def create_cnn_general(embedded_x, mycnf, max_len, embedding_size, inp_conv=False):
fv_len = 0
filter_sizes = mycnf['cnn_config']['filter_sizes']
num_filters = mycnf['cnn_config']['num_filters']
for i, fw in enumerate(filter_sizes):
conv = ConvolutionalActivation(
activation=Rectifier().apply,
filter_size=(fw, embedding_size),
num_filters=num_filters,
num_channels=1,
image_size=(max_len, embedding_size),
name="conv"+str(fw)+embedded_x.name)
pooling = MaxPooling((max_len-fw+1, 1), name="pool"+str(fw)+embedded_x.name)
initialize([conv])
if inp_conv:
convinp = embedded_x
else:
convinp = embedded_x.flatten().reshape((embedded_x.shape[0], 1, max_len, embedding_size))
onepool = pooling.apply(conv.apply(convinp)).flatten(2)
if i == 0:
outpools = onepool
else:
outpools = T.concatenate([outpools, onepool], axis=1)
fv_len += conv.num_filters
return outpools, fv_len
def test_convolutional_sequence():
x = tensor.tensor4('x')
num_channels = 4
pooling_size = 3
batch_size = 5
activation = Rectifier().apply
conv = ConvolutionalActivation(activation, (3, 3),
5,
weights_init=Constant(1.),
biases_init=Constant(5.))
pooling = MaxPooling(pooling_size=(pooling_size, pooling_size))
conv2 = ConvolutionalActivation(activation, (2, 2),
4,
weights_init=Constant(1.))
seq = ConvolutionalSequence([conv, pooling, conv2],
num_channels,
image_size=(17, 13))
seq.push_allocation_config()
assert conv.num_channels == 4
assert conv2.num_channels == 5
conv2.convolution.use_bias = False
y = seq.apply(x)
seq.initialize()
func = function([x], y)
x_val = numpy.ones((batch_size, 4, 17, 13), dtype=theano.config.floatX)
y_val = (numpy.ones((batch_size, 4, 4, 2)) * (9 * 4 + 5) * 4 * 5)
assert_allclose(func(x_val), y_val)
def test_convolutional_activation_use_bias():
act = ConvolutionalActivation(Rectifier().apply, (3, 3),
5,
4,
image_size=(9, 9),
use_bias=False)
act.allocate()
assert not act.convolution.use_bias
assert len(ComputationGraph([act.apply(tensor.tensor4())]).parameters) == 1
def __init__(self,
n_layers,
n_hidden,
spatial_width,
n_colors,
filter_size=3):
"""
A brick implementing a multi-layer convolutional network.
TODO make this multi-scale multi-layer convolution
"""
super(MultiLayerConvolution, self).__init__()
self.filter_size = filter_size
self.children = []
num_channels = n_colors
for ii in xrange(n_layers):
conv_layer = ConvolutionalActivation(
activation=conv_nonlinearity.apply,
filter_size=(filter_size, filter_size),
num_filters=n_hidden,
num_channels=num_channels,
image_size=(spatial_width, spatial_width),
# assume images are spatially smooth -- in which case output magnitude scales with
# # filter pixels rather than square root of # filter pixels, so initialize
# accordingly.
weights_init=IsotropicGaussian(std=np.sqrt(1. / (n_hidden)) /
filter_size**2),
biases_init=Constant(0),
border_mode='full',
name="conv%d" % ii)
self.children.append(conv_layer)
num_channels = n_hidden
def test_border_mode_not_pushed():
layers = [
Convolutional(border_mode='full'),
ConvolutionalActivation(Rectifier().apply),
ConvolutionalActivation(Rectifier().apply, border_mode='valid'),
ConvolutionalActivation(Rectifier().apply, border_mode='full')
]
stack = ConvolutionalSequence(layers)
stack.push_allocation_config()
assert stack.children[0].border_mode == 'full'
assert stack.children[1].border_mode == 'valid'
assert stack.children[2].border_mode == 'valid'
assert stack.children[3].border_mode == 'full'
stack2 = ConvolutionalSequence(layers, border_mode='full')
stack2.push_allocation_config()
assert stack2.children[0].border_mode == 'full'
assert stack2.children[1].border_mode == 'full'
assert stack2.children[2].border_mode == 'full'
assert stack2.children[3].border_mode == 'full'
def __init__(self,
conv_activations,
num_channels,
image_shape,
filter_sizes,
feature_maps,
pooling_sizes,
top_mlp_activations,
top_mlp_dims,
conv_step=None,
border_mode='valid',
**kwargs):
if conv_step is None:
self.conv_step = (1, 1)
else:
self.conv_step = conv_step
self.num_channels = num_channels
self.image_shape = image_shape
self.top_mlp_activations = top_mlp_activations
self.top_mlp_dims = top_mlp_dims
self.border_mode = border_mode
conv_parameters = zip(conv_activations, filter_sizes, feature_maps)
# Construct convolutional layers with corresponding parameters
self.layers = list(
interleave([(ConvolutionalActivation(filter_size=filter_size,
num_filters=num_filter,
activation=activation.apply,
step=self.conv_step,
border_mode=self.border_mode,
name='conv_{}'.format(i))
for i, (activation, filter_size,
num_filter) in enumerate(conv_parameters)),
(MaxPooling(size, name='pool_{}'.format(i))
for i, size in enumerate(pooling_sizes))]))
self.conv_sequence = ConvolutionalSequence(self.layers,
num_channels,
image_size=image_shape)
# Construct a top MLP
self.top_mlp = MLP(top_mlp_activations, top_mlp_dims)
# We need to flatten the output of the last convolutional layer.
# This brick accepts a tensor of dimension (batch_size, ...) and
# returns a matrix (batch_size, features)
self.flattener = Flattener()
application_methods = [
self.conv_sequence.apply, self.flattener.apply, self.top_mlp.apply
]
super(LeNet, self).__init__(application_methods, **kwargs)
def test_convolutional_sequence_use_bias():
cnn = ConvolutionalSequence([
ConvolutionalActivation(
activation=Rectifier().apply, filter_size=(1, 1), num_filters=1)
for _ in range(3)
],
num_channels=1,
image_size=(1, 1),
use_bias=False)
cnn.allocate()
x = tensor.tensor4()
y = cnn.apply(x)
params = ComputationGraph(y).parameters
assert len(params) == 3 and all(param.name == 'W' for param in params)
def apply_cnn(self, l_emb1, l_size1, l_emb2, l_size2, r_emb1, r_size1,
r_emb2, r_size2, embedding_size, mycnf):
assert l_size1 == r_size1
assert l_size2 == r_size2
assert l_size1 == l_size1
max_len = l_size1
fv_len = 0
filter_sizes = mycnf['cnn_config']['filter_sizes']
num_filters = mycnf['cnn_config']['num_filters']
for i, fw in enumerate(filter_sizes):
conv_left = ConvolutionalActivation(
activation=Rectifier().apply,
filter_size=(fw, embedding_size),
num_filters=num_filters,
num_channels=1,
image_size=(max_len, embedding_size),
name="conv" + str(fw) + l_emb1.name,
seed=self.curSeed)
conv_right = ConvolutionalActivation(
activation=Rectifier().apply,
filter_size=(fw, embedding_size),
num_filters=num_filters,
num_channels=1,
image_size=(max_len, embedding_size),
name="conv" + str(fw) + r_emb1.name,
seed=self.curSeed)
pooling = MaxPooling((max_len - fw + 1, 1), name="pool" + str(fw))
initialize([conv_left, conv_right])
l_convinp1 = l_emb1.flatten().reshape(
(l_emb1.shape[0], 1, max_len, embedding_size))
l_convinp2 = l_emb2.flatten().reshape(
(l_emb2.shape[0], 1, max_len, embedding_size))
l_pool1 = pooling.apply(conv_left.apply(l_convinp1)).flatten(2)
l_pool2 = pooling.apply(conv_left.apply(l_convinp2)).flatten(2)
r_convinp1 = r_emb1.flatten().reshape(
(r_emb1.shape[0], 1, max_len, embedding_size))
r_convinp2 = r_emb2.flatten().reshape(
(r_emb2.shape[0], 1, max_len, embedding_size))
r_pool1 = pooling.apply(conv_right.apply(r_convinp1)).flatten(2)
r_pool2 = pooling.apply(conv_right.apply(r_convinp2)).flatten(2)
onepools1 = T.concatenate([l_pool1, r_pool1], axis=1)
onepools2 = T.concatenate([l_pool2, r_pool2], axis=1)
fv_len += conv_left.num_filters * 2
if i == 0:
outpools1 = onepools1
outpools2 = onepools2
else:
outpools1 = T.concatenate([outpools1, onepools1], axis=1)
outpools2 = T.concatenate([outpools2, onepools2], axis=1)
return outpools1, outpools2, fv_len
def __init__(self,
num_channels,
num_filters,
spatial_width,
num_scales,
filter_size,
downsample_method='meanout',
name=""):
"""
A brick implementing a single layer in a multi-scale convolutional network.
"""
super(MultiScaleConvolution, self).__init__()
self.num_scales = num_scales
self.filter_size = filter_size
self.num_filters = num_filters
self.spatial_width = spatial_width
self.downsample_method = downsample_method
self.children = []
print "adding MultiScaleConvolution layer"
# for scale in range(self.num_scales-1, -1, -1):
for scale in range(self.num_scales):
print "scale %d" % scale
conv_layer = ConvolutionalActivation(
activation=conv_nonlinearity.apply,
filter_size=(filter_size, filter_size),
num_filters=num_filters,
num_channels=num_channels,
image_size=(spatial_width / 2**scale,
spatial_width / 2**scale),
# assume images are spatially smooth -- in which case output magnitude scales with
# # filter pixels rather than square root of # filter pixels, so initialize
# accordingly.
weights_init=IsotropicGaussian(
std=np.sqrt(1. / (num_filters)) / filter_size**2),
biases_init=Constant(0),
border_mode='full',
name=name + "scale%d" % scale)
self.children.append(conv_layer)
def test_convolutional_activation_use_bias():
act = ConvolutionalActivation(Rectifier().apply, (3, 3), 5, 4, image_size=(9, 9), use_bias=False)
act.allocate()
assert not act.convolution.use_bias
assert len(ComputationGraph([act.apply(tensor.tensor4())]).parameters) == 1
mlp_hiddens = [1000]
output_size = 2
activation = [Rectifier().apply for _ in num_filter]
mlp_activation = [Rectifier().apply for _ in mlp_hiddens] + [Softmax().apply]
#Create the symbolics variable
x = tensor.tensor4('image_features')
y = tensor.lmatrix('targets')
#Get the parameters
conv_parameters = zip(activation, filter_size, num_filter)
#Create the convolutions layers
conv_layers = list(
interleave([(ConvolutionalActivation(filter_size=filter_size,
num_filters=num_filter,
activation=activation,
name='conv_{}'.format(i))
for i, (activation, filter_size,
num_filter) in enumerate(conv_parameters)),
(MaxPooling(size, name='pool_{}'.format(i))
for i, size in enumerate(pooling_sizes))]))
#Create the sequence
conv_sequence = ConvolutionalSequence(conv_layers,
num_channels,
image_size=image_shape,
weights_init=Uniform(width=0.2),
biases_init=Constant(0.))
#Initialize the convnet
conv_sequence.initialize()
#Add the MLP