def test_networks():
# testbed をつくる
# bed = TestBed("../data/lane.180000.3.xml", r=2, d=1)
bed = TestBed("../../data/cross3ltl_full_3/lane.129600.1.xml",
r=2,
d=1,
debug_level=0)
n_input = bed.get_n_input()
n_output = bed.get_n_outupt()
# 実験用のネットワークを作る
exp1 = theanets.Experiment(theanets.feedforward.Regressor,
layers=(n_input,
dict(size=100,
activation='linear'), n_output),
optimize='sgd',
activation='linear')
exp2 = theanets.Experiment(theanets.feedforward.Regressor,
layers=(n_input,
dict(size=100, activation='linear'),
dict(size=100,
activation='linear'), n_output),
optimize='sgd',
activation='linear')
exp3 = theanets.Experiment(theanets.feedforward.Regressor,
layers=(n_input,
dict(size=100, activation='linear'),
dict(size=100, activation='linear'),
dict(size=100,
activation='linear'), n_output),
optimize='sgd',
activation='linear')
# exp4 = theanets.Experiment(
# theanets.feedforward.Regressor,
# layers=(
# n_input,
# dict(size=400, activation='linear'),
# dict(size=400, activation='linear'),
# dict(size=400, activation='linear'),
# n_output
# ),
# optimize='sgd',
# activation='linear'
# )
# 実験する
bed.test(exp1, 20)
bed.test(exp2, 20)
bed.test(exp3, 20, block=True)
def do_regression():
climate.enable_default_logging()
train, test = create_datasets()
x_train = train[:, 0:6]
y_train = train[:, 6]
y_train = np.reshape(y_train, (y_train.shape[0], 1))
y_test = test[:, 6]
y_test = np.reshape(y_test, (y_test.shape[0], 1))
exp = theanets.Experiment(theanets.Regressor, layers=(6, 6, 1))
exp.train([x_train, y_train])
#do the testing
x_test = test[:, 0:6]
y_test = test[:, 6]
yp = exp.network.predict(x_test)
xi = [(i + 1) for i in range(x_test.shape[0])]
pb.scatter(xi, y_test, color="red")
pb.scatter(xi, yp, color="blue")
pb.show()
return
def train_strategy(stock, ratio=0.8, min_improvement=0.001):
train, valid = load_dataset(stock)
n, n_input = train[0].shape
exp = theanets.Experiment(
theanets.Classifier,
layers=(n_input, n_input * 2, 2),
)
exp.train(train,
valid,
min_improvement=min_improvement,
algo='sgd',
learning_rate=0.01,
momentum=0.5,
hidden_l1=0.001,
weight_l2=0.001,
num_updates=100)
print('training:')
evaluate(exp, train)
print('validation:')
evaluate(exp, valid)
exp.save('%s.nn' % stock)
return exp
def partial_fit(self, X, y, sample_weight=None, keep_trainer=True, **trainer):
"""
Train the regressor by training the existing regressor again.
:param pandas.DataFrame X: data shape [n_samples, n_features]
:param y: values - array-like of shape [n_samples] (or [n_samples, n_targets])
:param bool keep_trainer: True if the trainer is not stored in self.trainers
:param dict trainer: parameters of the training algorithm we want to use now
:return: self
.. note:: if `trainer['optimize'] == 'pretrain'` (unsupervised training)
`y` can be any vector just with information about number of targets `numpy.shape(y)`
"""
allow_multiple_targets = False if len(numpy.shape(y)) == 1 else True
X, y = self._prepare_for_partial_fit(X, y, allow_multiple_targets=allow_multiple_targets,
keep_trainer=keep_trainer, **trainer)
if self.exp is None:
layers = self._construct_layers(X.shape[1], 1 if len(numpy.shape(y)) == 1 else numpy.shape(y)[1])
self.exp = tnt.Experiment(tnt.Regressor, layers=layers,
rng=self._reproducibilize(), **self._prepare_network_params())
self._reproducibilize()
if len(numpy.shape(y)) == 1:
y = y.reshape(len(y), 1)
if trainer.get('optimize') == 'pretrain':
self.exp.train([X.astype(numpy.float32)], **trainer)
else:
self.exp.train([X.astype(numpy.float32), y], **trainer)
return self
def partial_fit(self, X, y, keep_trainer=True, **trainer):
"""
Train the classifier by training the existing classifier again.
:param pandas.DataFrame X: data shape [n_samples, n_features]
:param y: values - array-like of shape [n_samples]
:param bool keep_trainer: True if the trainer is not stored in self.trainers
:param dict trainer: parameters of the training algorithm we want to use now
:return: self
.. note:: if `trainer['optimize'] == 'pretrain'` (unsupervised training)
`y` can be any vector just with information `numpy.unique(y) == classes`
"""
X, y = self._prepare_for_partial_fit(X, y, keep_trainer=keep_trainer, **trainer)
if self.exp is None:
self._set_classes(y)
layers = self._construct_layers(X.shape[1], len(self.classes_))
self.exp = tnt.Experiment(tnt.Classifier, layers=layers,
rng=self._reproducibilize(), **self._prepare_network_params())
self._reproducibilize()
if trainer.get('optimize', None) == 'pretrain':
self.exp.train([X.astype(numpy.float32)], **trainer)
else:
self.exp.train([X.astype(numpy.float32), y.astype(numpy.int32)], **trainer)
return self
def partial_fit(self,
X,
y,
sample_weight=None,
keep_trainer=True,
**trainer):
X, y, sample_weight = self._prepare_for_partial_fit(
X,
y,
sample_weight=sample_weight,
keep_trainer=keep_trainer,
**trainer)
if self.exp is None:
self._set_classes(y)
layers = self._construct_layers(X.shape[1], len(self.classes_))
self.exp = tnt.Experiment(tnt.Classifier,
layers=layers,
weighted=True)
params = self._prepare_network_params()
params.update(**trainer)
if trainer.get('algo', None) == 'pretrain':
self.exp.train([X.astype(numpy.float32)], **params)
else:
self.exp.train([
X.astype(numpy.float32),
y.astype(numpy.int32),
sample_weight.astype(numpy.float32)
], **params)
return self
def main(args):
train, valid, _ = load_cifar()
whiten, color = pca(train)
feat = args.features or int(np.sqrt(4 * K))
e = theanets.Experiment(
theanets.Autoencoder,
layers=(K, feat**2, K),
)
e.train(whiten(train), whiten(valid), input_noise=1)
plot_layers([
color(e.network.find(1, 0).get_value().T).T,
color(e.network.find('out', 0).get_value())
],
channels=3)
plt.tight_layout()
plt.show()
valid = whiten(valid[:100])
plot_images(color(valid), 121, 'Sample data', channels=3)
plot_images(color(e.network.predict(valid)),
122,
'Reconstructed data',
channels=3)
plt.tight_layout()
plt.show()
def do_classification():
climate.enable_default_logging()
df = pd.read_csv("glass.csv")
#recode the class variable to go from 0 through 5
df["GT"] = map(recode, df["GT"])
train, valid = train_test_split(df, test_size=0.3)
train_X = train.ix[:, 0:9].as_matrix()
train_Y = train.ix[:, 9].as_matrix()
valid_X = valid.ix[:, 0:9].as_matrix()
valid_Y = valid.ix[:, 9].as_matrix()
train_X = train_X.astype('f')
train_Y = train_Y.astype('i')
valid_X = valid_X.astype('f')
valid_Y = valid_Y.astype('i')
t0 = (train_X, train_Y)
t1 = (valid_X, valid_Y)
exp = theanets.Experiment(theanets.Classifier, layers=(9, 18, 18, 6))
exp.train(t0, t1, algorithm='sgd',\
learning_rate=1e-4, momentum=0.1,\
hidden_l1=0.001, weight_l2=0.001)
cm = confusion_matrix(valid_Y, exp.network.predict(valid_X))
return cm
def test_sgd(self):
self.exp = theanets.Experiment(theanets.recurrent.Classifier,
layers=(self.NUM_INPUTS, (10, 'rnn'),
self.NUM_CLASSES),
weighted=True)
self.assert_progress('sgd',
[self.INPUTS, self.CLASSES, self.CLASS_WEIGHTS])
def test_sgd(self):
self.exp = theanets.Experiment(theanets.recurrent.Regressor,
layers=(self.NUM_INPUTS, (10, 'rnn'),
self.NUM_OUTPUTS),
weighted=True)
self.assert_progress('sgd',
[self.INPUTS, self.OUTPUTS, self.OUTPUT_WEIGHTS])
def theanets_load(key, theanonets_type):
file_name_network = os.path.join(c["CACHE_DIR"], key + ".network.pkl")
file_name_network_structure = os.path.join(c["CACHE_DIR"], key + ".network.structure.pkl")
e = None
with open(file_name_network_structure, "r") as f:
e = theanets.Experiment(theanonets_type, **pickle.load(f))
e.load(file_name_network)
return e
def main():
training_data, validation_data, test_data, std_scale = load_training_data()
climate.enable_default_logging()
targets = ['esgd','layerwise','rmsprop','nag','rprop','sgd','sample','adadelta']
layers = [(93, dict(size=512, sparsity=0.2, activation='relu'),
dict(size=512, sparsity=0.2, activation='relu'),
dict(size=512, sparsity=0.2, activation='relu'),
9)]
for l in layers:
for t in targets:
exp = theanets.Experiment(
theanets.Classifier,
layers=l,
weighted=True,
output_activation='softmax'
)
exp.train(training_data,
validation_data,
optimize=t,
)
exp.train(training_data,
validation_data,
optimize=t,
)
exp.train(training_data,
validation_data,
optimize=t,
)
#get an prediction of the accuracy from the test_data
test_results = exp.network.predict(test_data[0])
loss = multiclass_log_loss(test_data[1], test_results)
print 'Test multiclass log loss:', loss
out_file = 'results/' + str(loss) + t
exp.save(out_file + '.pkl')
#save the kaggle results
kaggle_test_features = load_test_data(std_scale)
results = exp.network.predict(kaggle_test_features)
save_results(out_file + '.csv', kaggle_test_features, results)
def load_strategy(name, verbose=False):
print("loading %s trained strategy" % name)
train, valid = load_dataset(name)
n, n_input = train[0].shape
exp = theanets.Experiment(
theanets.Classifier,
layers=(n_input, n_input * 2, 2),
)
exp.load('%s.nn' % name)
if verbose:
print('training:')
evaluate(exp, train)
print('validation:')
evaluate(exp, valid)
return exp
def test_save_load(self):
exp = theanets.Experiment(theanets.Autoencoder, layers=(10, 3, 4, 10))
net = exp.network
f, p = tempfile.mkstemp(suffix='pkl')
os.close(f)
os.unlink(p)
try:
exp.save(p)
assert os.path.isfile(p)
exp.load(p)
assert exp.network is not net
assert exp.network.layers == (10, 3, 4, 10)
finally:
if os.path.exists(p):
os.unlink(p)
def main(args):
train, valid, _ = load_mnist()
e = theanets.Experiment(theanets.Autoencoder,
layers=(784, args.features**2, 784))
e.train(train, valid)
plot_layers([e.network.find(1, 0), e.network.find(2, 0)])
plt.tight_layout()
plt.show()
v = valid[:100]
plot_images(v, 121, 'Sample data')
plot_images(e.network.predict(v), 122, 'Reconstructed data')
plt.tight_layout()
plt.show()
def test_save_load(self):
exp = theanets.Experiment(theanets.Autoencoder, layers=(10, 3, 4, 10))
net = exp.network
f, p = tempfile.mkstemp(suffix='pkl')
os.close(f)
os.unlink(p)
try:
exp.save(p)
assert os.path.isfile(p)
exp.load(p)
assert exp.network is not net
for lo, ln in zip(net.layers, exp.network.layers):
assert lo.name == ln.name
assert lo.inputs == ln.inputs
assert lo.size == ln.size
finally:
if os.path.exists(p):
os.unlink(p)
def __setstate__(self, dictionary):
"""
Required for pickle.load working, because theanets objects can't be unpickled by default.
:param dict dictionary: the structure representing a TheanetsClassifier
"""
self.__dict__ = dictionary
if dictionary['dumped_exp'] is None:
self.exp = None
else:
with tempfile.NamedTemporaryFile() as dump:
with open(dump.name, 'wb') as dumpfile:
dumpfile.write(dictionary['dumped_exp'])
assert os.path.exists(dump.name), 'there is no such file: {}'.format(dump.name)
layers = [1] + self.layers + [1]
self.exp = tnt.Experiment(tnt.Classifier, layers=layers, rng=self._reproducibilize(),
**self.network_params)
self.exp.load(dump.name)
del dictionary['dumped_exp']
def partial_fit(self, X, y, sample_weight=None, new_trainer=True, **trainer):
"""
Train the regressor by training the existing regressor again.
:param pandas.DataFrame X: data shape [n_samples, n_features]
:param y: values - array-like of shape [n_samples]
:param bool new_trainer: True if the trainer is not stored in self.trainers
:param dict trainer: parameters of the training algorithm we want to use now
:return: self
"""
X, y = self._prepare_for_partial_fit(X, y, new_trainer, **trainer)
if self.exp is None:
layers = self._construct_layers(X.shape[1], 1)
print(layers)
self.exp = tnt.Experiment(tnt.feedforward.Regressor, layers=layers,
rng=self._reproducibilize(), **self.network_params)
self._reproducibilize()
self.exp.train([X.astype(numpy.float32), y.reshape(len(y), 1)], **trainer)
return self
def main(args):
train, valid, _ = load_cifar()
e = theanets.Experiment(theanets.Autoencoder,
layers=(3072, args.features**2, 3072))
e.train(train, valid)
plot_layers(e.network.weights, channels=3)
plt.tight_layout()
plt.show()
valid = valid[:100]
plot_images(valid, 121, 'Sample data', channels=3)
plot_images(e.network.predict(valid),
122,
'Reconstructed data',
channels=3)
plt.tight_layout()
plt.show()
def test_save_load(self):
exp = theanets.Experiment(theanets.Autoencoder, layers=(10, 3, 4, 10))
net = exp.network
f, p = tempfile.mkstemp(suffix='pkl')
os.close(f)
os.unlink(p)
try:
exp.save(p)
assert os.path.isfile(p)
exp.load(p)
assert exp.network is not net
for lo, ln in zip(net.layers, exp.network.layers):
assert lo.name == ln.name
assert lo._input_shapes == ln._input_shapes
for po, pn in zip(net.params, exp.network.params):
assert po.name == pn.name
assert np.allclose(po.get_value(), pn.get_value())
finally:
if os.path.exists(p):
os.unlink(p)
def __setstate__(self, dictionary):
"""
Required for pickle.load working, because theanets objects can't be unpickled by default.
:param dict dictionary: the structure representing a TheanetsClassifier or TheanetsRegressor
"""
dumped_exp = dictionary.pop('dumped_exp')
self.__dict__ = dictionary.copy()
if dumped_exp is None:
self.exp = None
else:
with tempfile.NamedTemporaryFile() as dump:
with open(dump.name, 'wb') as dumpfile:
dumpfile.write(dumped_exp)
assert os.path.exists(dump.name), 'something strange in unpickling: there is no such file: {}'.format(
dump.name)
dummy_layers = [1] + self.layers + [1]
estimator_object = tnt.Classifier if self._model_type == 'classification' else tnt.Regressor
self.exp = tnt.Experiment(estimator_object, layers=dummy_layers, rng=self._reproducibilize(),
**self._prepare_network_params())
self.exp.load(dump.name)
def partial_fit(self, X, y, new_trainer=True, **trainer):
"""
Train the classifier by training the existing classifier again.
:param pandas.DataFrame X: data shape [n_samples, n_features]
:param y: values - array-like of shape [n_samples]
:param bool new_trainer: True if the trainer is not stored in self.trainers
:param dict trainer: parameters of the training algorithm we want to use now
:return: self
"""
X, y = self._prepare_for_partial_fit(X, y, new_trainer, **trainer)
self.classes_ = numpy.unique(y)
if self.exp is None:
layers = self._construct_layers(X.shape[1], len(self.classes_))
print(layers)
self.exp = tnt.Experiment(tnt.Classifier, layers=layers,
rng=self._reproducibilize(), **self.network_params)
self._reproducibilize()
self.exp.train((X.astype(numpy.float32), y.astype(numpy.int32)),
**trainer)
return self
def test_networks():
# experiment をつくる
bed = Experiment(r=2, d=1, debug_level=2)
# データ準備
bed.setTrainData("../data/cross3ltl_full_3/lane.129600.1.xml")
bed.setTestData("../data/cross3ltl_full_3/lane.129600.2.xml")
# ネットワーク作成準備
n_input = bed.get_n_input()
n_output = bed.get_n_outupt()
# 実験用のネットワークを作る
exp1 = theanets.Experiment(
theanets.feedforward.Regressor,
layers=(
n_input,
dict(size=120, activation='relu'),
dict(size=120, activation='relu'),
dict(size=120, activation='relu'),
n_output
),
optimize='sgd',
activation='sigmoid'
)
# 事前学習
bed.pretrain(exp1)
# 学習と評価
bed.train(exp1, 1)
bed.test(exp1, False)
bed.train(exp1, 1)
bed.test(exp1, False)
bed.train(exp1, 1)
bed.test(exp1, True)
def train(self):
architecture = self.__architecture
dataset = self.__dataset
cut = int(0.9 * len(dataset)
) # select 90% of data for training, 10% for validation
idx = range(len(dataset))
np.random.shuffle(idx)
train = idx[:cut]
train_set = [dataset[train, :-1], dataset[train, -1:]]
valid = idx[cut:]
valid_set = [dataset[valid, :-1], dataset[valid, -1:]]
e = theanets.Experiment(theanets.feedforward.Regressor,
layers=architecture,
optimize='sgd',
hidden_activation='tanh',
output_activation='linear',
learning_rate=0.01)
e.train(train_set, valid_set)
self.e = e
def main(args):
# load up the MNIST digit dataset.
train, valid, _ = load_mnist()
e = theanets.Experiment(theanets.Autoencoder,
layers=(784, args.features**2, 784))
e.train(train,
valid,
input_noise=0.1,
weight_l2=0.0001,
algorithm='rmsprop',
momentum=0.9,
min_improvement=0.1)
plot_layers([e.network.find('hid1', 'w'), e.network.find('out', 'w')])
plt.tight_layout()
plt.show()
v = valid[:100]
plot_images(v, 121, 'Sample data')
plot_images(e.network.predict(v), 122, 'Reconstructed data')
plt.tight_layout()
plt.show()
def partial_fit(self,
X,
y,
sample_weight=None,
keep_trainer=True,
**trainer):
allow_multiple_targets = False if len(numpy.shape(y)) == 1 else True
X, y, sample_weight = self._prepare_for_partial_fit(
X,
y,
sample_weight=sample_weight,
allow_multiple_targets=allow_multiple_targets,
keep_trainer=keep_trainer,
**trainer)
if self.exp is None:
layers = self._construct_layers(
X.shape[1],
1 if len(numpy.shape(y)) == 1 else numpy.shape(y)[1])
self.exp = tnt.Experiment(tnt.Regressor,
layers=layers,
weighted=True)
params = self._prepare_network_params()
params.update(**trainer)
if len(numpy.shape(y)) == 1:
y = y.reshape(len(y), 1)
if len(numpy.shape(sample_weight)) == 1:
sample_weight = numpy.repeat(sample_weight, y.shape[1])
sample_weight = sample_weight.reshape(y.shape)
if trainer.get('algo') == 'pretrain':
self.exp.train([X.astype(numpy.float32)], **params)
else:
self.exp.train([
X.astype(numpy.float32), y,
sample_weight.astype(numpy.float32)
], **params)
return self
return np.dot(x, np.dot(vecs, np.diag(1. / vals)))
def color(z):
return np.dot(z, np.dot(np.diag(vals), vecs.T))
# now train our model on the whitened dataset.
N = 16
e = theanets.Experiment(
RICA,
layers=(K, N * N, K),
activation='linear',
hidden_l1=0.2,
no_learn_biases=True,
tied_weights=True,
train_batches=100,
weight_inverse=0.01,
)
e.run(whiten(train), whiten(valid))
# color the network weights so they are viewable as digits.
plot_layers([color(e.network.weights[0].get_value().T).T], tied_weights=True)
plt.tight_layout()
plt.show()
plot_images(valid[:N * N], 121, 'Sample data')
plot_images(color(e.network.predict(whiten(valid[:N * N]))), 122,
'Reconstructed data')
plt.tight_layout()
import logging
import matplotlib.pyplot as plt
import numpy as np
import theanets
climate.enable_default_logging()
TIME = 10
BITS = 3
BATCH_SIZE = 32
mask = np.ones((TIME, BATCH_SIZE, 1), bool)
mask[:TIME - BITS] = 0
e = theanets.Experiment(theanets.recurrent.Regressor,
layers=(1, ('rnn', 10), 1),
weighted=True)
def generate():
s, t = np.random.randn(2, TIME, BATCH_SIZE, 1).astype('f')
s[:BITS] = t[-BITS:] = np.random.randn(BITS, BATCH_SIZE, 1)
return s, t, mask
src, tgt, msk = generate()
logging.info('data batches: %s -> %s @ %s', src.shape, tgt.shape, msk.shape)
e.train(generate, batch_size=BATCH_SIZE)
predict = e.network.predict(src)[:, :, 0]
#!/usr/bin/env python
import matplotlib.pyplot as plt
import theanets
from utils import load_mnist, plot_layers, plot_images
train, valid, _ = load_mnist()
e = theanets.Experiment(
theanets.Autoencoder,
layers=(784, 256, 100, 64, ('tied', 100), ('tied', 256), ('tied', 784)),
)
e.train(train, valid,
algorithm='layerwise',
patience=1,
min_improvement=0.05,
train_batches=100)
e.train(train, valid, min_improvment=0.01, train_batches=100)
plot_layers([e.network.find(i, 'w') for i in (1, 2, 3)], tied_weights=True)
plt.tight_layout()
plt.show()
valid = valid[:16*16]
plot_images(valid, 121, 'Sample data')
plot_images(e.network.predict(valid), 122, 'Reconstructed data')
plt.tight_layout()
plt.show()
def train_nn(speech_data, speech_alignment):
vta = MLFMFCCOnlineAlignedArray(usec0=usec0,
n_last_frames=last_frames,
usedelta=usedelta,
useacc=useacc,
mel_banks_only=mel_banks_only)
sil_count = 0
speech_count = 0
for sd, sa in zip(speech_data, speech_alignment):
mlf_speech = load_mlf(sa, max_files, max_frames_per_segment)
vta.append_mlf(mlf_speech)
vta.append_trn(sd)
sil_count += mlf_speech.count_length('sil')
speech_count += mlf_speech.count_length('speech')
print "The length of sil segments: ", sil_count
print "The length of speech segments: ", speech_count
mfcc = vta.__iter__().next()
print "MFCC length:", len(mfcc[0])
input_size = len(mfcc[0])
e = theanets.Experiment(
theanets.Classifier,
layers=(input_size, hidden_units, hidden_units, hidden_units,
hidden_units, 2),
optimize="hf",
num_updates=30,
validate=1,
initial_lambda=0.1,
preconditioner=True if preconditioner else False,
hidden_dropouts=hidden_dropouts,
weight_l2=weight_l2,
batch_size=500,
)
random.seed(0)
print "Generating the cross-validation and train MFCC features"
crossvalid_x = []
crossvalid_y = []
train_x = []
train_y = []
i = 0
for frame, label in vta:
frame = frame - (10.0 if mel_banks_only else 0.0)
if i % (max_frames / 10) == 0:
print "Already processed: %.2f%% of data" % (100.0 * i /
max_frames)
if i > max_frames:
break
if random.random() < float(crossvalid_frames) / max_frames:
# sample validation (test) data
crossvalid_x.append(frame)
if label == "sil":
crossvalid_y.append(0)
else:
crossvalid_y.append(1)
else:
train_x.append(frame)
if label == "sil":
train_y.append(0)
else:
train_y.append(1)
i += 1
crossvalid = [
np.array(crossvalid_x),
np.array(crossvalid_y).astype('int32')
]
train = [np.array(train_x), np.array(train_y).astype('int32')]
print
print "The length of training data: ", len(train_x)
print "The length of test data: ", len(crossvalid_x)
print
dc_acc = deque(maxlen=20)
dt_acc = deque(maxlen=20)
epoch = 0
while True:
predictions_y = e.network.predict(crossvalid_x)
c_acc, c_sil = get_accuracy(crossvalid_y, predictions_y)
predictions_y = e.network.predict(train_x)
t_acc, t_sil = get_accuracy(train_y, predictions_y)
print
print "max_frames, max_files, max_frames_per_segment, trim_segments, max_epoch, hidden_units, last_frames, crossvalid_frames, usec0, usedelta, useacc, mel_banks_only, preconditioner, hidden_dropouts, weight_l2"
print max_frames, max_files, max_frames_per_segment, trim_segments, max_epoch, hidden_units, last_frames, crossvalid_frames, usec0, usedelta, useacc, mel_banks_only, preconditioner, hidden_dropouts, weight_l2
print "Epoch: %d" % (epoch, )
print
print "Cross-validation stats"
print "------------------------"
print "Epoch predictive accuracy: %0.2f" % c_acc
print "Last epoch accs:", ["%.2f" % x for x in dc_acc]
print "Epoch sil bias: %0.2f" % c_sil
print
print "Training stats"
print "------------------------"
print "Epoch predictive accuracy: %0.2f" % t_acc
print "Last epoch accs:", ["%.2f" % x for x in dt_acc]
print "Epoch sil bias: %0.2f" % t_sil
if epoch == max_epoch:
break
epoch += 1
e.run(train, crossvalid)
dc_acc.append(c_acc)
dt_acc.append(t_acc)
nn = ffnn.FFNN()
for w, b in zip(e.network.weights, e.network.biases):
nn.add_layer(w.get_value(), b.get_value())
nn.save(file_name = "model_voip/vad_sds_mfcc_is%d_hu%d_lf%d_mfr%d_mfl%d_mfps%d_ts%d_usec0%d_usedelta%d_useacc%d_mbo%d.nn" % \
(input_size, hidden_units, last_frames, max_frames, max_files, max_frames_per_segment, trim_segments, usec0, usedelta, useacc, mel_banks_only))
