def build_model(input_var,dro=0.5):
net = {}
net['input'] = InputLayer((None, 3, 299, 299),input_var=input_var)
print(net['input'])
net['conv1/7x7_s2'] = ConvLayer(
net['input'], 64, 7, stride=2, pad=3, flip_filters=False)
print(net['conv1/7x7_s2'])
net['pool1/3x3_s2'] = PoolLayer(
net['conv1/7x7_s2'], pool_size=3, stride=2, ignore_border=False)
net['pool1/norm1'] = LRNLayer(net['pool1/3x3_s2'], alpha=0.00002, k=1)
net['conv2/3x3_reduce'] = ConvLayer(
net['pool1/norm1'], 64, 1, flip_filters=False)
net['conv2/3x3'] = ConvLayer(
net['conv2/3x3_reduce'], 192, 3, pad=1, flip_filters=False)
net['conv2/norm2'] = LRNLayer(net['conv2/3x3'], alpha=0.00002, k=1)
net['pool2/3x3_s2'] = PoolLayer(
net['conv2/norm2'], pool_size=3, stride=2, ignore_border=False)
net.update(build_inception_module('inception_3a',
net['pool2/3x3_s2'],
[32, 64, 96, 128, 16, 32]))
net.update(build_inception_module('inception_3b',
net['inception_3a/output'],
[64, 128, 128, 192, 32, 96]))
net['pool3/3x3_s2'] = PoolLayer(
net['inception_3b/output'], pool_size=3, stride=2, ignore_border=False)
net.update(build_inception_module('inception_4a',
net['pool3/3x3_s2'],
[64, 192, 96, 208, 16, 48],dro))
net.update(build_inception_module('inception_4b',
net['inception_4a/output'],
[64, 160, 112, 224, 24, 64],dro))
net.update(build_inception_module('inception_4c',
net['inception_4b/output'],
[64, 128, 128, 256, 24, 64],dro))
net.update(build_inception_module('inception_4d',
net['inception_4c/output'],
[64, 112, 144, 288, 32, 64],dro))
net.update(build_inception_module('inception_4e',
net['inception_4d/output'],
[128, 256, 160, 320, 32, 128],dro))
net['pool4/3x3_s2'] = PoolLayer(
net['inception_4e/output'], pool_size=3, stride=2, ignore_border=False)
net.update(build_inception_module('inception_5a',
net['pool4/3x3_s2'],
[128, 256, 160, 320, 32, 128]))
net.update(build_inception_module('inception_5b',
net['inception_5a/output'],
[128, 384, 192, 384, 48, 128]))
net['pool5/7x7_s1'] = GlobalPoolLayer(net['inception_5b/output'])
net['pool5/7x7_s1_dropout'] = DropoutLayer(net['pool5/7x7_s1'], p=dro)
net['loss3/classifier'] = DenseLayer(net['pool5/7x7_s1_dropout'],
num_units=1,
nonlinearity=linear)
net['prob'] = NonlinearityLayer(net['loss3/classifier'],
nonlinearity=sigmoid)
return net
def build_cnn(input_var=None):
network = lasagne.layers.InputLayer(shape=(BATCHSIZE, 3, 256, 256),
input_var=input_var)
network = Conv2DLayer(network, num_filters=64, filter_size=7, stride=2, pad=3)
network = lasagne.layers.MaxPool2DLayer(network,pool_size=3, stride=2, ignore_border=False)
network= LRNLayer(network, alpha=0.00002, k=1)
network = lasagne.layers.NINLayer(network, num_units=64, W=lasagne.init.Orthogonal(1), b=lasagne.init.Constant(0))
network = Conv2DLayer(network, 192, 3, pad=1)
network = LRNLayer(network, alpha=0.00002, k=1)
network = lasagne.layers.MaxPool2DLayer(network, pool_size=3, stride=2)
network = inception_module(
network,pool_filters=32, num_1x1=64, reduce_3x3=96, num_3x3=128, reduce_5x5=16, num_5x5=32)
network = inception_module(
network,pool_filters=64, num_1x1=128, reduce_3x3=128, num_3x3=192, reduce_5x5=32, num_5x5=96)
network = lasagne.layers.MaxPool2DLayer(network, pool_size=3, stride=2)
network = inception_module(
network,pool_filters=64, num_1x1=192, reduce_3x3=96, num_3x3=208, reduce_5x5=16, num_5x5=48)
network = inception_module(
network,pool_filters=64, num_1x1=160, reduce_3x3=112, num_3x3=224, reduce_5x5=24, num_5x5=64)
network = inception_module(
network,pool_filters=64, num_1x1=128, reduce_3x3=128, num_3x3=256, reduce_5x5=24, num_5x5=64)
network = inception_module(
network,pool_filters=64, num_1x1=112, reduce_3x3=144, num_3x3=288, reduce_5x5=32, num_5x5=64)
network = inception_module(
network,pool_filters=128, num_1x1=256, reduce_3x3=160, num_3x3=320, reduce_5x5=32, num_5x5=128)
network = lasagne.layers.MaxPool2DLayer(network, pool_size=3, stride=2)
network = inception_module(
network,pool_filters=128, num_1x1=256, reduce_3x3=160, num_3x3=320, reduce_5x5=32, num_5x5=128)
network = inception_module(
network,pool_filters=128, num_1x1=384, reduce_3x3=192, num_3x3=384, reduce_5x5=48, num_5x5=128)
network = GlobalPoolLayer(network)
network = lasagne.layers.DenseLayer(
lasagne.layers.dropout(network, p=.4),
num_units=344,
nonlinearity=lasagne.nonlinearities.linear)
network = lasagne.layers.NonlinearityLayer(network,nonlinearity=lasagne.nonlinearities.softmax)
return network
def build_model_LeNet(input_var=None):
net = {}
net['input'] = InputLayer(shape=(None, 1, 224, 224), input_var=input_var)
net['conv1/7x7_s2'] = ConvLayer(net['input'], 64, 7, stride=2, pad=3)
net['pool1/3x3_s2'] = PoolLayer(net['conv1/7x7_s2'],
pool_size=3,
stride=2,
ignore_border=False)
net['pool1/norm1'] = LRNLayer(net['pool1/3x3_s2'], alpha=0.00002, k=1)
net['conv2/3x3_reduce'] = ConvLayer(net['pool1/norm1'], 64, 1)
net['conv2/3x3'] = ConvLayer(net['conv2/3x3_reduce'], 192, 3, pad=1)
net['conv2/norm2'] = LRNLayer(net['conv2/3x3'], alpha=0.00002, k=1)
net['pool2/3x3_s2'] = PoolLayer(net['conv2/norm2'], pool_size=3, stride=2)
net.update(
build_inception_module('inception_3a', net['pool2/3x3_s2'],
[32, 64, 96, 128, 16, 32]))
net.update(
build_inception_module('inception_3b', net['inception_3a/output'],
[64, 128, 128, 192, 32, 96]))
net['pool3/3x3_s2'] = PoolLayer(net['inception_3b/output'],
pool_size=3,
stride=2)
net.update(
build_inception_module('inception_4a', net['pool3/3x3_s2'],
[64, 192, 96, 208, 16, 48]))
net.update(
build_inception_module('inception_4b', net['inception_4a/output'],
[64, 160, 112, 224, 24, 64]))
net.update(
build_inception_module('inception_4c', net['inception_4b/output'],
[64, 128, 128, 256, 24, 64]))
net.update(
build_inception_module('inception_4d', net['inception_4c/output'],
[64, 112, 144, 288, 32, 64]))
net.update(
build_inception_module('inception_4e', net['inception_4d/output'],
[128, 256, 160, 320, 32, 128]))
net['pool4/3x3_s2'] = PoolLayer(net['inception_4e/output'],
pool_size=3,
stride=2)
net.update(
build_inception_module('inception_5a', net['pool4/3x3_s2'],
[128, 256, 160, 320, 32, 128]))
net.update(
build_inception_module('inception_5b', net['inception_5a/output'],
[128, 384, 192, 384, 48, 128]))
net['pool5/7x7_s1'] = GlobalPoolLayer(net['inception_5b/output'])
net['loss3/classifier'] = DenseLayer(net['pool5/7x7_s1'],
num_units=num_categories,
nonlinearity=linear)
net['prob'] = NonlinearityLayer(net['loss3/classifier'],
nonlinearity=softmax)
#print(net)
return net['prob']
def CNN_model():
net = {}
net['input'] = InputLayer((None, 3, 64, 64))
net['conv1/3x3_s1'] = ConvLayer(net['input'], 16, 3, stride=2, pad='same', W=lasagne.init.HeNormal(gain='relu')) # 16*64*64
# net['pool1/3x3_s2'] = PoolLayer(net['conv1/3x3_s1'], pool_size=3, stride=2, ignore_border=False) # 16*32*32
net['pool1/norm1'] = LRNLayer(net['conv1/3x3_s2'], alpha=0.00002, k=1)
net['conv2/3x3_s1'] = ConvLayer(net['pool1/norm1'], 32, 3, stride=1, pad='same', W=lasagne.init.HeNormal(gain='relu')) # 32*32*32
net['pool2/3x3_s2'] = PoolLayer(net['conv2/3x3_s1'], pool_size=3, stride=2, ignore_border=False) # 32*16*16
net['pool2/norm1'] = LRNLayer(net['pool2/3x3_s2'], alpha=0.00002, k=1)
net['conv3/3x3_s1'] = ConvLayer(net['pool2/norm1'], 64, 3, stride=1, pad='same', W=lasagne.init.HeNormal(gain='relu')) # 64*16*16
net['pool3/3x3_s2'] = PoolLayer(net['conv3/3x3_s1'], pool_size=3, stride=2, ignore_border=False) # 64*8*8
net['pool3/norm1'] = LRNLayer(net['pool3/3x3_s2'], alpha=0.00002, k=1)
net['conv4/3x3_s1'] = ConvLayer(net['pool3/norm1'], 128, 3, stride=1, pad='same', W=lasagne.init.HeNormal(gain='relu')) # 128*8*8
net['pool4/3x3_s2'] = PoolLayer(net['conv4/3x3_s1'], pool_size=3, stride=2, ignore_border=False) # 128*4*4
net['pool4/norm1'] = LRNLayer(net['pool4/3x3_s2'], alpha=0.00002, k=1)
net['conv5/3x3_s1'] = ConvLayer(net['pool4/norm1'], 128, 3, stride=1, pad='same', W=lasagne.init.HeNormal(gain='relu'))
net['pool5/norm1'] = LRNLayer(net['conv5/3x3_s1'], alpha=0.00002, k=1)
net['conv6/3x3_s1'] = ConvLayer(net['pool5/norm1'], 128, 3, stride=1, pad='same', W=lasagne.init.HeNormal(gain='relu'))
net['pool6/norm1'] = LRNLayer(net['conv6/3x3_s1'], alpha=0.00002, k=1)
net['conv7/3x3_s1'] = ConvLayer(net['pool6/norm1'], 128, 3, stride=1, pad='same', W=lasagne.init.HeNormal(gain='relu'))
net['pool7/norm1'] = LRNLayer(net['conv7/3x3_s1'], alpha=0.00002, k=1)
net['conv8/3x3_s1'] = ConvLayer(net['pool7/norm1'], 256, 3, stride=1, pad='same', W=lasagne.init.HeNormal(gain='relu')) # 256*4*4
net['pool8/4x4_s1'] = GlobalPoolLayer(net['conv8/3x3_s1']); # 256
return net
def build_model(input_shape):
net = {}
net['input'] = InputLayer(input_shape)
net['conv1'] = ConvLayer(net['input'],
num_filters=96,
filter_size=7,
stride=2,
flip_filters=False)
net['norm1'] = LRNLayer(
net['conv1'], alpha=0.0001) # caffe has alpha = alpha * pool_size
net['pool1'] = PoolLayer(net['norm1'],
pool_size=3,
stride=3,
ignore_border=False)
net['conv2'] = ConvLayer(net['pool1'],
num_filters=256,
filter_size=5,
flip_filters=False)
net['pool2'] = PoolLayer(net['conv2'],
pool_size=2,
stride=2,
ignore_border=False)
net['conv3'] = ConvLayer(net['pool2'],
num_filters=512,
filter_size=3,
pad=1,
flip_filters=False)
net['conv4'] = ConvLayer(net['conv3'],
num_filters=512,
filter_size=3,
pad=1,
flip_filters=False)
net['conv5'] = ConvLayer(net['conv4'],
num_filters=512,
filter_size=3,
pad=1,
flip_filters=False)
net['pool5'] = PoolLayer(net['conv5'],
pool_size=3,
stride=3,
ignore_border=False)
net['fc6'] = DenseLayer(net['pool5'], num_units=4096)
net['drop6'] = DropoutLayer(net['fc6'], p=0.5)
net['fc7'] = DenseLayer(net['drop6'], num_units=4096)
net['drop7'] = DropoutLayer(net['fc7'], p=0.5)
net['fc8'] = DenseLayer(net['drop7'],
num_units=1000,
nonlinearity=lasagne.nonlinearities.softmax)
for layer in net.values():
print str(lasagne.layers.get_output_shape(layer))
return net
def build_model(input_var):
from lasagne.layers import InputLayer
from lasagne.layers import DenseLayer
from lasagne.layers import ConcatLayer
from lasagne.layers import NonlinearityLayer
from lasagne.layers import GlobalPoolLayer
from lasagne.layers.dnn import Conv2DDNNLayer as ConvLayer
from lasagne.layers.dnn import MaxPool2DDNNLayer as PoolLayerDNN
from lasagne.layers import MaxPool2DLayer as PoolLayer
from lasagne.layers import LocalResponseNormalization2DLayer as LRNLayer
from lasagne.nonlinearities import softmax, linear
def build_inception_module(name, input_layer, nfilters):
# nfilters: (pool_proj, 1x1, 3x3_reduce, 3x3, 5x5_reduce, 5x5)
net = dict()
net['pool'] = PoolLayerDNN(input_layer, pool_size=3, stride=1, pad=1)
net['pool_proj'] = ConvLayer(net['pool'],
nfilters[0],
1,
flip_filters=False)
net['1x1'] = ConvLayer(input_layer, nfilters[1], 1, flip_filters=False)
net['3x3_reduce'] = ConvLayer(input_layer,
nfilters[2],
1,
flip_filters=False)
net['3x3'] = ConvLayer(net['3x3_reduce'],
nfilters[3],
3,
pad=1,
flip_filters=False)
net['5x5_reduce'] = ConvLayer(input_layer,
nfilters[4],
1,
flip_filters=False)
net['5x5'] = ConvLayer(net['5x5_reduce'],
nfilters[5],
5,
pad=2,
flip_filters=False)
net['output'] = ConcatLayer(
[net['1x1'], net['3x3'], net['5x5'], net['pool_proj']])
return {'{}/{}'.format(name, k): v for k, v in net.items()}
net = dict()
net['input'] = InputLayer((None, 3, None, None), input_var)
net['conv1/7x7_s2'] = ConvLayer(net['input'],
64,
7,
stride=2,
pad=3,
flip_filters=False)
net['pool1/3x3_s2'] = PoolLayer(net['conv1/7x7_s2'],
pool_size=3,
stride=2,
ignore_border=False)
net['pool1/norm1'] = LRNLayer(net['pool1/3x3_s2'], alpha=0.00002, k=1)
net['conv2/3x3_reduce'] = ConvLayer(net['pool1/norm1'],
64,
1,
flip_filters=False)
net['conv2/3x3'] = ConvLayer(net['conv2/3x3_reduce'],
192,
3,
pad=1,
flip_filters=False)
net['conv2/norm2'] = LRNLayer(net['conv2/3x3'], alpha=0.00002, k=1)
net['pool2/3x3_s2'] = PoolLayer(net['conv2/norm2'],
pool_size=3,
stride=2,
ignore_border=False)
net.update(
build_inception_module('inception_3a', net['pool2/3x3_s2'],
[32, 64, 96, 128, 16, 32]))
net.update(
build_inception_module('inception_3b', net['inception_3a/output'],
[64, 128, 128, 192, 32, 96]))
net['pool3/3x3_s2'] = PoolLayer(net['inception_3b/output'],
pool_size=3,
stride=2,
ignore_border=False)
net.update(
build_inception_module('inception_4a', net['pool3/3x3_s2'],
[64, 192, 96, 208, 16, 48]))
net.update(
build_inception_module('inception_4b', net['inception_4a/output'],
[64, 160, 112, 224, 24, 64]))
net.update(
build_inception_module('inception_4c', net['inception_4b/output'],
[64, 128, 128, 256, 24, 64]))
net.update(
build_inception_module('inception_4d', net['inception_4c/output'],
[64, 112, 144, 288, 32, 64]))
net.update(
build_inception_module('inception_4e', net['inception_4d/output'],
[128, 256, 160, 320, 32, 128]))
net['pool4/3x3_s2'] = PoolLayer(net['inception_4e/output'],
pool_size=3,
stride=2,
ignore_border=False)
net.update(
build_inception_module('inception_5a', net['pool4/3x3_s2'],
[128, 256, 160, 320, 32, 128]))
net.update(
build_inception_module('inception_5b', net['inception_5a/output'],
[128, 384, 192, 384, 48, 128]))
net['pool5/7x7_s1'] = GlobalPoolLayer(net['inception_5b/output'])
net['loss3/classifier'] = DenseLayer(net['pool5/7x7_s1'],
num_units=1000,
nonlinearity=linear)
net['prob'] = NonlinearityLayer(net['loss3/classifier'],
nonlinearity=softmax)
return net
def build_model(self, incoming, out_size):
net = {}
#net['input'] = InputLayer((batch_size, 3, 224, 224))
net['conv1/7x7_s2'] = ConvLayer(incoming,
64,
7,
stride=2,
pad=3,
flip_filters=False)
net['pool1/3x3_s2'] = PoolLayer(net['conv1/7x7_s2'],
pool_size=3,
stride=2,
ignore_border=False)
net['pool1/norm1'] = LRNLayer(net['pool1/3x3_s2'], alpha=0.00002, k=1)
net['conv2/3x3_reduce'] = ConvLayer(net['pool1/norm1'],
64,
1,
flip_filters=False)
net['conv2/3x3'] = ConvLayer(net['conv2/3x3_reduce'],
192,
3,
pad=1,
flip_filters=False)
net['conv2/norm2'] = LRNLayer(net['conv2/3x3'], alpha=0.00002, k=1)
net['pool2/3x3_s2'] = PoolLayer(net['conv2/norm2'],
pool_size=3,
stride=2,
ignore_border=False)
net.update(
self.build_inception_module('inception_3a', net['pool2/3x3_s2'],
[32, 64, 96, 128, 16, 32]))
net.update(
self.build_inception_module('inception_3b',
net['inception_3a/output'],
[64, 128, 128, 192, 32, 96]))
net['pool3/3x3_s2'] = PoolLayer(net['inception_3b/output'],
pool_size=3,
stride=2,
ignore_border=False)
net.update(
self.build_inception_module('inception_4a', net['pool3/3x3_s2'],
[64, 192, 96, 208, 16, 48]))
net.update(
self.build_inception_module('inception_4b',
net['inception_4a/output'],
[64, 160, 112, 224, 24, 64]))
net.update(
self.build_inception_module('inception_4c',
net['inception_4b/output'],
[64, 128, 128, 256, 24, 64]))
net.update(
self.build_inception_module('inception_4d',
net['inception_4c/output'],
[64, 112, 144, 288, 32, 64]))
net.update(
self.build_inception_module('inception_4e',
net['inception_4d/output'],
[128, 256, 160, 320, 32, 128]))
net['pool4/3x3_s2'] = PoolLayer(net['inception_4e/output'],
pool_size=3,
stride=2,
ignore_border=False)
net.update(
self.build_inception_module('inception_5a', net['pool4/3x3_s2'],
[128, 256, 160, 320, 32, 128]))
net.update(
self.build_inception_module('inception_5b',
net['inception_5a/output'],
[128, 384, 192, 384, 48, 128]))
net['pool5/7x7_s1'] = GlobalPoolLayer(net['inception_5b/output'])
net['loss3/classifier'] = DenseLayer(net['pool5/7x7_s1'],
num_units=1000,
nonlinearity=linear)
net['prob'] = NonlinearityLayer(net['loss3/classifier'],
nonlinearity=softmax)
return net
def alexnet(library):
if library == 'keras':
model = Sequential()
# First convolutional layer
model.add(Convolution2D(96, 11, 11, border_mode='valid', subsample=(4, 4),
input_shape=(3, 224, 224)))
model.add(Activation('relu'))
model.add(BatchNormalization(alpha=0.000002, k=1))
model.add(MaxPooling2D(pool_size=(3, 3), strides=(2, 2)))
# Second convolutional layer
model.add(Convolution2D(256, 5, 5, border_mode='valid'))
model.add(Activation('relu'))
model.add(BatchNormalization(alpha=0.000002, k=1))
model.add(MaxPooling2D(pool_size=(3, 3), strides=(2, 2)))
# Third convolutional layer
model.add(Convolution2D(384, 3, 3, border_mode='valid'))
model.add(Activation('relu'))
# Fourth convolutional layer
model.add(Convolution2D(384, 3, 3, border_mode='valid'))
model.add(Activation('relu'))
# Fifth convolutional layer
model.add(Convolution2D(256, 3, 3, border_mode='valid'))
model.add(Activation('relu'))
# First fully connected layer
model.add(Flatten())
model.add(Dense(4096))
model.add(Activation('relu'))
# Second fully connected layer
model.add(Dense(4096))
model.add(Activation('relu'))
# Softmax prediction layer
model.add(Dense(1000))
model.add(Activation('softmax'))
return model
else:
net = {}
net['input'] = InputLayer((3, 244, 244))
# First convolutional layer
net['conv1'] = ConvLayer(
net['input'], 96, 11, stride=(4, 4), flip_filters=False)
net['norm1'] = LRNLayer(net['conv1'])
net['pool1'] = PoolLayer(net['norm1'], 3, stride=2)
# Second convolutional layer
net['conv2'] = ConvLayer(net['pool1'], 256, 5, flip_filters=False)
net['norm2'] = LRNLayer(net['conv2'])
net['pool2'] = PoolLayer(net['norm2'], 3, stride=2)
# Third convolutional layer
net['conv3'] = ConvLayer(net['pool2'], 384, 3, flip_filters=False)
# Fourth convolutional layer
net['conv4'] = ConvLayer(net['conv3'], 384, 3, flip_filters=False)
# Fifth convolutional layer
net['conv5'] = ConvLayer(net['conv4'], 256, 3, flip_filters=False)
# First fully connected layer
net['fc1'] = DenseLayer(net['conv5'], num_units=4096)
# Second fully connected layer
net['fc2'] = DenseLayer(net['fc1'], num_units=4096)
# Softmax prediction layer
net['fc3'] = DenseLayer(net['fc2'], num_units=1000, nonlinearity=None)
net['prob'] = NonlinearityLayer(net['fc3'], softmax)
return net
def googlenet(library):
if library == 'keras':
model = Sequential()
# First convolutional layer
model.add(ZeroPadding2D(padding=(3, 3), input_shape=(3, 244, 244)))
model.add(Convolution2D(64, 7, 7, subsample=(2, 2)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(3, 3), strides=(2, 2)))
model.add(BatchNormalization(alpha=0.000002, k=1))
# Second convolutional layer
model.add(Convolution2D(64, 1, 11))
model.add(Activation('relu'))
model.add(ZeroPadding2D(padding=(1, 1)))
model.add(Convolution2D(192, 3, 3))
model.add(Activation('relu'))
model.add(BatchNormalization(alpha=0.000002, k=1))
model.add(MaxPooling2D(pool_size=(3, 3), strides=(2, 2)))
# First set of inception modules
create_inception_module_keras(model, [32, 64, 96, 128, 16, 32])
create_inception_module_keras(model, [64, 128, 128, 192, 32, 96])
model.add(MaxPooling2D(pool_size=(3, 3), strides=(2, 2)))
# Second set of inception modules
create_inception_module_keras(model, [64, 192, 96, 208, 16, 48])
create_inception_module_keras(model, [64, 160, 112, 224, 24, 64])
create_inception_module_keras(model, [64, 128, 128, 256, 24, 64])
create_inception_module_keras(model, [64, 122, 144, 288, 32, 64])
create_inception_module_keras(model, [128, 256, 160, 320, 32, 128])
model.add(pool_size=(3, 3), strides=(2, 2))
# Third set of inception modules
create_inception_module_keras(model, [128, 256, 160, 320, 32, 128])
create_inception_module_keras(model, [128, 384, 192, 384, 48, 128])
# Average pooling layer
model.add(AveragePooling2D())
# Softmax prediction layer
model.add(Flatten())
model.add(Dense(1000))
model.add(Activation('softmax'))
return model
else:
net = {}
net['input'] = InputLayer((None, 3, 244, 244))
# First convolutional layer
net['conv1/7x7_s2'] = ConvLayer(
net['input'], 64, 7, stride=2, pad=3, flip_filters=False)
net['pool1/3x3_s2'] = PoolLayer(
net['conv1/7x7_s2'], pool_size=3, stride=2, ignore_border=False)
net['pool1/norm1'] = LRNLayer(net['pool1/3x3_s2'], alpha=0.000002, k=1)
# Second convolutional layer
net['conv2/3x3_reduce'] = ConvLayer(
net['pool1/norm1'], 64, 1, flip_filters=False)
net['conv2/3x3'] = ConvLayer(
net['conv2/3x3_reduce'], 192, 3, pad=1, flip_filters=False)
net['conv2/norm2'] = LRNLayer(net['conv2/3x3'], alpha=0.00002, k=1)
net['pool2/3x3_s2'] = PoolLayer(
net['conv2/norm2'], pool_size=3, stride=2, ignore_border=False)
# First set of inception modules
net.update(create_inception_module_lasagne('inception_3a',
net['pool2/3x3_s2'],
[32, 64, 96, 128, 16, 32]))
net.update(create_inception_module_lasagne('inception_3b',
net['inception_3a/output'],
[64, 128, 128, 192, 32, 96]))
net['pool3/3x3_s2'] = PoolLayer(
net['inception_3b/output'], pool_size=3, stride=2, ignore_border=False)
# Second set of inception modules
net.update(create_inception_module_lasagne('inception_4a',
net['pool3/3x3_s2'],
[64, 192, 96, 208, 16, 48]))
net.update(create_inception_module_lasagne('inception_4b',
net['inception_4a/output'],
[64, 160, 112, 224, 24, 64]))
net.update(create_inception_module_lasagne('inception_4c',
net['inception_4b/output'],
[64, 128, 128, 256, 24, 64]))
net.update(create_inception_module_lasagne('inception_4d',
net['inception_4c/output'],
[64, 122, 144, 288, 32, 64]))
net.update(create_inception_module_lasagne('inception_4e',
net['inception_4d/output'],
[128, 256, 160, 320, 32, 128]))
net['pool4/3x3_s2'] = PoolLayer(
net['inception_4e/output'], pool_size=3, stride=2, ignore_border=False)
# Third set of inception modules
net.update(create_inception_module_lasagne('inception_5a',
net['pool4/3x3_s2'],
[128, 256, 160, 320, 32, 128]))
net.update(create_inception_module_lasagne('inception_5b',
net['inception_5a/output'],
[128, 384, 192, 384, 48, 128]))
net['pool5/7x7_s1'] = GlobalPoolLayer(net['inception_5b/output'])
net['loss3/classifier'] = DenseLayer(net['pool5/7x7_s1'], num_units=1000,
nonlinearity=None)
net['prob'] = NonlinearityLayer(net['loss3/classifier'],
nonlinearity=softmax)
return net
def build_model(self, input_var, forward, dropout):
net = dict()
net['input'] = InputLayer((None, 3, None, None), input_var=input_var)
net['conv1/7x7_s2'] = ConvLayer(net['input'],
64,
7,
stride=2,
pad=3,
flip_filters=False)
net['pool1/3x3_s2'] = PoolLayer(net['conv1/7x7_s2'],
pool_size=3,
stride=2,
ignore_border=False)
net['pool1/norm1'] = LRNLayer(net['pool1/3x3_s2'], alpha=0.00002, k=1)
net['conv2/3x3_reduce'] = ConvLayer(net['pool1/norm1'],
64,
1,
flip_filters=False)
net['conv2/3x3'] = ConvLayer(net['conv2/3x3_reduce'],
192,
3,
pad=1,
flip_filters=False)
net['conv2/norm2'] = LRNLayer(net['conv2/3x3'], alpha=0.00002, k=1)
net['pool2/3x3_s2'] = PoolLayerDNN(net['conv2/norm2'],
pool_size=3,
stride=2)
net.update(
self.build_inception_module('inception_3a', net['pool2/3x3_s2'],
[32, 64, 96, 128, 16, 32]))
net.update(
self.build_inception_module('inception_3b',
net['inception_3a/output'],
[64, 128, 128, 192, 32, 96]))
net['pool3/3x3_s2'] = PoolLayerDNN(net['inception_3b/output'],
pool_size=3,
stride=2)
net.update(
self.build_inception_module('inception_4a', net['pool3/3x3_s2'],
[64, 192, 96, 208, 16, 48]))
net.update(
self.build_inception_module('inception_4b',
net['inception_4a/output'],
[64, 160, 112, 224, 24, 64]))
net.update(
self.build_inception_module('inception_4c',
net['inception_4b/output'],
[64, 128, 128, 256, 24, 64]))
net.update(
self.build_inception_module('inception_4d',
net['inception_4c/output'],
[64, 112, 144, 288, 32, 64]))
net.update(
self.build_inception_module('inception_4e',
net['inception_4d/output'],
[128, 256, 160, 320, 32, 128]))
net['pool4/3x3_s2'] = PoolLayerDNN(net['inception_4e/output'],
pool_size=3,
stride=2)
net.update(
self.build_inception_module('inception_5a', net['pool4/3x3_s2'],
[128, 256, 160, 320, 32, 128]))
net.update(
self.build_inception_module('inception_5b',
net['inception_5a/output'],
[128, 384, 192, 384, 48, 128]))
net['pool5/7x7_s1'] = GlobalPoolLayer(net['inception_5b/output'])
if forward:
#net['fc6'] = DenseLayer(net['pool5/7x7_s1'], num_units=1000)
net['prob'] = DenseLayer(net['pool5/7x7_s1'],
num_units=4,
nonlinearity=softmax)
else:
net['dropout1'] = DropoutLayer(net['pool5/7x7_s1'], p=dropout)
#net['fc6'] = DenseLayer(net['dropout1'], num_units=1000)
#net['dropout2'] = DropoutLayer(net['fc6'], p=dropout)
net['prob'] = DenseLayer(net['dropout1'],
num_units=4,
nonlinearity=softmax)
return net
def build_model(x):
print 'build googlenet'
net = {}
net['input'] = InputLayer((None, 3, None, None), x)
net['conv1/7x7_s2'] = ConvLayer(
net['input'], 64, 7, stride=2, pad=3, flip_filters=False)
net['pool1/3x3_s2'] = PoolLayer(
net['conv1/7x7_s2'], pool_size=3, stride=2, ignore_border=False)
net['pool1/norm1'] = LRNLayer(net['pool1/3x3_s2'], alpha=0.00002, k=1)
net['conv2/3x3_reduce'] = ConvLayer(
net['pool1/norm1'], 64, 1, flip_filters=False)
net['conv2/3x3'] = ConvLayer(
net['conv2/3x3_reduce'], 192, 3, pad=1, flip_filters=False)
net['conv2/norm2'] = LRNLayer(net['conv2/3x3'], alpha=0.00002, k=1)
net['pool2/3x3_s2'] = PoolLayer(net['conv2/norm2'], pool_size=3, stride=2)
net.update(build_inception_module('inception_3a',
net['pool2/3x3_s2'],
[32, 64, 96, 128, 16, 32]))
net.update(build_inception_module('inception_3b',
net['inception_3a/output'],
[64, 128, 128, 192, 32, 96]))
net['pool3/3x3_s2'] = PoolLayer(net['inception_3b/output'],
pool_size=3, stride=2)
net.update(build_inception_module('inception_4a',
net['pool3/3x3_s2'],
[64, 192, 96, 208, 16, 48]))
net.update(build_inception_module('inception_4b',
net['inception_4a/output'],
[64, 160, 112, 224, 24, 64]))
net.update(build_inception_module('inception_4c',
net['inception_4b/output'],
[64, 128, 128, 256, 24, 64]))
net.update(build_inception_module('inception_4d',
net['inception_4c/output'],
[64, 112, 144, 288, 32, 64]))
net.update(build_inception_module('inception_4e',
net['inception_4d/output'],
[128, 256, 160, 320, 32, 128]))
net['pool4/3x3_s2'] = PoolLayer(net['inception_4e/output'],
pool_size=3, stride=2)
net.update(build_inception_module('inception_5a',
net['pool4/3x3_s2'],
[128, 256, 160, 320, 32, 128]))
net.update(build_inception_module('inception_5b',
net['inception_5a/output'],
[128, 384, 192, 384, 48, 128]))
net['pool5/7x7_s1'] = GlobalPoolLayer(net['inception_5b/output'])
net['loss3/classifier'] = DenseLayer(net['pool5/7x7_s1'],
num_units=1000,
nonlinearity=linear)
net['prob'] = NonlinearityLayer(net['loss3/classifier'],
nonlinearity=softmax)
'''
download from
https://s3.amazonaws.com/lasagne/recipes/pretrained/imagenet/blvc_googlenet.pkl
'''
file = open(googlenet_file, 'r')
vals = pickle.load(file)
values = pickle.load(
open(googlenet_file))['param values']
lasagne.layers.set_all_param_values(
net['prob'], [v.astype(np.float32) for v in values])
return net, vals['synset words']
def build_model(data_size, num_classes, batch_norm=True):
net = {}
net['input'] = InputLayer(data_size)
net['conv1/7x7_s2'] = ConvLayer(net['input'],
64,
7,
stride=2,
pad=3,
flip_filters=False)
if batch_norm:
net['conv1/7x7_s2'] = normalization.batch_norm(net['conv1/7x7_s2'])
net['pool1/3x3_s2'] = PoolLayer(net['conv1/7x7_s2'],
pool_size=3,
stride=2,
ignore_border=False)
net['pool1/norm1'] = LRNLayer(net['pool1/3x3_s2'], alpha=0.00002, k=1)
net['conv2/3x3_reduce'] = ConvLayer(net['pool1/norm1'],
64,
1,
flip_filters=False)
if batch_norm:
net['conv2/3x3_reduce'] = normalization.batch_norm(
net['conv2/3x3_reduce'])
net['conv2/3x3'] = ConvLayer(net['conv2/3x3_reduce'],
192,
3,
pad=1,
flip_filters=False)
if batch_norm:
net['conv2/3x3'] = normalization.batch_norm(net['conv2/3x3'])
net['conv2/norm2'] = LRNLayer(net['conv2/3x3'], alpha=0.00002, k=1)
net['pool2/3x3_s2'] = PoolLayer(net['conv2/norm2'],
pool_size=3,
stride=2,
ignore_border=False)
net.update(
build_inception_module('inception_3a', net['pool2/3x3_s2'],
[32, 64, 96, 128, 16, 32], batch_norm))
net.update(
build_inception_module('inception_3b', net['inception_3a/output'],
[64, 128, 128, 192, 32, 96], batch_norm))
net['pool3/3x3_s2'] = PoolLayer(net['inception_3b/output'],
pool_size=3,
stride=2,
ignore_border=False)
net.update(
build_inception_module('inception_4a', net['pool3/3x3_s2'],
[64, 192, 96, 208, 16, 48], batch_norm))
net.update(
build_inception_module('inception_4b', net['inception_4a/output'],
[64, 160, 112, 224, 24, 64], batch_norm))
net.update(
build_inception_module('inception_4c', net['inception_4b/output'],
[64, 128, 128, 256, 24, 64], batch_norm))
net.update(
build_inception_module('inception_4d', net['inception_4c/output'],
[64, 112, 144, 288, 32, 64], batch_norm))
net.update(
build_inception_module('inception_4e', net['inception_4d/output'],
[128, 256, 160, 320, 32, 128], batch_norm))
net['pool4/3x3_s2'] = PoolLayer(net['inception_4e/output'],
pool_size=3,
stride=2,
ignore_border=False)
net.update(
build_inception_module('inception_5a', net['pool4/3x3_s2'],
[128, 256, 160, 320, 32, 128], batch_norm))
net.update(
build_inception_module('inception_5b', net['inception_5a/output'],
[128, 384, 192, 384, 48, 128], batch_norm))
net['pool5/7x7_s1'] = GlobalPoolLayer(net['inception_5b/output'])
net['dropout'] = DropoutLayer(net['pool5/7x7_s1'], p=0.4)
net['loss3/classifier'] = DenseLayer(net['dropout'],
num_units=1000,
nonlinearity=linear)
if batch_norm:
net['loss3/classifier'] = normalization.batch_norm(
net['loss3/classifier'])
net['output'] = DenseLayer(net['loss3/classifier'],
num_units=num_classes,
nonlinearity=softmax)
if batch_norm:
net['output'] = normalization.batch_norm(net['output'])
return net
def CNN_model():
net = {}
net['input'] = InputLayer((None, 3, 64, 64))
net['conv0/3x3_s1'] = ConvLayer(
net['input'],
16,
3,
stride=1,
pad='same',
W=lasagne.init.HeNormal(gain='relu')) # 16*64*64
net['conv0/norm1'] = LRNLayer(net['conv0/3x3_s1'], alpha=0.00002, k=1)
net['conv1/3x3_s1'] = ConvLayer(
net['conv0/norm1'],
32,
3,
stride=2,
pad='same',
W=lasagne.init.HeNormal(gain='relu')) # 32*32*32
net['conv1/norm1'] = LRNLayer(net['conv1/3x3_s1'], alpha=0.00002, k=1)
net['conv2/3x3_s1'] = ConvLayer(
net['conv1/norm1'],
32,
3,
stride=1,
pad='same',
W=lasagne.init.HeNormal(gain='relu')) # 32*32*32
net['conv2/norm1'] = LRNLayer(net['conv2/3x3_s1'], alpha=0.00002, k=1)
net['conv3/3x3_s1'] = ConvLayer(
net['conv2/norm1'],
64,
3,
stride=2,
pad='same',
W=lasagne.init.HeNormal(gain='relu')) # 64*16*16
net['conv3/norm1'] = LRNLayer(net['conv3/3x3_s1'], alpha=0.00002, k=1)
net['conv4/3x3_s1'] = ConvLayer(
net['conv3/norm1'],
64,
3,
stride=1,
pad='same',
W=lasagne.init.HeNormal(gain='relu')) # 64*16*16
net['conv4/norm1'] = LRNLayer(net['conv4/3x3_s1'], alpha=0.00002, k=1)
net['conv5/3x3_s1'] = ConvLayer(
net['conv4/norm1'],
128,
3,
stride=2,
pad='same',
W=lasagne.init.HeNormal(gain='relu')) # 128*8*8
net['conv5/norm1'] = LRNLayer(net['conv5/3x3_s1'], alpha=0.00002, k=1)
net['conv6/3x3_s1'] = ConvLayer(
net['conv5/norm1'],
128,
3,
stride=1,
pad='same',
W=lasagne.init.HeNormal(gain='relu')) # 128*8*8
net['conv6/norm1'] = LRNLayer(net['conv6/3x3_s1'], alpha=0.00002, k=1)
net['conv7/3x3_s1'] = ConvLayer(
net['conv6/norm1'],
256,
3,
stride=2,
pad='same',
W=lasagne.init.HeNormal(gain='relu')) # 256*4*4
net['conv7/norm1'] = LRNLayer(net['conv7/3x3_s1'], alpha=0.00002, k=1)
net['conv8/3x3_s1'] = ConvLayer(
net['conv7/norm1'],
256,
3,
stride=1,
pad='same',
W=lasagne.init.HeNormal(gain='relu')) # 256*4*4
net['conv8/norm1'] = LRNLayer(net['conv8/3x3_s1'], alpha=0.00002, k=1)
net['conv9/3x3_s1'] = ConvLayer(
net['conv8/norm1'],
512,
3,
stride=2,
pad='same',
W=lasagne.init.HeNormal(gain='relu')) # 512*2*2
net['conv9/norm1'] = LRNLayer(net['conv9/3x3_s1'], alpha=0.00002, k=1)
net['conv10/3x3_s1'] = ConvLayer(
net['conv9/norm1'],
512,
3,
stride=1,
pad='same',
W=lasagne.init.HeNormal(gain='relu')) # 512*2*2
net['conv10/norm1'] = LRNLayer(net['conv10/3x3_s1'], alpha=0.00002, k=1)
net['conv11/3x3_s1'] = ConvLayer(
net['conv10/norm1'],
1024,
3,
stride=2,
pad='same',
W=lasagne.init.HeNormal(gain='relu')) # 1024*1*1
# net['conv11/norm1'] = LRNLayer(net['conv11/3x3_s1'], alpha=0.00002, k=1)
return net
def build_model_vgg_cnn_s(input_shape, verbose):
'''
See Lasagne Modelzoo:
https://github.com/Lasagne/Recipes/blob/master/modelzoo/vgg_cnn_s.py
'''
if verbose: print 'VGG_cnn_s (from lasagne model zoo)'
net = {}
net['input'] = InputLayer(input_shape)
net['conv1'] = ConvLayer(net['input'],
num_filters=96,
filter_size=7,
stride=2,
flip_filters=False)
net['norm1'] = LRNLayer(
net['conv1'], alpha=0.0001) # caffe has alpha = alpha * pool_size
net['pool1'] = PoolLayer(net['norm1'],
pool_size=3,
stride=3,
ignore_border=False)
net['conv2'] = ConvLayer(net['pool1'],
num_filters=256,
filter_size=5,
flip_filters=False)
net['pool2'] = PoolLayer(net['conv2'],
pool_size=2,
stride=2,
ignore_border=False)
net['conv3'] = ConvLayer(net['pool2'],
num_filters=512,
filter_size=3,
pad=1,
flip_filters=False)
net['conv4'] = ConvLayer(net['conv3'],
num_filters=512,
filter_size=3,
pad=1,
flip_filters=False)
net['conv5'] = ConvLayer(net['conv4'],
num_filters=512,
filter_size=3,
pad=1,
flip_filters=False)
net['pool5'] = PoolLayer(net['conv5'],
pool_size=3,
stride=3,
ignore_border=False)
net['fc6'] = DenseLayer(net['pool5'], num_units=4096)
net['drop6'] = DropoutLayer(net['fc6'], p=0.5)
net['fc7'] = DenseLayer(net['drop6'], num_units=4096)
net['drop7'] = DropoutLayer(net['fc7'], p=0.5)
net['fc8'] = DenseLayer(net['drop7'],
num_units=1000,
nonlinearity=lasagne.nonlinearities.softmax)
if verbose:
for layer in net.values():
print str(lasagne.layers.get_output_shape(layer))
return net
def build_model(input,
prefix,
lastClassNum=67,
dropoutratio=0.4,
classificationFlag=False):
net = {}
net[prefix + 'input'] = input
net[prefix + 'conv1/7x7_s2'] = ConvLayer(net[prefix + 'input'],
64,
7,
stride=2,
pad=3,
flip_filters=False,
name=prefix + 'conv1/7x7_s2')
net[prefix + 'pool1/3x3_s2'] = PoolLayer(net[prefix + 'conv1/7x7_s2'],
pool_size=3,
stride=2,
ignore_border=False,
name=prefix + 'pool1/3x3_s2')
net[prefix + 'pool1/norm1'] = LRNLayer(net[prefix + 'pool1/3x3_s2'],
alpha=0.00002,
k=1,
name=prefix + 'pool1/norm1')
net[prefix + 'conv2/3x3_reduce'] = ConvLayer(net[prefix + 'pool1/norm1'],
64,
1,
flip_filters=False,
name=prefix +
'conv2/3x3_reduce')
net[prefix + 'conv2/3x3'] = ConvLayer(net[prefix + 'conv2/3x3_reduce'],
192,
3,
pad=1,
flip_filters=False,
name=prefix + 'conv2/3x3')
net[prefix + 'conv2/norm2'] = LRNLayer(net[prefix + 'conv2/3x3'],
alpha=0.00002,
k=1,
name=prefix + 'conv2/norm2')
net[prefix + 'pool2/3x3_s2'] = PoolLayer(net[prefix + 'conv2/norm2'],
pool_size=3,
stride=2,
ignore_border=False,
name=prefix + 'pool2/3x3_s2')
net.update(
build_inception_module('inception_3a', prefix,
net[prefix + 'pool2/3x3_s2'],
[32, 64, 96, 128, 16, 32]), )
net.update(
build_inception_module('inception_3b', prefix,
net[prefix + 'inception_3a/output'],
[64, 128, 128, 192, 32, 96]))
net[prefix + 'pool3/3x3_s2'] = PoolLayer(net[prefix +
'inception_3b/output'],
pool_size=3,
stride=2,
ignore_border=False,
name=prefix + 'pool3/3x3_s2')
net.update(
build_inception_module('inception_4a', prefix,
net[prefix + 'pool3/3x3_s2'],
[64, 192, 96, 208, 16, 48]))
net.update(
build_inception_module('inception_4b', prefix,
net[prefix + 'inception_4a/output'],
[64, 160, 112, 224, 24, 64]))
net.update(
build_inception_module('inception_4c', prefix,
net[prefix + 'inception_4b/output'],
[64, 128, 128, 256, 24, 64]))
net.update(
build_inception_module('inception_4d', prefix,
net[prefix + 'inception_4c/output'],
[64, 112, 144, 288, 32, 64]))
net.update(
build_inception_module('inception_4e', prefix,
net[prefix + 'inception_4d/output'],
[128, 256, 160, 320, 32, 128]))
net[prefix + 'pool4/3x3_s2'] = PoolLayer(net[prefix +
'inception_4e/output'],
pool_size=3,
stride=2,
ignore_border=False,
name=prefix + 'pool4/3x3_s2')
net.update(
build_inception_module('inception_5a', prefix,
net[prefix + 'pool4/3x3_s2'],
[128, 256, 160, 320, 32, 128]))
net.update(
build_inception_module('inception_5b', prefix,
net[prefix + 'inception_5a/output'],
[128, 384, 192, 384, 48, 128]))
net[prefix + 'pool5/7x7_s1'] = GlobalPoolLayer(
net[prefix + 'inception_5b/output'], name=prefix + 'pool5/7x7_s1')
net[prefix + 'dropout'] = lasagne.layers.DropoutLayer(
net[prefix + 'pool5/7x7_s1'], p=dropoutratio, name=prefix + 'dropout')
if classificationFlag == True:
net[prefix + 'loss3/classifier'] = DenseLayer(net[prefix + 'dropout'],
num_units=lastClassNum,
nonlinearity=None,
name=prefix +
'loss3/classifier')
net[prefix + 'prob'] = NonlinearityLayer(net[prefix +
'loss3/classifier'],
nonlinearity=softmax,
name=prefix + 'prob')
return net
def build_model():
net = {}
net['input'] = InputLayer((None, 3, image_size, image_size))
net['conv1/7x7_s2'] = ConvLayer(net['input'],
64,
7,
stride=2,
pad=3,
flip_filters=False)
net['pool1/3x3_s2'] = PoolLayer(net['conv1/7x7_s2'],
pool_size=3,
stride=2,
ignore_border=False)
net['pool1/norm1'] = LRNLayer(net['pool1/3x3_s2'], alpha=0.00002, k=1)
net['conv2/3x3_reduce'] = ConvLayer(net['pool1/norm1'],
64,
1,
flip_filters=False)
net['conv2/3x3'] = ConvLayer(net['conv2/3x3_reduce'],
192,
3,
pad=1,
flip_filters=False)
net['conv2/norm2'] = LRNLayer(net['conv2/3x3'], alpha=0.00002, k=1)
net['pool2/3x3_s2'] = PoolLayer(net['conv2/norm2'], pool_size=3, stride=2)
net.update(
build_inception_module('inception_3a', net['pool2/3x3_s2'],
[32, 64, 96, 128, 16, 32]))
net.update(
build_inception_module('inception_3b', net['inception_3a/output'],
[64, 128, 128, 192, 32, 96]))
net['pool3/3x3_s2'] = PoolLayer(net['inception_3b/output'],
pool_size=3,
stride=2)
net.update(
build_inception_module('inception_4a', net['pool3/3x3_s2'],
[64, 192, 96, 208, 16, 48]))
net.update(
build_inception_module('inception_4b', net['inception_4a/output'],
[64, 160, 112, 224, 24, 64]))
net.update(
build_inception_module('inception_4c', net['inception_4b/output'],
[64, 128, 128, 256, 24, 64]))
net.update(
build_inception_module('inception_4d', net['inception_4c/output'],
[64, 112, 144, 288, 32, 64]))
net.update(
build_inception_module('inception_4e', net['inception_4d/output'],
[128, 256, 160, 320, 32, 128]))
net['pool4/3x3_s2'] = PoolLayer(net['inception_4e/output'],
pool_size=3,
stride=2)
net.update(
build_inception_module('inception_5a', net['pool4/3x3_s2'],
[128, 256, 160, 320, 32, 128]))
net.update(
build_inception_module('inception_5b', net['inception_5a/output'],
[128, 384, 192, 384, 48, 128]))
net['pool5/7x7_s1'] = GlobalPoolLayer(net['inception_5b/output'])
import pickle
model = pickle.load(open(root + 'models/blvc_googlenet.pkl'))
set_all_param_values(net['pool5/7x7_s1'], model['param values'][:114])
net['loss3/classifier'] = DenseLayer(net['pool5/7x7_s1'],
num_units=196,
nonlinearity=linear)
net['prob'] = NonlinearityLayer(net['loss3/classifier'],
nonlinearity=softmax)
model = NeuralNet(
layers=net['prob'],
#use_label_encoder=False,
#objective_l2=1e-4, #1e-3
#update=lasagne.updates.adam,
#update_learning_rate=1e-4,
update=lasagne.updates.nesterov_momentum,
update_momentum=0.9,
update_learning_rate=theano.shared(float32(0.03)), # 1e-4
train_split=TrainSplit(0.1, random_state=42, stratify=False),
#batch_iterator_train=train_iterator,
#batch_iterator_test=test_iterator,
on_epoch_finished=[
save_weights,
save_training_history,
plot_training_history,
early_stopping,
#StepDecay('update_learning_rate', start=1e-2, stop=1e-3)
],
verbose=1,
max_epochs=200,
#custom_score = ('CRPS', CRPS)
)
return model
def build_model():
net = {}
net['input'] = InputLayer((None, 3, None, None))
net['conv1/7x7_s2'] = ConvLayer(net['input'],
64,
7,
stride=2,
pad=3,
flip_filters=False)
net['pool1/3x3_s2'] = PoolLayer(net['conv1/7x7_s2'],
pool_size=3,
stride=2,
ignore_border=False)
net['pool1/norm1'] = LRNLayer(net['pool1/3x3_s2'], alpha=0.00002, k=1)
net['conv2/3x3_reduce'] = ConvLayer(net['pool1/norm1'],
64,
1,
flip_filters=False)
net['conv2/3x3'] = ConvLayer(net['conv2/3x3_reduce'],
192,
3,
pad=1,
flip_filters=False)
net['conv2/norm2'] = LRNLayer(net['conv2/3x3'], alpha=0.00002, k=1)
net['pool2/3x3_s2'] = PoolLayer(net['conv2/norm2'],
pool_size=3,
stride=2,
ignore_border=False)
net.update(
build_inception_module('inception_3a', net['pool2/3x3_s2'],
[32, 64, 96, 128, 16, 32]))
net.update(
build_inception_module('inception_3b', net['inception_3a/output'],
[64, 128, 128, 192, 32, 96]))
net['pool3/3x3_s2'] = PoolLayer(net['inception_3b/output'],
pool_size=3,
stride=2,
ignore_border=False)
net.update(
build_inception_module('inception_4a', net['pool3/3x3_s2'],
[64, 192, 96, 208, 16, 48]))
net.update(
build_inception_module('inception_4b', net['inception_4a/output'],
[64, 160, 112, 224, 24, 64]))
net.update(
build_inception_module('inception_4c', net['inception_4b/output'],
[64, 128, 128, 256, 24, 64]))
net.update(
build_inception_module('inception_4d', net['inception_4c/output'],
[64, 112, 144, 288, 32, 64]))
net.update(
build_inception_module('inception_4e', net['inception_4d/output'],
[128, 256, 160, 320, 32, 128]))
net['pool4/3x3_s2'] = PoolLayer(net['inception_4e/output'],
pool_size=3,
stride=2,
ignore_border=False)
net.update(
build_inception_module('inception_5a', net['pool4/3x3_s2'],
[128, 256, 160, 320, 32, 128]))
net.update(
build_inception_module('inception_5b', net['inception_5a/output'],
[128, 384, 192, 384, 48, 128]))
net['pool5/7x7_s1'] = GlobalPoolLayer(net['inception_5b/output'])
# added a dropout layer (mentioned in the publication, missing in the lasagne-code)
net['pool5_drop_7x7_s1'] = DropoutLayer(net['pool5/7x7_s1'], p=0.5)
net['loss3/classifier'] = DenseLayer(net['pool5_drop_7x7_s1'],
num_units=1000,
nonlinearity=linear)
net['prob'] = NonlinearityLayer(net['loss3/classifier'],
nonlinearity=softmax)
return net
def build_GoogLeNet(width, height):
# Download pretrained weights from:
# https://s3.amazonaws.com/lasagne/recipes/pretrained/imagenet/blvc_googlenet.pkl
net = {}
net['input'] = InputLayer((None, 3, width, height))
net['conv1/7x7_s2'] = ConvLayer(net['input'],
64,
7,
stride=2,
pad=3,
nonlinearity=elu,
flip_filters=False)
net['pool1/3x3_s2'] = PoolLayer(net['conv1/7x7_s2'], pool_size=3, stride=2)
net['pool1/norm1'] = LRNLayer(net['pool1/3x3_s2'], alpha=0.00002, k=1)
net['conv2/3x3_reduce'] = DropoutLayer(ConvLayer(net['pool1/norm1'],
64,
1,
nonlinearity=elu,
flip_filters=False),
p=0.2)
net['conv2/3x3'] = ConvLayer(net['conv2/3x3_reduce'],
192,
3,
pad=1,
nonlinearity=elu,
flip_filters=False)
net['conv2/norm2'] = LRNLayer(net['conv2/3x3'], alpha=0.00002, k=1)
net['pool2/3x3_s2'] = PoolLayer(net['conv2/norm2'], pool_size=3, stride=2)
net.update(
build_inception_module('inception_3a', net['pool2/3x3_s2'],
[32, 64, 96, 128, 16, 32]))
net.update(
build_inception_module('inception_3b', net['inception_3a/output'],
[64, 128, 128, 192, 32, 96]))
net['pool3/3x3_s2'] = PoolLayer(net['inception_3b/output'],
pool_size=3,
stride=2)
net.update(
build_inception_module('inception_4a', net['pool3/3x3_s2'],
[64, 192, 96, 208, 16, 48]))
net.update(
build_inception_module('inception_4b', net['inception_4a/output'],
[64, 160, 112, 224, 24, 64]))
net.update(
build_inception_module('inception_4c', net['inception_4b/output'],
[64, 128, 128, 256, 24, 64]))
net.update(
build_inception_module('inception_4d', net['inception_4c/output'],
[64, 112, 144, 288, 32, 64]))
net.update(
build_inception_module('inception_4e', net['inception_4d/output'],
[128, 256, 160, 320, 32, 128]))
net['pool4/3x3_s2'] = PoolLayer(net['inception_4e/output'],
pool_size=3,
stride=2)
net.update(
build_inception_module('inception_5a', net['pool4/3x3_s2'],
[128, 256, 160, 320, 32, 128]))
net.update(
build_inception_module('inception_5b', net['inception_5a/output'],
[128, 384, 192, 384, 48, 128]))
net['pool5/7x7_s1'] = GlobalPoolLayer(net['inception_5b/output'])
net['loss3/classifier'] = DenseLayer(net['pool5/7x7_s1'],
num_units=2,
nonlinearity=linear)
net['softmax'] = NonlinearityLayer(net['loss3/classifier'],
nonlinearity=softmax)
return net
