net1 = nn(
layers = [
('input', layers.InputLayer),
('conv1', layers.Conv2DLayer),
('pool1', layers.MaxPool2DLayer),
# ('dropout1', layers.DropoutLayer),
('conv2', layers.Conv2DLayer),
('pool2', layers.MaxPool2DLayer),
('dropout2', layers.DropoutLayer),
('hidden3', layers.DenseLayer),
('dropout3', layers.DropoutLayer),
('hidden4', layers.DenseLayer),
('dropout4', layers.DropoutLayer),
('output', layers.DenseLayer),
],
input_shape=(None, 1, 28, 28),
conv1_num_filters = 128, conv1_filter_size=(5,5), pool1_pool_size=(2,2),
conv1_W=lasagne.init.GlorotUniform(),
conv1_nonlinearity=leaky_rectify,
# dropout1_p=0.3,
conv2_num_filters = 64, conv2_filter_size=(3,3), pool2_pool_size=(2,2),
conv2_W=lasagne.init.GlorotUniform(),
conv2_nonlinearity=leaky_rectify,
dropout2_p=0.4,
hidden3_num_units = 2000, hidden4_num_units=2000,
hidden3_nonlinearity=leaky_rectify,
hidden4_nonlinearity=leaky_rectify,
dropout3_p=0.5,
dropout4_p=0.5,
batch_iterator_train=BatchIterator(batch_size=256),
output_num_units=10,
output_nonlinearity=softmax,
# objective_loss_function=categorical_crossentropy,
regression=False,
use_label_encoder=True,
max_epochs=30,
update=nesterov_momentum,
update_learning_rate = theano.shared(float32(0.1)),
update_momentum = theano.shared(float32(0.9)),
on_epoch_finished = [
AdjustVariable('update_learning_rate', start=0.1, stop=0.001),
AdjustVariable('update_momentum', start=0.9, stop=0.9999),
EarlyStopping(patience=6),
],
verbose=1
)
net9 = nn(
layers=[
('input', layers.InputLayer),
('conv1', layers.Conv2DLayer),
('pool1', layers.MaxPool2DLayer),
('dropout1', layers.DropoutLayer), # !
('conv2', layers.Conv2DLayer),
('pool2', layers.MaxPool2DLayer),
('dropout2', layers.DropoutLayer), # !
('conv3', layers.Conv2DLayer),
('pool3', layers.MaxPool2DLayer),
('dropout3', layers.DropoutLayer), # !
('hidden4', layers.DenseLayer),
('dropout4', layers.DropoutLayer), # !
('hidden5', layers.DenseLayer),
('output', layers.DenseLayer),
],
input_shape=(None, 1, 96, 96),
conv1_num_filters=64, conv1_filter_size=(5, 5), pool1_pool_size=(2, 2),
conv1_nonlinearity=leaky_rectify,
dropout1_p=0.1, # !
conv2_num_filters=128, conv2_filter_size=(3, 3), pool2_pool_size=(2, 2),
conv2_nonlinearity=leaky_rectify,
dropout2_p=0.2, # !
conv3_num_filters=256, conv3_filter_size=(2, 2), pool3_pool_size=(2, 2),
conv3_nonlinearity=leaky_rectify,
dropout3_p=0.3, # !
hidden4_num_units=2048,
hidden4_nonlinearity=leaky_rectify,
dropout4_p=0.5, # !
hidden5_num_units=1000,
hidden5_nonlinearity=leaky_rectify,
output_num_units=30, output_nonlinearity=None,
update_learning_rate=theano.shared(float32(0.03)),
update_momentum=theano.shared(float32(0.9)),
regression=True,
batch_iterator_train=FlipBatchIterator(batch_size=128),
on_epoch_finished=[
AdjustVariable('update_learning_rate', start=0.01, stop=0.0007),
AdjustVariable('update_momentum', start=0.9, stop=0.999),
EarlyStopping(patience=20),
],
max_epochs=5000,
verbose=1,
)
net = nn(
layers = [
('input', layers.InputLayer),
('conv1', layers.Conv2DLayer),
('pool1', layers.MaxPool2DLayer),
('dropout1', layers.DropoutLayer),
('conv2', layers.Conv2DLayer),
('pool2', layers.MaxPool2DLayer),
('dropout2', layers.DropoutLayer),
# ('conv3', layers.Conv2DLayer),
# ('pool3', layers.MaxPool2DLayer),
# ('dropout3', layers.DropoutLayer),
# ('conv4', layers.Conv2DLayer),
# ('pool4', layers.MaxPool2DLayer),
# ('dropout4', layers.DropoutLayer),
('hidden5', layers.DenseLayer),
('dropout5', layers.DropoutLayer),
('hidden6', layers.DenseLayer),
('dropout6', layers.DropoutLayer),
# ('hidden7', layers.DenseLayer),
# ('dropout7', layers.DropoutLayer),
('output', layers.DenseLayer),
],
input_shape=(None, 3, 32, 32),
conv1_num_filters = 256, conv1_filter_size=(5,5), pool1_pool_size=(2,2),
conv1_W=lasagne.init.GlorotUniform(),
conv1_nonlinearity=LeakyRectify(leakiness=.03),
dropout1_p=0.2,
conv2_num_filters = 128, conv2_filter_size=(2,2), pool2_pool_size=(2,2),
conv2_W=lasagne.init.GlorotUniform(),
conv2_nonlinearity=LeakyRectify(leakiness=.03),
dropout2_p=0.3,
# conv3_num_filters = 128, conv3_filter_size=(2,2), pool3_pool_size=(2,2),
# conv3_W=lasagne.init.GlorotUniform(),
# conv3_nonlinearity=LeakyRectify,
# dropout3_p=0.4,
# conv4_num_filters = 256, conv4_filter_size=(2,2), pool4_pool_size=(2,2),
# conv4_W=lasagne.init.GlorotUniform(),
# conv4_nonlinearity=LeakyRectify,
# dropout4_p=0.4,
hidden5_num_units=4096,
hidden5_nonlinearity=LeakyRectify(leakiness=.03),
dropout5_p=0.5,
hidden6_num_units = 4096,
hidden6_nonlinearity=LeakyRectify(leakiness=.03),
dropout6_p=0.5,
# hidden7_num_units = 1000,
# hidden7_nonlinearity=LeakyRectify,
# dropout7_p=0.2,
batch_iterator_train=BatchIterator(batch_size=128),
train_split=TrainSplit(eval_size=0.08),
output_num_units=10,
output_nonlinearity=softmax,
# objective_loss_function=categorical_crossentropy,
regression=False,
use_label_encoder=True,
max_epochs=10000,
update=nesterov_momentum,
update_learning_rate = theano.shared(float32(0.03)),
update_momentum = theano.shared(float32(0.9)),
on_epoch_finished = [
AdjustVariable('update_learning_rate', start=0.01, stop=0.0001),
AdjustVariable('update_momentum', start=0.9, stop=0.9999),
EarlyStopping(patience=200),
],
verbose=1
)