def __call__(self, nn, train_history):
#if self.ls is None:
# self.ls = np.linspace(self.start, self.stop, nn.max_epochs)
## detect if NN_EPOCHS has changed
#if Settings.NN_EPOCHS != nn.max_epochs:
# changeEpochs = utils.radio_question("[!]", "Change in NN_EPOCHS detected. Change net?", None, ["Yes", "No"], [True, False])
# if changeEpochs:
# #getattr(nn, "max_epochs").set_value(Settings.NN_EPOCHS)
# nn.max_epochs = Settings.NN_EPOCHS
# self.ls = np.linspace(self.start, self.stop, Settings.NN_EPOCHS)
# else:
# raise StopIteration()
epoch = train_history[-1]['epoch']
stop = self.start * 10e-2 * 2
stop2 = stop * 10e-4 * 2
ls = np.linspace(self.start, stop, 50)
ls2 = np.linspace(stop, stop2, nn.max_epochs - 100)
if epoch <= 50:
new_value = utils.to_float32(ls[epoch - 1])
elif epoch <= 100:
new_value = utils.to_float32(ls[-1])
else:
new_value = utils.to_float32(ls2[epoch - 1 - 100])
getattr(nn, self.name).set_value(new_value)
def __init__(self, outputShape, testData, modelSaver):
self.set_network_specific_settings()
modelSaver.model = self
self.net = NeuralNet(
layers=[
('input', layers.InputLayer),
('hidden1', layers.DenseLayer),
('hidden2', layers.DenseLayer),
('output', layers.DenseLayer),
],
# Layer parameter
input_shape=(None, Settings.NN_CHANNELS,
Settings.NN_INPUT_SHAPE[0], Settings.NN_INPUT_SHAPE[1]
), # variable batch size, single row shape
hidden1_num_units=500,
hidden2_num_units=50,
output_num_units=outputShape,
output_nonlinearity=lasagne.nonlinearities.softmax,
# optimization method:
update=nesterov_momentum,
update_learning_rate=theano.shared(
utils.to_float32(Settings.NN_START_LEARNING_RATE)),
update_momentum=theano.shared(
utils.to_float32(Settings.NN_START_MOMENTUM)),
batch_iterator_train=AugmentingLazyBatchIterator(
Settings.NN_BATCH_SIZE,
testData,
"train",
True,
loadingSize=(50, 50)),
batch_iterator_test=LazyBatchIterator(Settings.NN_BATCH_SIZE,
testData,
"valid",
False,
newSegmentation=False),
train_split=TrainSplit(
eval_size=0.0), # we cross validate on our own
regression=False, # classification problem
on_epoch_finished=[
AdjustVariable('update_learning_rate',
start=Settings.NN_START_LEARNING_RATE,
stop=0.0001),
AdjustVariable('update_momentum',
start=Settings.NN_START_MOMENTUM,
stop=0.999),
TrainingHistory("?", str(self), [], modelSaver),
EarlyStopping(150),
modelSaver,
],
max_epochs=Settings.NN_EPOCHS,
verbose=1,
)
def __init__(self, outputShape, modelSaver):
self.set_network_specific_settings()
modelSaver.model = self
self.net = NeuralNet(
layers=[
('input', layers.InputLayer),
('conv1', layers.Conv2DLayer),
('pool1', layers.MaxPool2DLayer),
('conv2', layers.Conv2DLayer),
('pool2', layers.MaxPool2DLayer),
('conv3', layers.Conv2DLayer),
('pool3', layers.MaxPool2DLayer),
('hidden4', layers.DenseLayer),
('hidden5', layers.DenseLayer),
('output', layers.DenseLayer),
],
input_shape=(None, Settings.NN_CHANNELS, Settings.NN_INPUT_SHAPE[0], Settings.NN_INPUT_SHAPE[1]), # variable batch size, 3 color shape row shape
conv1_num_filters=32, conv1_filter_size=(3, 3), pool1_pool_size=(2, 2),
conv2_num_filters=64, conv2_filter_size=(2, 2), pool2_pool_size=(2, 2),
conv3_num_filters=128, conv3_filter_size=(2, 2), pool3_pool_size=(2, 2),
hidden4_num_units=500,
hidden5_num_units=500,
output_num_units=outputShape,
output_nonlinearity=lasagne.nonlinearities.softmax,
# optimization method:
update=nesterov_momentum,
update_learning_rate=theano.shared(utils.to_float32(Settings.NN_START_LEARNING_RATE)),
update_momentum=theano.shared(utils.to_float32(Settings.NN_START_MOMENTUM)),
regression=False, # classification problem
on_epoch_finished=[
AdjustVariable('update_learning_rate', start=Settings.NN_START_LEARNING_RATE, stop=0.0001),
AdjustVariable('update_momentum', start=Settings.NN_START_MOMENTUM, stop=0.999),
TrainingHistory("?", str(self), [1], modelSaver),
EarlyStopping(30),
modelSaver,
],
max_epochs=Settings.NN_EPOCHS,
verbose=1,
)
def __init__(self, outputShape):
self.set_network_specific_settings()
self.net = NeuralNet(
layers=[
('input', layers.InputLayer),
('conv1', layers.Conv2DLayer),
('pool1', layers.MaxPool2DLayer),
('conv2', layers.Conv2DLayer),
('pool2', layers.MaxPool2DLayer),
('conv3', layers.Conv2DLayer),
('pool3', layers.MaxPool2DLayer),
('hidden4', layers.DenseLayer),
('hidden5', layers.DenseLayer),
('output', layers.DenseLayer),
],
input_shape=(None, Settings.NN_CHANNELS,
Settings.NN_INPUT_SHAPE[0], Settings.NN_INPUT_SHAPE[1]
), # variable batch size, 3 color shape row shape
conv1_num_filters=32,
conv1_filter_size=(3, 3),
pool1_pool_size=(2, 2),
conv2_num_filters=64,
conv2_filter_size=(2, 2),
pool2_pool_size=(2, 2),
conv3_num_filters=128,
conv3_filter_size=(2, 2),
pool3_pool_size=(2, 2),
hidden4_num_units=500,
hidden5_num_units=500,
output_num_units=outputShape,
output_nonlinearity=lasagne.nonlinearities.softmax,
# optimization method:
update=nesterov_momentum,
update_learning_rate=theano.shared(
utils.to_float32(Settings.NN_START_LEARNING_RATE)),
update_momentum=theano.shared(
utils.to_float32(Settings.NN_START_MOMENTUM)),
regression=False, # classification problem
batch_iterator_train=AugmentingBatchIterator(
batch_size=Settings.NN_BATCH_SIZE),
max_epochs=Settings.NN_EPOCHS,
verbose=1,
)
def __init__(self, outputShape):
self.set_network_specific_settings()
self.net = NeuralNet(
layers=[
('input', layers.InputLayer),
('conv2d1', layers.Conv2DLayer),
('maxpool1', layers.MaxPool2DLayer),
('conv2d2', layers.Conv2DLayer),
('maxpool2', layers.MaxPool2DLayer),
('dense', layers.DenseLayer),
('output', layers.DenseLayer),
],
input_shape=(None, Settings.NN_CHANNELS,
Settings.NN_INPUT_SHAPE[0], Settings.NN_INPUT_SHAPE[1]
), # variable batch size, 3 color shape row shape
conv2d1_num_filters=20,
conv2d1_filter_size=(5, 5),
conv2d1_stride=(1, 1),
conv2d1_pad=(2, 2),
conv2d1_nonlinearity=lasagne.nonlinearities.rectify,
maxpool1_pool_size=(2, 2),
conv2d2_num_filters=20,
conv2d2_filter_size=(5, 5),
conv2d2_stride=(1, 1),
conv2d2_pad=(2, 2),
conv2d2_nonlinearity=lasagne.nonlinearities.rectify,
maxpool2_pool_size=(2, 2),
dense_num_units=1000,
dense_nonlinearity=lasagne.nonlinearities.rectify,
output_num_units=outputShape,
output_nonlinearity=lasagne.nonlinearities.softmax,
# optimization method:
update=nesterov_momentum,
update_learning_rate=theano.shared(
utils.to_float32(Settings.NN_START_LEARNING_RATE)),
update_momentum=theano.shared(
utils.to_float32(Settings.NN_START_MOMENTUM)),
regression=False, # classification problem
max_epochs=Settings.NN_EPOCHS,
verbose=1,
)
def __init__(self, outputShape, testData, modelSaver):
self.set_network_specific_settings()
modelSaver.model = self
self.net = NeuralNet(
layers=[
('input', layers.InputLayer),
('conv1', layers.Conv2DLayer),
('pool1', layers.MaxPool2DLayer),
('conv2', layers.Conv2DLayer),
('pool2', layers.MaxPool2DLayer),
('conv3', layers.Conv2DLayer),
('conv4', layers.Conv2DLayer),
('conv5', layers.Conv2DLayer),
('pool3', layers.MaxPool2DLayer),
('hidden6', layers.DenseLayer),
('dropout1', layers.DropoutLayer),
('hidden7', layers.DenseLayer),
('dropout2', layers.DropoutLayer),
('output', layers.DenseLayer),
],
input_shape=(None, Settings.NN_CHANNELS, Settings.NN_INPUT_SHAPE[0], Settings.NN_INPUT_SHAPE[1]), # variable batch size, 3 color shape row shape
conv1_num_filters=96, conv1_filter_size=(11, 11), conv1_stride=(4, 4),
pool1_pool_size=(5, 5),
conv2_num_filters=256, conv2_filter_size=(5, 5),
pool2_pool_size=(3, 3),
conv3_num_filters=384, conv3_filter_size=(3, 3), conv3_pad = (1,1),
conv4_num_filters=384, conv4_filter_size=(3, 3), conv4_pad = (1,1),
conv5_num_filters=256, conv5_filter_size=(3, 3), conv5_pad = (1,1),
pool3_pool_size=(2, 2),
hidden6_num_units=4096,
dropout1_p=0.5,
hidden7_num_units=4096,
dropout2_p=0.5,
output_num_units=outputShape,
output_nonlinearity=lasagne.nonlinearities.softmax,
# optimization method:
update=nesterov_momentum,
update_learning_rate=theano.shared(utils.to_float32(Settings.NN_START_LEARNING_RATE)),
update_momentum=theano.shared(utils.to_float32(Settings.NN_START_MOMENTUM)),
batch_iterator_train=AugmentingLazyBatchIterator(Settings.NN_BATCH_SIZE, testData, "train", False, newSegmentation=False, loadingSize=(256,256)),
batch_iterator_test=LazyBatchIterator(Settings.NN_BATCH_SIZE, testData, "valid", False, newSegmentation=False, loadingInputShape=Settings.NN_INPUT_SHAPE),
train_split=TrainSplit(eval_size=0.0), # we cross validate on our own
regression=False, # classification problem
on_epoch_finished=[
AdjustVariable('update_learning_rate', start=Settings.NN_START_LEARNING_RATE, stop=0.0001),
AdjustVariable('update_momentum', start=Settings.NN_START_MOMENTUM, stop=0.999),
TrainingHistory("Krizhevsky", str(self), [], modelSaver),
EarlyStopping(150),
modelSaver,
],
max_epochs=Settings.NN_EPOCHS,
verbose=1,
)
def __init__(self, outputShape, testData, modelSaver):
super(net_cifar_cnn_do, self).__init__(outputShape)
modelSaver.model = self
self.net = NeuralNet(
layers=[
('input', layers.InputLayer),
('conv2d1', layers.Conv2DLayer),
('maxpool1', layers.MaxPool2DLayer),
('conv2d2', layers.Conv2DLayer),
('maxpool2', layers.MaxPool2DLayer),
('dropout1', layers.DropoutLayer),
('dense', layers.DenseLayer),
('dropout2', layers.DropoutLayer),
('output', layers.DenseLayer),
],
input_shape=(None, Settings.NN_CHANNELS,
Settings.NN_INPUT_SHAPE[0], Settings.NN_INPUT_SHAPE[1]
), # variable batch size, 3 color shape row shape
conv2d1_num_filters=20,
conv2d1_filter_size=(5, 5),
conv2d1_stride=(1, 1),
conv2d1_pad=(2, 2),
conv2d1_nonlinearity=lasagne.nonlinearities.rectify,
maxpool1_pool_size=(2, 2),
conv2d2_num_filters=20,
conv2d2_filter_size=(5, 5),
conv2d2_stride=(1, 1),
conv2d2_pad=(2, 2),
conv2d2_nonlinearity=lasagne.nonlinearities.rectify,
maxpool2_pool_size=(2, 2),
dropout1_p=0.5,
dense_num_units=1000,
dense_nonlinearity=lasagne.nonlinearities.rectify,
dropout2_p=0.5,
output_num_units=outputShape,
output_nonlinearity=lasagne.nonlinearities.softmax,
# optimization method:
update=nesterov_momentum,
update_learning_rate=theano.shared(
utils.to_float32(Settings.NN_START_LEARNING_RATE)),
update_momentum=theano.shared(
utils.to_float32(Settings.NN_START_MOMENTUM)),
#batch_iterator_train=AugmentingLazyBatchIterator(Settings.NN_BATCH_SIZE, testData, "train", False, newSegmentation=False, loadingSize=(120,120)),
#batch_iterator_test=LazyBatchIterator(Settings.NN_BATCH_SIZE, testData, "valid", False, newSegmentation=False, loadingInputShape=Settings.NN_INPUT_SHAPE),
#train_split=TrainSplit(eval_size=0.0), # we cross validate on our own
regression=False, # classification problem
on_epoch_finished=[
AdjustVariable('update_learning_rate',
start=Settings.NN_START_LEARNING_RATE,
stop=0.0001),
AdjustVariable('update_momentum',
start=Settings.NN_START_MOMENTUM,
stop=0.999),
TrainingHistory("Cifar10-do1", str(self), [1], modelSaver),
EarlyStopping(300),
modelSaver,
],
max_epochs=Settings.NN_EPOCHS,
verbose=1,
)
