def on_train_finish(**variables):
global plot
if plot is not None:
learning_utils.plot_lines(data=[train_errors, valid_errors], labels=["Training error", "Validation error"],
xlabel="number of epochs",
ylabel=config['loss_type'],
title="training and validation error", suptitle=None, conf=config, additionals=[
[best_train_i, variables['best_train_error']], [best_valid_i, variables['best_valid_error']]],
path=plot)
# joblib.dump(valid_errors, 'test.pkl', compress=1) todo
if config['unit_range']:
unit_iter_train_error.append(variables['avg_train_error'])
unit_iter_valid_error.append(variables['avg_valid_error'])
unit_iter_best_train_error.append(variables['best_train_error'])
unit_iter_best_valid_error.append(variables['best_valid_error'])
def train_custom(conf, plot_path, debug, verbose, gs_params=None, callbacks=default_callbacks):
global config, plot, train_errors, valid_errors
plot = plot_path
train_errors = np.zeros(conf['epochs'])
valid_errors = np.zeros(conf['epochs'])
all_ = list()
for x in ['all', 'md', 'mir', 'feat_sel', 'random']:
all = list()
if x in ('all', 'md', 'mir'):
training_data, training_targets, valid_data, valid_targets = getData(conf['datasets'], 0, None,
None, type=x)
elif x == 'random':
from utils import features
import random
conf['features'] = random.sample(np.hstack(features.values()), random.randint(1, 115))
training_data, training_targets, valid_data, valid_targets = getData(conf['datasets'], 0, conf['features'],
balanced=conf['balanced'], type=x)
else:
training_data, training_targets, valid_data, valid_targets = getData(conf['datasets'], 0, conf['features'],
balanced=conf['balanced'], type=x)
# if there is not enough data available
conf['datasets'] = training_data.shape[0]
if conf['units'] is None:
conf['units'] = [int(math.ceil((training_data.shape[1] + 7) / 2))]
conf['n_input'] = training_data.shape[1]
config = conf
units = [int(math.ceil((training_data.shape[1] + 7) / 2))]
for i in range(1, 101):
net = getNet(units, conf['learning_rate'], conf['epochs'], conf['learning_rule'],
conf['batch_size'], conf['weight_decay'], conf['dropout_rate'],
conf['loss_type'], n_stable=conf['n_stable'], debug=debug, verbose=verbose,
callbacks=callbacks,
# valid_set=(valid_data, valid_targets)
valid_size=conf['ratio']
)
pipeline = learning_utils.getPipeline(training_data, net)
pipeline.fit(training_data, training_targets)
all.append(valid_errors)
all_.append(np.array(all).mean(axis=0))
learning_utils.plot_lines(data=all_, labels=["all", "md", "mir", "feat_sel", "random"],
xlabel="number of epochs",
ylabel=config['loss_type'],
title="mean training and validation error", suptitle=None,
path="learning/nn/plots/comb/test.png")
def train(conf, plot_path, debug, verbose, gs_params=None, callbacks=default_callbacks):
global config, plot, train_errors, valid_errors
plot = plot_path
if 'neural_network__n_iter' in gs_params:
train_errors = np.zeros(20)
valid_errors = np.zeros(20)
else:
train_errors = np.zeros(conf['epochs'])
valid_errors = np.zeros(conf['epochs'])
training_data, training_targets, valid_data, valid_targets, ids = getData(conf['datasets'], conf['ratio'],
conf['features'],
balanced=conf['balanced'],
type=conf['type'], return_ids=True)
# if there is not enough data available
conf['datasets'] = training_data.shape[0]
if conf['units'] is None:
conf['units'] = [int(math.ceil((training_data.shape[1] + 7) / 2))]
conf['n_input'] = training_data.shape[1]
config = conf
if conf['unit_range'] is not None:
del conf['units']
for units in range(conf['unit_range'][0], conf['unit_range'][1] + 1):
net = getNet([units], conf['learning_rate'], conf['epochs'], conf['learning_rule'],
conf['batch_size'], conf['weight_decay'], conf['dropout_rate'],
conf['loss_type'], n_stable=conf['n_stable'], debug=debug, verbose=verbose,
callbacks=callbacks,
valid_size=conf['ratio']
)
pipeline = learning_utils.getPipeline(training_data, net, 'neural_network')
pipeline.fit(training_data, training_targets)
learning_utils.plot_lines(data=[unit_iter_train_error, unit_iter_valid_error],
labels=["Training error", "Validation error"],
xlabel="number of hidden units",
ylabel=config['loss_type'],
title="training and validation error", suptitle=None, conf=config, additionals=[
[np.array(unit_iter_train_error).argmin() + conf['unit_range'][0],
np.array(unit_iter_train_error).min()],
[np.array(unit_iter_valid_error).argmin() + conf['unit_range'][0],
np.array(unit_iter_valid_error).min()]],
begin=conf['unit_range'][0],
path="learning/nn/plots/unit_iter/{}_{}.png".format(conf['unit_range'],
conf['epochs']))
learning_utils.plot_lines(data=[unit_iter_best_train_error, unit_iter_best_valid_error],
labels=["Training error", "Validation error"],
xlabel="number of hidden units",
ylabel=config['loss_type'],
title="training and validation error", suptitle=None, conf=config, additionals=[
[np.array(unit_iter_best_train_error).argmin() + conf['unit_range'][0],
np.array(unit_iter_best_train_error).min()],
[np.array(unit_iter_best_valid_error).argmin() + conf['unit_range'][0],
np.array(unit_iter_best_valid_error).min()]],
begin=conf['unit_range'][0],
path="learning/nn/plots/unit_iter/{}_{}_{}.png".format(conf['unit_range'],
conf['epochs'],
"best"))
else:
net = getNet(conf['units'], conf['learning_rate'], conf['epochs'], conf['learning_rule'],
conf['batch_size'], conf['weight_decay'], conf['dropout_rate'],
conf['loss_type'], n_stable=conf['n_stable'], debug=debug, verbose=verbose,
callbacks=callbacks,
# valid_set=(np.array(valid_data), valid_targets),
valid_size=conf['ratio']
)
pipeline = learning_utils.getPipeline(training_data, net, 'neural_network')
model = pipeline.fit(training_data, training_targets)
model.ids = ids
return model