def layer(self, input_layer):
from lasagne.layers.normalization import\
LocalResponseNormalization2DLayer
layer = LocalResponseNormalization2DLayer(input_layer,
alpha=1.5,
k=2,
beta=0.75,
n=5)
return layer
def test_even_n_fails(self, input_layer):
from lasagne.layers.normalization import\
LocalResponseNormalization2DLayer
with pytest.raises(NotImplementedError):
LocalResponseNormalization2DLayer(input_layer, n=4)
def googlenet(batch_shape):
"""
Create a googlenet, with the parameters from https://github.com/BVLC/caffe/tree/master/models/bvlc_googlenet
:param batch_shape: The shape of the input images. This should be of size (N, 3, 224, 224)
:return: a struct with the input layer, the logit layer (before the final softmax) and the output layer.
"""
l_in = lasagne.layers.InputLayer(
shape=batch_shape,
name='input',
)
l_conv1 = dnn.Conv2DDNNLayer(
l_in,
num_filters=64,
filter_size=(7, 7),
pad=3,
stride=(2, 2),
nonlinearity=lasagne.nonlinearities.rectify,
name='conv1/7x7_s2',
)
l_pool1 = flip(
dnn.MaxPool2DDNNLayer(
flip(l_conv1),
pool_size=(3, 3), # pool_size
stride=(2, 2),
pad=(1, 1),
name='pool1/3x3_s2'))
lrn = LocalResponseNormalization2DLayer(
l_pool1,
alpha=0.0001 / 5,
beta=0.75,
k=1,
n=5,
name='pool1/norm1',
)
l_conv2 = dnn.Conv2DDNNLayer(
lrn,
num_filters=64,
filter_size=(1, 1),
pad=0,
stride=(1, 1),
nonlinearity=lasagne.nonlinearities.rectify,
name='conv2/3x3_reduce',
)
l_conv2b = dnn.Conv2DDNNLayer(
l_conv2,
num_filters=192,
filter_size=(3, 3),
pad=1,
stride=(1, 1),
nonlinearity=lasagne.nonlinearities.rectify,
name='conv2/3x3',
)
lrn2 = LocalResponseNormalization2DLayer(
l_conv2b,
alpha=0.0001 / 5,
beta=0.75,
k=1,
n=5,
name='conv2/norm2',
)
l_pool2 = flip(
dnn.MaxPool2DDNNLayer(
flip(lrn2),
pool_size=(3, 3), # pool_size
stride=(2, 2),
pad=(1, 1),
name='pool2/3x3_s2'))
def inception(layer,
name,
no_1x1=64,
no_3x3r=96,
no_3x3=128,
no_5x5r=16,
no_5x5=32,
no_pool=32):
l_conv_inc = dnn.Conv2DDNNLayer(
layer,
num_filters=no_1x1,
filter_size=(1, 1),
pad=0,
stride=(1, 1),
nonlinearity=lasagne.nonlinearities.rectify,
name=name + '/1x1',
)
l_conv_inc2 = dnn.Conv2DDNNLayer(
layer,
num_filters=no_3x3r,
filter_size=(1, 1),
pad=0,
stride=(1, 1),
nonlinearity=lasagne.nonlinearities.rectify,
name=name + '/3x3_reduce',
)
l_conv_inc2b = dnn.Conv2DDNNLayer(
l_conv_inc2,
num_filters=no_3x3,
filter_size=(3, 3),
pad=1,
stride=(1, 1),
nonlinearity=lasagne.nonlinearities.rectify,
name=name + '/3x3',
)
l_conv_inc2c = dnn.Conv2DDNNLayer(
layer,
num_filters=no_5x5r,
filter_size=(1, 1),
pad=0,
stride=(1, 1),
nonlinearity=lasagne.nonlinearities.rectify,
name=name + '/5x5_reduce',
)
l_conv_inc2d = dnn.Conv2DDNNLayer(
l_conv_inc2c,
num_filters=no_5x5,
filter_size=(5, 5),
pad=2,
stride=(1, 1),
nonlinearity=lasagne.nonlinearities.rectify,
name=name + '/5x5',
)
l_pool2 = flip(
dnn.MaxPool2DDNNLayer(
flip(layer),
pool_size=(3, 3), # pool_size
stride=(1, 1),
pad=(1, 1),
name=name + '/pool'))
l_conv_inc2e = dnn.Conv2DDNNLayer(
l_pool2,
num_filters=no_pool,
filter_size=(1, 1),
pad=0,
stride=(1, 1),
nonlinearity=lasagne.nonlinearities.rectify,
name=name + '/pool_proj',
)
l_inc_out = nn.layers.concat(
[l_conv_inc, l_conv_inc2b, l_conv_inc2d, l_conv_inc2e])
return l_inc_out
l_inc_3a = inception(l_pool2,
'inception_3a',
no_1x1=64,
no_3x3r=96,
no_3x3=128,
no_5x5r=16,
no_5x5=32,
no_pool=32)
l_inc_3b = inception(l_inc_3a,
'inception_3b',
no_1x1=128,
no_3x3r=128,
no_3x3=192,
no_5x5r=32,
no_5x5=96,
no_pool=64)
l_pool3 = flip(
dnn.MaxPool2DDNNLayer(
flip(l_inc_3b),
pool_size=(3, 3), # pool_size
stride=(2, 2),
pad=1,
name='pool3/3x3_s2'))
l_inc_4a = inception(l_pool3,
'inception_4a',
no_1x1=192,
no_3x3r=96,
no_3x3=208,
no_5x5r=16,
no_5x5=48,
no_pool=64)
l_inc_4b = inception(l_inc_4a,
'inception_4b',
no_1x1=160,
no_3x3r=112,
no_3x3=224,
no_5x5r=24,
no_5x5=64,
no_pool=64)
l_inc_4c = inception(l_inc_4b,
'inception_4c',
no_1x1=128,
no_3x3r=128,
no_3x3=256,
no_5x5r=24,
no_5x5=64,
no_pool=64)
l_inc_4d = inception(l_inc_4c,
'inception_4d',
no_1x1=112,
no_3x3r=144,
no_3x3=288,
no_5x5r=32,
no_5x5=64,
no_pool=64)
l_inc_4e = inception(l_inc_4d,
'inception_4e',
no_1x1=256,
no_3x3r=160,
no_3x3=320,
no_5x5r=32,
no_5x5=128,
no_pool=128)
l_pool4 = flip(
dnn.MaxPool2DDNNLayer(
flip(l_inc_4e),
pool_size=(3, 3), # pool_size
stride=(2, 2),
pad=1,
name='pool4/3x3_s2'))
l_inc_5a = inception(l_pool4,
'inception_5a',
no_1x1=256,
no_3x3r=160,
no_3x3=320,
no_5x5r=32,
no_5x5=128,
no_pool=128)
l_inc_5b = inception(l_inc_5a,
'inception_5b',
no_1x1=384,
no_3x3r=192,
no_3x3=384,
no_5x5r=48,
no_5x5=128,
no_pool=128)
l_pool5 = flip(
dnn.Pool2DDNNLayer(
flip(l_inc_5b),
pool_size=(7, 7), # pool_size
stride=(1, 1),
pad=0,
mode='average',
name='pool5/7x7_s1'))
l_logit = dnn.Conv2DDNNLayer(
l_pool5,
num_filters=1000,
filter_size=(1, 1),
pad=0,
stride=(1, 1),
nonlinearity=lasagne.nonlinearities.linear,
name='prob',
)
l_logit_flat = lasagne.layers.FlattenLayer(l_logit)
l_dense = lasagne.layers.NonlinearityLayer(
l_logit_flat, nonlinearity=lasagne.nonlinearities.softmax)
l_out = l_dense
filename = os.path.join(os.path.dirname(os.path.abspath(__file__)),
'googlenet.pkl')
if os.path.exists(filename):
with open(filename, 'r') as f:
nn.layers.set_all_param_values(l_dense, pickle.load(f))
return utils.struct(
input=l_in,
logit=l_logit,
out=l_out,
interesting=l_inc_4c,
)
def get_googlenet(batchsize=10, n_channels=3, img_height=224, img_width=224):
l_in = lasagne.layers.InputLayer(
shape=(batchsize, n_channels, img_height, img_width),
name='input',
)
l_conv1 = dnn.Conv2DDNNLayer(
l_in,
num_filters=64,
filter_size=(7, 7),
pad=3,
stride=(2, 2),
nonlinearity=lasagne.nonlinearities.rectify,
name='conv1/7x7_s2',
)
l_pool1 = flip(dnn.MaxPool2DDNNLayer(
flip(l_conv1),
pool_size=(3, 3),#pool_size
stride=(2, 2),
pad=(1,1),
name='pool1/3x3_s2'
))
lrn = LocalResponseNormalization2DLayer(
l_pool1,
alpha= 0.0001 / 5,
beta= 0.75,
k=1,
n=5,
name='pool1/norm1',
)
l_conv2 = dnn.Conv2DDNNLayer(
lrn,
num_filters=64,
filter_size=(1, 1),
pad=0,
stride=(1, 1),
nonlinearity=lasagne.nonlinearities.rectify,
name='conv2/3x3_reduce',
)
l_conv2b = dnn.Conv2DDNNLayer(
l_conv2,
num_filters=192,
filter_size=(3, 3),
pad=1,
stride=(1, 1),
nonlinearity=lasagne.nonlinearities.rectify,
name='conv2/3x3',
)
lrn2 = LocalResponseNormalization2DLayer(
l_conv2b,
alpha= 0.0001 / 5,
beta= 0.75,
k=1,
n=5,
name='conv2/norm2',
)
l_pool2 = flip(dnn.MaxPool2DDNNLayer(
flip(lrn2),
pool_size=(3, 3),#pool_size
stride=(2, 2),
pad=(1,1),
name='pool2/3x3_s2'
))
def inception(layer, name, no_1x1=64, no_3x3r=96, no_3x3=128, no_5x5r=16, no_5x5=32, no_pool=32):
l_conv_inc = dnn.Conv2DDNNLayer(
layer,
num_filters=no_1x1,
filter_size=(1, 1),
pad=0,
stride=(1, 1),
nonlinearity=lasagne.nonlinearities.rectify,
name=name+'/1x1',
)
l_conv_inc2 = dnn.Conv2DDNNLayer(
layer,
num_filters=no_3x3r,
filter_size=(1, 1),
pad=0,
stride=(1, 1),
nonlinearity=lasagne.nonlinearities.rectify,
name=name+'/3x3_reduce',
)
l_conv_inc2b = dnn.Conv2DDNNLayer(
l_conv_inc2,
num_filters=no_3x3,
filter_size=(3, 3),
pad=1,
stride=(1, 1),
nonlinearity=lasagne.nonlinearities.rectify,
name=name+'/3x3',
)
l_conv_inc2c = dnn.Conv2DDNNLayer(
layer,
num_filters=no_5x5r,
filter_size=(1, 1),
pad=0,
stride=(1, 1),
nonlinearity=lasagne.nonlinearities.rectify,
name=name+'/5x5_reduce',
)
l_conv_inc2d = dnn.Conv2DDNNLayer(
l_conv_inc2c,
num_filters=no_5x5,
filter_size=(5, 5),
pad=2,
stride=(1, 1),
nonlinearity=lasagne.nonlinearities.rectify,
name=name+'/5x5',
)
l_pool2 = flip(dnn.MaxPool2DDNNLayer(
flip(layer),
pool_size=(3, 3),#pool_size
stride=(1, 1),
pad=(1, 1),
name=name+'/pool'
))
l_conv_inc2e = dnn.Conv2DDNNLayer(
l_pool2,
num_filters=no_pool,
filter_size=(1, 1),
pad=0,
stride=(1, 1),
nonlinearity=lasagne.nonlinearities.rectify,
name=name+'/pool_proj',
)
l_inc_out = nn.layers.concat([l_conv_inc, l_conv_inc2b, l_conv_inc2d, l_conv_inc2e])
return l_inc_out
l_inc_3a = inception(l_pool2, 'inception_3a', no_1x1=64, no_3x3r=96, no_3x3=128, no_5x5r=16, no_5x5=32, no_pool=32)
l_inc_3b = inception(l_inc_3a, 'inception_3b', no_1x1=128, no_3x3r=128, no_3x3=192, no_5x5r=32, no_5x5=96, no_pool=64)
l_pool3 = flip(dnn.MaxPool2DDNNLayer(
flip(l_inc_3b),
pool_size=(3, 3),#pool_size
stride=(2, 2),
pad=1,
name='pool3/3x3_s2'
))
l_inc_4a = inception(l_pool3, 'inception_4a', no_1x1=192, no_3x3r=96, no_3x3=208, no_5x5r=16, no_5x5=48, no_pool=64)
l_inc_4b = inception(l_inc_4a, 'inception_4b', no_1x1=160, no_3x3r=112, no_3x3=224, no_5x5r=24, no_5x5=64, no_pool=64)
l_inc_4c = inception(l_inc_4b, 'inception_4c', no_1x1=128, no_3x3r=128, no_3x3=256, no_5x5r=24, no_5x5=64, no_pool=64)
l_inc_4d = inception(l_inc_4c, 'inception_4d', no_1x1=112, no_3x3r=144, no_3x3=288, no_5x5r=32, no_5x5=64, no_pool=64)
l_inc_4e = inception(l_inc_4d, 'inception_4e', no_1x1=256, no_3x3r=160, no_3x3=320, no_5x5r=32, no_5x5=128, no_pool=128)
l_pool4 = flip(dnn.MaxPool2DDNNLayer(
flip(l_inc_4e),
pool_size=(3, 3),#pool_size
stride=(2, 2),
pad=1,
name='pool4/3x3_s2'
))
l_inc_5a = inception(l_pool4, 'inception_5a', no_1x1=256, no_3x3r=160, no_3x3=320, no_5x5r=32, no_5x5=128, no_pool=128)
l_inc_5b = inception(l_inc_5a, 'inception_5b', no_1x1=384, no_3x3r=192, no_3x3=384, no_5x5r=48, no_5x5=128, no_pool=128)
l_pool5 = flip(dnn.Pool2DDNNLayer(
flip(l_inc_5b),
pool_size=(7, 7),#pool_size
stride=(1, 1),
pad=0,
mode='average',
name='pool5/7x7_s1'
))
l_logit = dnn.Conv2DDNNLayer(
l_pool5,
num_filters=1000,
filter_size=(1, 1),
pad=0,
stride=(1, 1),
nonlinearity=lasagne.nonlinearities.linear,
name='prob',
)
l_logit_flat = lasagne.layers.FlattenLayer(l_logit)
l_dense = lasagne.layers.NonlinearityLayer(l_logit_flat, nonlinearity=lasagne.nonlinearities.softmax)
l_out = l_dense
filename = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'googlenet.pkl')
if os.path.exists(filename):
with open(filename, 'r') as f:
nn.layers.set_all_param_values(l_dense, pickle.load(f))
return utils.struct(
input=l_in,
logit=l_logit,
out=l_out
)