def __init__(self):
CNN.__init__(
self,
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),
('output', layers.DenseLayer),
],
# input
input_shape=(None, 3, 75, 75),
# conv2d + pool + dropout
conv1_filter_size=(3, 3),
conv1_num_filters=64,
pool1_pool_size=(2, 2),
dropout1_p=0.2,
# conv2d + pool + dropout
conv2_filter_size=(3, 3),
conv2_num_filters=48,
pool2_pool_size=(2, 2),
dropout2_p=0.2,
# conv2d + pool + dropout
conv3_filter_size=(3, 3),
conv3_num_filters=48,
pool3_pool_size=(2, 2),
dropout3_p=0.2,
# conv2d + pool + dropout
conv4_filter_size=(3, 3),
conv4_num_filters=48,
pool4_pool_size=(2, 2),
dropout4_p=0.2,
# dense layer 1
hidden5_num_units=512,
hidden5_nonlinearity=nonlinearities.rectify,
dropout5_p=0.5,
# dense layer 2
output_num_units=2,
output_nonlinearity=nonlinearities.softmax)
def __init__(self, num_label):
self.layers = [
['conv', 512, 64, 1, 0],
['norm'],
['relu'],
]
self._make_layers(BlockType.BASIC, 64, 64, 2)
self._make_layers(BlockType.BASIC, 64, 128, 2)
self._make_layers(BlockType.BASIC, 128, 256, 2)
self._make_layers(BlockType.BASIC, 256, 512, 2)
self.layers.extend([['avgPool', 4, 1], ['flatten'],
['fc', 512, num_label]])
CNN.__init__(self)
def __init__(self, num_label):
self.layers = [
['conv', 3, 64, 3, 1],
]
out_dim = self.add_res_layers()
self.layers.extend([
#['norm'], ['relu'],
['avgPool', 4, 1],
#['conv', 512, 256, 1, 0], ['norm'], ['relu'],
#['dropout', 0.5],
#['conv', 256, 256, 1, 0], ['norm'], ['relu'],
#['dropout', 0.5],
['flatten'],
['fc', out_dim, num_label]
])
CNN.__init__(self)
def __init__(self, dataset: AbstractDataset):
H, W, C = dataset.num_dims()
out_dim = H // 8
self.layers = [['conv', C, 64, 3, 1], ['norm'], ['relu'],
['conv', 64, 64, 3, 1], ['norm'], ['relu'],
['maxPool', 2, 2], ['conv', 64, 128, 3, 1], ['norm'],
['relu'], ['conv', 128, 128, 3, 1], ['norm'], ['relu'],
['conv', 128, 128, 3, 1], ['norm'], ['relu'],
['maxPool', 2, 2], ['conv', 128, 256, 3, 1], ['norm'],
['relu'], ['conv', 256, 256, 3, 1], ['norm'], ['relu'],
['conv', 256, 256, 3, 1], ['norm'], ['relu'],
['maxPool', 2, 2], ['conv', 256, 512, 3, 1], ['norm'],
['relu'], ['conv', 512, 512, 3, 1], ['norm'], ['relu'],
['conv', 512, 512, 3, 1], ['norm'], ['relu'],
['avgPool', out_dim, 1], ['flatten'],
['fc', 512, dataset.num_labels()]]
CNN.__init__(self)
def __init__(self, num_label):
self.layers = [['conv', 3, 64, 3, 1], ['norm'], ['relu'],
['conv', 64, 64, 3, 1], ['norm'], ['relu'],
['nonlocal', 64], ['maxPool', 2, 2],
['conv', 64, 128, 3, 1], ['norm'], ['relu'],
['conv', 128, 128, 3, 1], ['norm'], ['relu'],
['conv', 128, 128, 3, 1], ['norm'], ['relu'],
['maxPool', 2, 2], ['conv', 128, 256, 3, 1], ['norm'],
['relu'], ['conv', 256, 256, 3, 1], ['norm'], ['relu'],
['conv', 256, 256, 3, 1], ['norm'], ['relu'],
['maxPool', 2, 2], ['conv', 256, 512, 3, 1], ['norm'],
['relu'], ['conv', 512, 512, 3, 1], ['norm'], ['relu'],
['conv', 512, 512, 3, 1], ['norm'], ['relu'],
['avgPool', 4, 1], ['conv', 512, 256, 1, 0], ['norm'],
['relu'], ['dropout', 0.5], ['conv', 256, 256, 1, 0],
['norm'], ['relu'], ['dropout', 0.5], ['flatten'],
['fc', 256, num_label]]
CNN.__init__(self)
def __init__(self):
CNN.__init__(
self,
layers=[
('input', layers.InputLayer),
('conv1', layers.Conv2DLayer),
('pool1', layers.MaxPool2DLayer),
('conv2', layers.Conv2DLayer),
('pool2', layers.MaxPool2DLayer),
('conv3', layers.Conv2DLayer),
('pool3', layers.MaxPool2DLayer),
('hidden3', layers.DenseLayer),
('dropout3', layers.DropoutLayer),
('output', layers.DenseLayer),
],
# input
input_shape=(None, 4, 75, 75),
# conv2d + pool + dropout
conv1_filter_size=(13, 13),
conv1_num_filters=16,
pool1_pool_size=(2, 2),
# conv2d + pool + dropout
conv2_filter_size=(13, 13),
conv2_num_filters=16,
pool2_pool_size=(2, 2),
# conv2d + pool + dropout
conv3_filter_size=(13, 13),
conv3_num_filters=16,
pool3_pool_size=(2, 2),
# dense layer 1
hidden3_num_units=256,
hidden3_nonlinearity=nonlinearities.rectify,
dropout3_p=0.5,
# dense layer 2
output_num_units=2,
output_nonlinearity=nonlinearities.softmax)
def __init__(self):
CNN.__init__(
self,
layers=[
('image_input', layers.InputLayer),
('image_conv1', layers.Conv2DLayer),
('image_pool1', layers.MaxPool2DLayer),
('image_conv2', layers.Conv2DLayer),
('image_pool2', layers.MaxPool2DLayer),
('prob_input', layers.InputLayer),
('prob_conv1', layers.Conv2DLayer),
('prob_pool1', layers.MaxPool2DLayer),
('prob_conv2', layers.Conv2DLayer),
('prob_pool2', layers.MaxPool2DLayer),
('binary_input', layers.InputLayer),
('binary_conv1', layers.Conv2DLayer),
('binary_pool1', layers.MaxPool2DLayer),
('binary_conv2', layers.Conv2DLayer),
('binary_pool2', layers.MaxPool2DLayer),
('merge', layers.ConcatLayer),
('hidden3', layers.DenseLayer),
('output', layers.DenseLayer),
],
# input
image_input_shape=(None, 1, 75, 75),
# conv2d + pool + dropout
image_conv1_filter_size=(13, 13),
image_conv1_num_filters=16,
image_conv1_nonlinearity=nonlinearities.rectify,
image_pool1_pool_size=(2, 2),
# conv2d + pool + dropout
image_conv2_filter_size=(13, 13),
image_conv2_num_filters=16,
image_conv2_nonlinearity=nonlinearities.rectify,
image_pool2_pool_size=(2, 2),
prob_input_shape=(None, 1, 75, 75),
# conv2d + pool + dropout
prob_conv1_filter_size=(13, 13),
prob_conv1_num_filters=16,
prob_conv1_nonlinearity=nonlinearities.rectify,
prob_pool1_pool_size=(2, 2),
# conv2d + pool + dropout
prob_conv2_filter_size=(13, 13),
prob_conv2_num_filters=16,
prob_conv2_nonlinearity=nonlinearities.rectify,
prob_pool2_pool_size=(2, 2),
binary_input_shape=(None, 1, 75, 75),
# conv2d + pool + dropout
binary_conv1_filter_size=(13, 13),
binary_conv1_num_filters=16,
binary_conv1_nonlinearity=nonlinearities.rectify,
binary_pool1_pool_size=(2, 2),
# conv2d + pool + dropout
binary_conv2_filter_size=(13, 13),
binary_conv2_num_filters=16,
binary_conv2_nonlinearity=nonlinearities.rectify,
binary_pool2_pool_size=(2, 2),
# concat
merge_incomings=['image_pool2', 'prob_pool2', 'binary_pool2'],
# dense layer 1
hidden3_num_units=256,
hidden3_nonlinearity=nonlinearities.rectify,
# dense layer 2
output_num_units=2,
output_nonlinearity=nonlinearities.softmax)
def __init__(self):
CNN.__init__(
self,
layers=[
('image_input', layers.InputLayer),
('image_conv1', layers.Conv2DLayer),
('image_pool1', layers.MaxPool2DLayer),
('image_dropout1', layers.DropoutLayer),
('image_conv2', layers.Conv2DLayer),
('image_pool2', layers.MaxPool2DLayer),
('image_dropout2', layers.DropoutLayer),
('prob_input', layers.InputLayer),
('prob_conv1', layers.Conv2DLayer),
('prob_pool1', layers.MaxPool2DLayer),
('prob_dropout1', layers.DropoutLayer),
('prob_conv2', layers.Conv2DLayer),
('prob_pool2', layers.MaxPool2DLayer),
('prob_dropout2', layers.DropoutLayer),
('binary_input', layers.InputLayer),
('binary_conv1', layers.Conv2DLayer),
('binary_pool1', layers.MaxPool2DLayer),
('binary_dropout1', layers.DropoutLayer),
('binary_conv2', layers.Conv2DLayer),
('binary_pool2', layers.MaxPool2DLayer),
('binary_dropout2', layers.DropoutLayer),
('border_input', layers.InputLayer),
('border_conv1', layers.Conv2DLayer),
('border_pool1', layers.MaxPool2DLayer),
('border_dropout1', layers.DropoutLayer),
('border_conv2', layers.Conv2DLayer),
('border_pool2', layers.MaxPool2DLayer),
('border_dropout2', layers.DropoutLayer),
('merge', layers.ConcatLayer),
('hidden3', layers.DenseLayer),
('dropout3', layers.DropoutLayer),
('output', layers.DenseLayer),
],
# input
image_input_shape=(None, 1, 75, 75),
# conv2d + pool + dropout
image_conv1_filter_size=(3, 3),
image_conv1_num_filters=64,
image_pool1_pool_size=(2, 2),
image_dropout1_p=0.2,
# conv2d + pool + dropout
image_conv2_filter_size=(3, 3),
image_conv2_num_filters=64,
image_pool2_pool_size=(2, 2),
image_dropout2_p=0.2,
prob_input_shape=(None, 1, 75, 75),
# conv2d + pool + dropout
prob_conv1_filter_size=(3, 3),
prob_conv1_num_filters=64,
prob_pool1_pool_size=(2, 2),
prob_dropout1_p=0.2,
# conv2d + pool + dropout
prob_conv2_filter_size=(3, 3),
prob_conv2_num_filters=64,
prob_pool2_pool_size=(2, 2),
prob_dropout2_p=0.2,
binary_input_shape=(None, 1, 75, 75),
# conv2d + pool + dropout
binary_conv1_filter_size=(3, 3),
binary_conv1_num_filters=64,
binary_pool1_pool_size=(2, 2),
binary_dropout1_p=0.2,
# conv2d + pool + dropout
binary_conv2_filter_size=(3, 3),
binary_conv2_num_filters=64,
binary_pool2_pool_size=(2, 2),
binary_dropout2_p=0.2,
border_input_shape=(None, 1, 75, 75),
# conv2d + pool + dropout
border_conv1_filter_size=(3, 3),
border_conv1_num_filters=64,
border_pool1_pool_size=(2, 2),
border_dropout1_p=0.2,
# conv2d + pool + dropout
border_conv2_filter_size=(3, 3),
border_conv2_num_filters=64,
border_pool2_pool_size=(2, 2),
border_dropout2_p=0.2,
# concat
# merge_incomings=['image_pool2','prob_pool2','binary_pool2','border_pool2'],
merge_incomings=[
'image_dropout2', 'prob_dropout2', 'binary_dropout2',
'border_dropout2'
],
# dense layer 1
hidden3_num_units=512,
hidden3_nonlinearity=nonlinearities.rectify,
dropout3_p=0.5,
# dense layer 2
output_num_units=2,
output_nonlinearity=nonlinearities.softmax)
def __init__(self, size=64, gray=False, gen_depth=16, dis_depth=2):
CNN.__init__(self, size=size, gray=gray, gen_depth=gen_depth)
self.dis_depth = dis_depth
self.D = None
self.combined = None
