time = time.seconds + time.microseconds / 1e6
return last_perf, min_perf, time
cbs = Tr.get_callbacks(plat=True, es=True)
head = ['iteration', 'seed']
head += pg.get_head()
head += ['last_perf', 'min_perf', 'time']
print(head)
log = Bu.CSVWriter(filename, head=head)
model_path = 'current_weights.h5'
while True:
seed = randint(0, 2**32-1)
set_random_seed(seed)
set_numpy_seed(seed)
for i_param, param in enumerate(parameters):
last_perfs = 0
min_perfs = 0
time = 0
for i_cv, cv in enumerate(cvs):
if i_param + i_cv == 0:
model_storage = 'save'
else:
model_storage = 'load'
last_perf, min_perf, dt = Tr.train_network(param, cv, seed=seed, callbacks=cbs, verbose=False, model_path=model_path, model_storage=model_storage)
last_perfs += last_perf
min_perfs += min_perf
time += dt