def train(self, X, Y, num_iterations=1000, learning_rate=0.01, batch_size=50):
assert Y.shape[1] == self.config['layers'][-1], 'output dimensions do not match.'
assert X.shape[1] == self.model.layers[0].get_weights().shape[0], 'input dimensions do not match.'
self.config.update({'iterations':num_iterations, 'learning_rate':learning_rate, 'batch_size':batch_size})
algorithm = pylearn2.training_algorithms.sgd.SGD(
learning_rate = self.config['learning_rate'],
#init_momentum = 0.5,
batch_size=self.config['batch_size'],
batches_per_iter=X.shape[0]//self.config['batch_size'],
cost = pylearn2.costs.mlp.Default(),
termination_criterion = pylearn2.termination_criteria.EpochCounter(self.config['iterations'])
)
data_train = pylearn2.datasets.DenseDesignMatrix(X=X, y=Y)
train = pylearn2.train.Train(
dataset = data_train,
model = self.model,
algorithm = algorithm,
save_path = 'tmp.pkl',
save_freq = 100)
train.main_loop()
def train(self, X, num_iterations=1000, learning_rate=0.01, batch_size=50):
self.config.update( {'iterations':num_iterations, 'learning_rate':learning_rate, 'batch_size':batch_size})
algorithm = pylearn2.training_algorithms.sgd.SGD(
learning_rate = self.config['learning_rate'],
batch_size = batch_size,
batches_per_iter=int(X.shape[0]/batch_size),
cost = pylearn2.costs.autoencoder.MeanSquaredReconstructionError(),
termination_criterion = pylearn2.termination_criteria.EpochCounter(self.config['iterations']),
)
data_train = pylearn2.datasets.DenseDesignMatrix(X=X[:,:])
train = pylearn2.train.Train(
dataset = data_train,
model = self.model,
algorithm = algorithm,
save_path = 'tmp.pkl',
save_freq = 100,
extensions=[pylearn2.training_algorithms.sgd.OneOverEpoch(self.config['iterations']//2, self.config['iterations']//8)]
)
train.main_loop()
def compute_objective(args):
'''Initializes neural network with specified arguments and trains.'''
# Train network.
train = init_train(args)
train.main_loop()
model = train.model
# Return objective to hyperopt.
loss = train.model.monitor.channels['valid_y_kl'].val_record[-1]
return loss
def compute_objective(args):
'''Initializes neural network with specified arguments and trains.'''
# Train network.
train = init_train(args)
train.main_loop()
model = train.model
# Return objective to hyperopt.
loss = train.model.monitor.channels['valid_y_kl'].val_record[-1]
return loss
def main(argv, freeze):
try:
opts, args = getopt.getopt(argv, '')
yaml = args[0]
model = args[1]
except getopt.GetoptError:
usage()
sys.exit(2)
# Load yaml
with open(yaml, "r") as sty:
train = serial.load_train_file(yaml)
#train = yaml_parse.load(sty)
# Load pretrained model with bad sigmoid output
with open(model, 'r') as fo:
model = pkl.load(fo)
# Remove the last layer, puts a real sigmoid instead
if freeze:
for i in range(0, len(model.layers) - 2):
model.freeze(model.layers[i].get_params())
### Add last conv elemwise
layer = ConvElemwise(layer_name= 'out',
output_channels= 1,
kernel_shape=(1,1),
irange=0.05,
nonlinearity=IdentityConvNonlinearity(),
max_kernel_norm= 7.9,
tied_b=1)
layer.set_mlp(model)
layer.set_input_space(model.layers[-3].get_output_space())
model.layers[-2] = layer
### Add Sigmoid
layer = SigmoidExtended(layer_name='y', n_classes=1)
layer.set_mlp(model)
layer.set_input_space(model.layers[-2].get_output_space())
model.layers[-1] = layer
#print model.layers
#model.monitor = train.model.monitor
#train.model = model
train.model = push_monitor(model, "old")
print train.model
#train = Train(train.dataset, model, train.algorithm, train.save_path,
# train.save_freq, train.extensions, train.allow_overwrite)
train.main_loop()
def main(argv, freeze):
try:
opts, args = getopt.getopt(argv, '')
yaml = args[0]
model = args[1]
except getopt.GetoptError:
usage()
sys.exit(2)
# Load yaml
with open(yaml, "r") as sty:
train = serial.load_train_file(yaml)
#train = yaml_parse.load(sty)
# Load pretrained model with bad sigmoid output
with open(model, 'r') as fo:
model = pkl.load(fo)
# Remove the last layer, puts a real sigmoid instead
if freeze:
for i in range(0, len(model.layers) - 2):
model.freeze(model.layers[i].get_params())
### Add last conv elemwise
layer = ConvElemwise(layer_name='out',
output_channels=1,
kernel_shape=(1, 1),
irange=0.05,
nonlinearity=IdentityConvNonlinearity(),
max_kernel_norm=7.9,
tied_b=1)
layer.set_mlp(model)
layer.set_input_space(model.layers[-3].get_output_space())
model.layers[-2] = layer
### Add Sigmoid
layer = SigmoidExtended(layer_name='y', n_classes=1)
layer.set_mlp(model)
layer.set_input_space(model.layers[-2].get_output_space())
model.layers[-1] = layer
#print model.layers
#model.monitor = train.model.monitor
#train.model = model
train.model = push_monitor(model, "old")
print train.model
#train = Train(train.dataset, model, train.algorithm, train.save_path,
# train.save_freq, train.extensions, train.allow_overwrite)
train.main_loop()
def train(self, X, num_iterations=1000, learning_rate=0.01, batch_size=50):
self.config.update( {'iterations':num_iterations, 'learning_rate':learning_rate, 'batch_size':batch_size})
data_train = pylearn2.datasets.DenseDesignMatrix(X=X[:,:])
self._build_algorithm(X.shape[0])
train = pylearn2.train.Train(
dataset = data_train,
model = self.model,
algorithm = self.algorithm,
save_path = 'tmp.pkl',
save_freq = 100,
extensions=[pylearn2.training_algorithms.sgd.OneOverEpoch(max(1,self.config['iterations']//2), max(1,self.config['iterations']//8))]
)
train.main_loop()
def Compute_Objective(args,conf, **kwargs):
'''Initializes neural network with specified arguments and
configuration and trains.
'''
# Train network.
old_stdout = sys.stdout
old_stderr = sys.stderr
dev_path = os.environ['DNN_PATH']
modelname = '%s_%d_%d_%d_%0.14f_%0.14f_%0.6f_%d_%0.6f' % tuple(args[:-1])
modelname = '%s_%s_%d' % (str( conf['ptype'] ),modelname,
int( conf['batch_size']) )
train = init_train(args,conf,**kwargs)
train.main_loop()
model = train.model
sys.stdout = old_stdout
sys.stderr = old_stderr
# Return objective to hyperopt.
loss = train.model.monitor.channels['valid_objective'].val_record
try:
fig = plt.figure()
fig.suptitle('Progress of Training', fontsize=20)
plt.xlabel('Number of Epochs')
plt.ylabel('Value from Loss-function')
plt.plot(loss)
plt.plot(loss,'b.')
plt.savefig('%s/figure/objective/%s.pdf' % (dev_path,modelname))
plt.close('all')
np.savetxt('%s/figure/objective/%s.dat' % (dev_path,modelname), loss)
except:
pass
return loss[-1]
input_include_probs={'hidden_0':1., 'hidden_1':1., 'hidden_2':1., 'hidden_3':1., 'y':0.5},
input_scales={ 'hidden_0': 1., 'hidden_1':1., 'hidden_2':1., 'hidden_3':1., 'y':2.}),
update_callbacks=pylearn2.training_algorithms.sgd.ExponentialDecay(
decay_factor=1.0000003, # Decreases by this factor every batch. (1/(1.000001^8000)^100
min_lr=.000001
));
## EXTENSION
extensions=[pylearn2.training_algorithms.learning_rule.MomentumAdjustor(start=0,
saturate=momentum_saturate,
final_momentum=.99)];
## TRAINING MODEL
train=pylearn2.train.Train(dataset=train_set,
model=model,
algorithm=algorithm,
save_path=save_path,
save_freq=100);
debug = False
logfile = os.path.splitext(train.save_path)[0] + '.log'
print 'Using=%s' % theano.config.device # Can use gpus.
print 'Writing to %s' % logfile
print 'Writing to %s' % train.save_path
sys.stdout = open(logfile, 'w')
train.main_loop();
