def build_model(input_shape, input_var, dense=True):
net = {}
net['input'] = InputLayer(input_shape, input_var=input_var)
net['input'].num_filters = input_shape[1]
net['conv1'] = ConvLayer(net['input'], num_filters=128, filter_size=3, nonlinearity=nonlinearities.leaky_rectify, pad='same')
net['conv2'] = ConvLayer(net['conv1'], num_filters=256, filter_size=3, nonlinearity=nonlinearities.leaky_rectify, pad='same')
net['pool1'] = ConvLayer(net['conv2'], num_filters=256, filter_size=3, stride=2, nonlinearity=nonlinearities.leaky_rectify, pad='same')
net['conv3'] = ConvLayer(net['pool1'], num_filters=512, filter_size=3, nonlinearity=nonlinearities.leaky_rectify, pad='same')
net['pool2'] = ConvLayer(net['conv3'], num_filters=512, filter_size=3, stride=2, nonlinearity=nonlinearities.leaky_rectify, pad='same')
if dense:
net['dense'] = dropout(DenseLayer(net['pool2'], num_units=1024, nonlinearity=nonlinearities.leaky_rectify), 0.5)
# Deconv
net['dense/inverse'] = inverse_dense_layer(net['dense'], net['dense'], net['pool2'].output_shape)
net['pool2/inverse'] = inverse_convolution_strided_layer(net['dense/inverse'], net['pool2'])
else:
net['pool2/inverse'] = inverse_convolution_strided_layer(net['pool2'], net['pool2'])
net['conv3/inverse'] = inverse_convolution_layer(net['pool2/inverse'], net['conv3'])
net['pool1/inverse'] = inverse_convolution_strided_layer(net['conv3/inverse'], net['pool1'])
net['conv2/inverse'] = inverse_convolution_layer(net['pool1/inverse'], net['conv2'])
net['conv1/inverse'] = inverse_convolution_layer(net['conv2/inverse'], net['conv1'])
net['conv0/inverse'] = ConvLayer(net['conv1/inverse'], num_filters=input_shape[1], filter_size=1, nonlinearity=nonlinearities.linear, pad='same')
net['prob'] = net['conv0/inverse']
for layer in get_all_layers(net['prob']):
print layer
print layer.output_shape
return net
def build_model_dense(input_shape, input_var):
net = {}
net['input'] = InputLayer(input_shape, input_var=input_var)
net['input'].num_filters = input_shape[1]
net['conv1'] = ConvLayer(net['input'], num_filters=256, filter_size=3, nonlinearity=nonlinearities.leaky_rectify, pad='same')
net['conv2'] = ConvLayer(net['conv1'], num_filters=256, filter_size=3, nonlinearity=nonlinearities.leaky_rectify, pad='same')
net['conv2/reshape'] = ReshapeLayer(net['conv2'], (-1, net['conv2'].output_shape[1] * net['conv2'].output_shape[2]))
net['dense'] = dropout(DenseLayer(net['conv2/reshape'], num_units=1024, nonlinearity=nonlinearities.leaky_rectify), 0.5)
net['dense/inverse'] = inverse_dense_layer(net['dense'], net['dense'], net['conv2'].output_shape)
net['conv2/inverse'] = inverse_convolution_layer(net['dense/inverse'], net['conv2'])
net['conv1/inverse'] = inverse_convolution_layer(net['conv2/inverse'], net['conv1'])
net['conv0/inverse'] = ConvLayer(net['conv1/inverse'], num_filters=input_shape[1], filter_size=1,nonlinearity=nonlinearities.linear, pad='same')
net['prob'] = net['conv0/inverse']
for layer in get_all_layers(net['prob']):
print layer
print layer.output_shape
return net
def build_model_small(input_shape, input_var):
net = {}
net['input'] = InputLayer(input_shape, input_var=input_var)
net['input'].num_filters = input_shape[1]
net['conv1'] = batch_norm(ConvLayer(net['input'], num_filters=256, filter_size=11, nonlinearity=nonlinearities.leaky_rectify, pad='same'))
net['pool1'] = dropout(PoolLayer(net['conv1'], 2, mode='max'), 0.5)
net['conv2'] = batch_norm(ConvLayer(net['pool1'], num_filters=256, filter_size=7, nonlinearity=nonlinearities.leaky_rectify, pad='same'))
net['pool2'] = dropout(PoolLayer(net['conv2'], 2, mode='max'), 0.5)
net['conv3'] = batch_norm(ConvLayer(net['pool2'], num_filters=396, filter_size=5, nonlinearity=nonlinearities.leaky_rectify, pad='same'))
net['pool3'] = dropout(PoolLayer(net['conv3'], 2, mode='max'), 0.5)
net['conv4'] = dropout(batch_norm(ConvLayer(net['pool3'], num_filters=512, filter_size=3, nonlinearity=nonlinearities.leaky_rectify, pad='same')), 0.5)
net['conv5'] = dropout(batch_norm(ConvLayer(net['conv4'], num_filters=1024, filter_size=1, nonlinearity=nonlinearities.leaky_rectify,pad='same')), 0.5)
net['dense1'] = dropout(batch_norm(DenseLayer(net['conv5'], num_units=1024, nonlinearity=nonlinearities.leaky_rectify)), 0.5)
net['dense2'] = DenseLayer(net['dense1'], num_units=11, nonlinearity=nonlinearities.softmax)
net['prob'] = net['dense2']
for layer in get_all_layers(net['prob']):
print layer
print layer.output_shape
return net
def build_model_small(input_shape, input_var):
net = {}
net['input'] = InputLayer(input_shape, input_var=input_var)
net['input'].num_filters = input_shape[1]
net['conv1'] = ConvLayer(net['input'], num_filters=256, filter_size=11, nonlinearity=nonlinearities.leaky_rectify, pad='same')
net['conv2'] = ConvLayer(net['conv1'], num_filters=256, filter_size=7, nonlinearity=nonlinearities.leaky_rectify, pad='same')
net['conv3'] = ConvLayer(net['conv2'], num_filters=396, filter_size=5, nonlinearity=nonlinearities.leaky_rectify, pad='same')
net['conv4'] = ConvLayer(net['conv3'], num_filters=512, filter_size=3, nonlinearity=nonlinearities.leaky_rectify, pad='same')
net['conv5'] = ConvLayer(net['conv4'], num_filters=1024, filter_size=1, nonlinearity=nonlinearities.leaky_rectify,pad='same')
net['conv5/inverse'] = inverse_convolution_layer(net['conv5'], net['conv5'])
net['conv4/inverse'] = inverse_convolution_layer(net['conv5/inverse'], net['conv4'])
net['conv3/inverse'] = inverse_convolution_layer(net['conv4/inverse'], net['conv3'])
net['conv2/inverse'] = inverse_convolution_layer(net['conv3/inverse'], net['conv2'])
net['conv1/inverse'] = inverse_convolution_layer(net['conv2/inverse'], net['conv1'])
net['conv0/inverse'] = ConvLayer(net['conv1/inverse'], num_filters=input_shape[1], filter_size=1,nonlinearity=nonlinearities.linear, pad='same')
net['prob'] = net['conv0/inverse']
for layer in get_all_layers(net['prob']):
print layer
print layer.output_shape
return net
def build_network_zeta(inputlist,
imgh=(50, 25, 25),
imgw=127,
convpooldictlist=None,
nhidden=None,
dropoutp=None,
noutputs=11,
depth=1
):
"""
here, `inputlist` should have img tensors for x, u, v, and for muon_data
here, `imgh` is a tuple of sizes for `(x, u, v)`. `imgw` is the same
for all three views.
also, the `convpooldictlist` here must be a dictionary of dictionaries,
with the set of convolution and pooling defined independently for 'x', 'u',
and 'v' - e.g., `convpooldictlist['x']` will be a dictionary similar to
the dictionaries used by network models like `beta`, etc.
"""
net = {}
# Input layer
input_var_x, input_var_u, input_var_v, input_var_muon = \
inputlist[0], inputlist[1], inputlist[2], inputlist[3]
net['input-x'] = InputLayer(shape=(None, depth, imgw, imgh[0]),
input_var=input_var_x)
net['input-u'] = InputLayer(shape=(None, depth, imgw, imgh[1]),
input_var=input_var_u)
net['input-v'] = InputLayer(shape=(None, depth, imgw, imgh[2]),
input_var=input_var_v)
net['input-muon-dat'] = InputLayer(shape=(None, 10),
input_var=input_var_muon)
if convpooldictlist is None:
raise Exception('Conv-pool dictionaries must be defined!')
if nhidden is None:
nhidden = 256
if dropoutp is None:
dropoutp = 0.5
net.update(
make_Nconvpool_1dense_branch('x', net['input-x'],
convpooldictlist['x'],
nhidden, dropoutp))
net.update(
make_Nconvpool_1dense_branch('u', net['input-u'],
convpooldictlist['u'],
nhidden, dropoutp))
net.update(
make_Nconvpool_1dense_branch('v', net['input-v'],
convpooldictlist['v'],
nhidden, dropoutp))
# put a softmax on the muon vars
net['softed-muon-dat'] = DenseLayer(
net['input-muon-dat'],
num_units=noutputs,
nonlinearity=lasagne.nonlinearities.softmax
)
logger.info("Softmax on muon dat with n_units = {}".format(noutputs))
# Concatenate the parallel inputs, include the muon data
net['concat'] = ConcatLayer((
net['dense-x'],
net['dense-u'],
net['dense-v'],
net['softed-muon-dat']
))
logger.info("Network: concat columns...")
# One more dense layer
net['dense-across'] = DenseLayer(
dropout(net['concat'], p=dropoutp),
num_units=(nhidden // 2),
nonlinearity=lasagne.nonlinearities.rectify)
logger.info("Dense {} with nhidden = {}, dropout = {}".format(
'dense-across', nhidden // 2, dropoutp))
# And, finally, the `noutputs`-unit output layer
net['output_prob'] = DenseLayer(
net['dense-across'],
num_units=noutputs,
nonlinearity=lasagne.nonlinearities.softmax
)
logger.info("Softmax output prob with n_units = {}".format(noutputs))
logger.info(
"n-parameters: %s" % lasagne.layers.count_params(net['output_prob'])
)
return net['output_prob']
def build_beta_single_view(inputlist, view='x', imgh=68, imgw=127,
convpooldictlist=None,
nhidden=None, dropoutp=None, noutputs=11,
depth=1
):
"""
This network is modeled after the 'triamese' (tri-columnar) beta model,
but is meant to operate on one view only.
This function has a differen signature than the rest of the functions
in this module, so it is really not meant to be used as a `build_cnn`
function in the runner scripts (although, in Python, that would work).
"""
net = {}
# Input layer
input_var = inputlist[0]
tshape = (None, depth, imgw, imgh)
input_name = 'input-' + view
net[input_name] = InputLayer(shape=tshape, input_var=input_var)
if convpooldictlist is None:
convpooldictlist = []
convpool1dict = {}
convpool1dict['nfilters'] = 32
convpool1dict['filter_size'] = (3, 3)
convpool1dict['pool_size'] = (2, 2)
convpooldictlist.append(convpool1dict)
convpool2dict = {}
convpool2dict['nfilters'] = 32
convpool2dict['filter_size'] = (3, 3)
convpool2dict['pool_size'] = (2, 2)
convpooldictlist.append(convpool2dict)
if nhidden is None:
nhidden = 256
if dropoutp is None:
dropoutp = 0.5
net.update(
make_Nconvpool_1dense_branch(view, net[input_name], convpooldictlist,
nhidden, dropoutp))
# One more dense layer
dense_name = 'dense-' + view
net['dense-across'] = DenseLayer(
dropout(net[dense_name], p=dropoutp),
num_units=(nhidden // 2),
nonlinearity=lasagne.nonlinearities.rectify)
logger.info("Dense {} with nhidden = {}, dropout = {}".format(
'dense-across', nhidden // 2, dropoutp))
# And, finally, the `noutputs`-unit output layer
net['output_prob'] = DenseLayer(
net['dense-across'],
num_units=noutputs,
nonlinearity=lasagne.nonlinearities.softmax
)
logger.info("Softmax output prob with n_units = {}".format(noutputs))
logger.info("n-parameters: {}".format(
lasagne.layers.count_params(net['output_prob']))
)
return net['output_prob']
def build_triamese_delta(inputlist, imgh=68, imgw=127, convpooldictlist=None,
nhidden=None, dropoutp=None, noutputs=67, depth=1):
"""
'triamese' (one branch for each view, feeding a fully-connected network),
model using two layers of convolutions and pooling.
This model is basically identical to the `beta` model, except we have
a softmax output of `noutputs` (def 67) for the full set of planecodes.
"""
net = {}
# Input layer
input_var_x, input_var_u, input_var_v = \
inputlist[0], inputlist[1], inputlist[2]
tshape = (None, depth, imgw, imgh)
net['input-x'] = InputLayer(shape=tshape, input_var=input_var_x)
net['input-u'] = InputLayer(shape=tshape, input_var=input_var_u)
net['input-v'] = InputLayer(shape=tshape, input_var=input_var_v)
if convpooldictlist is None:
convpooldictlist = []
convpool1dict = {}
convpool1dict['nfilters'] = 32
convpool1dict['filter_size'] = (3, 3)
convpool1dict['pool_size'] = (2, 2)
convpooldictlist.append(convpool1dict)
convpool2dict = {}
convpool2dict['nfilters'] = 32
convpool2dict['filter_size'] = (3, 3)
convpool2dict['pool_size'] = (2, 2)
convpooldictlist.append(convpool2dict)
if nhidden is None:
nhidden = 256
if dropoutp is None:
dropoutp = 0.5
net.update(
make_Nconvpool_1dense_branch('x', net['input-x'], convpooldictlist,
nhidden, dropoutp))
net.update(
make_Nconvpool_1dense_branch('u', net['input-u'], convpooldictlist,
nhidden, dropoutp))
net.update(
make_Nconvpool_1dense_branch('v', net['input-v'], convpooldictlist,
nhidden, dropoutp))
# Concatenate the two parallel inputs
net['concat'] = ConcatLayer((net['dense-x'],
net['dense-u'],
net['dense-v']))
logger.info("Network: concat columns...")
# One more dense layer
net['dense-across'] = DenseLayer(
dropout(net['concat'], p=dropoutp),
num_units=(nhidden // 2),
nonlinearity=lasagne.nonlinearities.rectify)
logger.info("Dense {} with nhidden = {}, dropout = {}".format(
'dense-across', nhidden // 2, dropoutp))
# And, finally, the `noutputs`-unit output layer
net['output_prob'] = DenseLayer(
net['dense-across'],
num_units=noutputs,
nonlinearity=lasagne.nonlinearities.softmax
)
logger.info("Softmax output prob with n_units = {}".format(noutputs))
logger.info("n-parameters: {}".format(
lasagne.layers.count_params(net['output_prob']))
)
return net['output_prob']
def build_triamese_gamma(inputlist, imgh=50, imgw=50, convpooldictlist=None,
nhidden=None, dropoutp=None, noutputs=11, depth=1):
"""
'triamese' (one branch for each view, feeding a fully-connected network),
model using two layers of convolutions - no pooling.
"""
net = {}
# Input layer
input_var_x, input_var_u, input_var_v = \
inputlist[0], inputlist[1], inputlist[2]
tshape = (None, depth, imgw, imgh)
net['input-x'] = InputLayer(shape=tshape, input_var=input_var_x)
net['input-u'] = InputLayer(shape=tshape, input_var=input_var_u)
net['input-v'] = InputLayer(shape=tshape, input_var=input_var_v)
if convpooldictlist is None:
convpooldictlist = []
convpool1dict = {}
convpool1dict['nfilters'] = 32
convpool1dict['filter_size'] = (3, 3)
convpooldictlist.append(convpool1dict)
convpool2dict = {}
convpool2dict['nfilters'] = 16
convpool2dict['filter_size'] = (3, 3)
convpooldictlist.append(convpool2dict)
if nhidden is None:
nhidden = 256
if dropoutp is None:
dropoutp = 0.5
def make_branch(view, input_layer, cpdictlist, nhidden=256, dropoutp=0.5):
"""
see: http://lasagne.readthedocs.org/en/latest/modules/layers.html
convolution only - no pooling
"""
net = {}
convname = ''
prev_layername = ''
for i, cpdict in enumerate(cpdictlist):
convname = 'conv-{}-{}'.format(view, i)
logger.info("Convpool {} params: {}".format(convname, cpdict))
# the first time through, use `input`, after use the last layer
# from the previous iteration - ah loose scoping rules...
if i == 0:
layer = input_layer
else:
layer = net[prev_layername]
net[convname] = Conv2DLayer(
layer, num_filters=cpdict['nfilters'],
filter_size=cpdict['filter_size'],
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.GlorotUniform())
prev_layername = convname
densename = 'dense-{}'.format(view)
net[densename] = DenseLayer(
dropout(net[convname], p=dropoutp),
num_units=nhidden,
nonlinearity=lasagne.nonlinearities.rectify)
logger.info("Dense {} with nhidden = {}, dropout = {}".format(
densename, nhidden, dropoutp))
return net
net.update(make_branch('x', net['input-x'], convpooldictlist,
nhidden, dropoutp))
net.update(make_branch('u', net['input-u'], convpooldictlist,
nhidden, dropoutp))
net.update(make_branch('v', net['input-v'], convpooldictlist,
nhidden, dropoutp))
# Concatenate the two parallel inputs
net['concat'] = ConcatLayer((net['dense-x'],
net['dense-u'],
net['dense-v']))
logger.info("Network: concat columns...")
# One more dense layer
net['dense-across'] = DenseLayer(
dropout(net['concat'], p=dropoutp),
num_units=(nhidden // 2),
nonlinearity=lasagne.nonlinearities.rectify)
logger.info("Dense {} with nhidden = {}, dropout = {}".format(
'dense-across', nhidden // 2, dropoutp))
# And, finally, the `noutputs`-unit output layer
net['output_prob'] = DenseLayer(
net['dense-across'],
num_units=noutputs,
nonlinearity=lasagne.nonlinearities.softmax
)
logger.info("Softmax output prob with n_units = {}".format(noutputs))
logger.info("n-parameters: {}".format(
lasagne.layers.count_params(net['output_prob']))
)
return net['output_prob']
def build_model():
#################
# Regular model #
#################
data_size = data_sizes["sliced:data:ax:noswitch"]
l0 = InputLayer(data_size)
l0r = batch_norm(reshape(l0, (-1, 1, ) + data_size[1:]))
# (batch, channel, axis, time, x, y)
# convolve over time
l1 = batch_norm(ConvolutionOverAxisLayer(l0r, num_filters=4, filter_size=(3,), axis=(3,), channel=1,
W=lasagne.init.Orthogonal(),
b=lasagne.init.Constant(0.0),
))
l1m = batch_norm(MaxPoolOverAxisLayer(l1, pool_size=(4,), axis=(3,)))
# convolve over x and y
l2a = batch_norm(ConvolutionOver2DAxisLayer(l1m, num_filters=8, filter_size=(3, 3),
axis=(4,5), channel=1,
W=lasagne.init.Orthogonal(),
b=lasagne.init.Constant(0.0),
))
l2b = batch_norm(ConvolutionOver2DAxisLayer(l2a, num_filters=8, filter_size=(3, 3),
axis=(4,5), channel=1,
W=lasagne.init.Orthogonal(),
b=lasagne.init.Constant(0.0),
))
l2m = batch_norm(MaxPoolOver2DAxisLayer(l2b, pool_size=(2, 2), axis=(4,5)))
# convolve over x, y, time
l3a = batch_norm(ConvolutionOver3DAxisLayer(l2m, num_filters=32, filter_size=(3, 3, 3),
axis=(3,4,5), channel=1,
W=lasagne.init.Orthogonal(),
b=lasagne.init.Constant(0.1),
))
l3b = batch_norm(ConvolutionOver2DAxisLayer(l3a, num_filters=32, filter_size=(3, 3),
axis=(4,5), channel=1,
W=lasagne.init.Orthogonal(),
b=lasagne.init.Constant(0.1),
))
l3m = batch_norm(MaxPoolOver2DAxisLayer(l3b, pool_size=(2, 2), axis=(4,5)))
# convolve over time
l4 = batch_norm(ConvolutionOverAxisLayer(l3m, num_filters=32, filter_size=(3,), axis=(3,), channel=1,
W=lasagne.init.Orthogonal(),
b=lasagne.init.Constant(0.1),
))
l4m = batch_norm(MaxPoolOverAxisLayer(l4, pool_size=(2,), axis=(2,)))
# maxpool over axis
l5 = batch_norm(MaxPoolOverAxisLayer(l3m, pool_size=(4,), axis=(2,)))
# convolve over x and y
l6a = batch_norm(ConvolutionOver2DAxisLayer(l5, num_filters=128, filter_size=(3, 3),
axis=(4,5), channel=1,
W=lasagne.init.Orthogonal(),
b=lasagne.init.Constant(0.1),
))
l6b = batch_norm(ConvolutionOver2DAxisLayer(l6a, num_filters=128, filter_size=(3, 3),
axis=(4,5), channel=1,
W=lasagne.init.Orthogonal(),
b=lasagne.init.Constant(0.1),
))
l6m = batch_norm(MaxPoolOver2DAxisLayer(l6b, pool_size=(2, 2), axis=(4,5)))
# convolve over time and x,y
l7 = ConvolutionOver3DAxisLayer(l6m, num_filters=128, filter_size=(3,3,3), axis=(3,4,5), channel=1,
W=lasagne.init.Orthogonal(),
b=lasagne.init.Constant(0.1),
)
l8 = lasagne.layers.DropoutLayer(l7, p=0.5)
l_systole = CumSumLayer(lasagne.layers.DenseLayer(l8,
num_units=600,
nonlinearity=lasagne.nonlinearities.softmax))
l_diastole = CumSumLayer(lasagne.layers.DenseLayer(l8,
num_units=600,
nonlinearity=lasagne.nonlinearities.softmax))
return {
"inputs":{
"sliced:data:ax:noswitch": l0
},
"outputs": {
"systole": l_systole,
"diastole": l_diastole,
},
"regularizable": {
}
}
def build_network(pooling='avg'):
net = {}
net['input'] = InputLayer((None, 3, 299, 299))
net['conv'] = bn_conv(net['input'],
num_filters=32,
filter_size=3,
stride=2)
net['conv_1'] = bn_conv(net['conv'], num_filters=32, filter_size=3)
net['conv_2'] = bn_conv(net['conv_1'],
num_filters=64,
filter_size=3,
pad=1)
net['pool'] = Pool2DLayer(net['conv_2'], pool_size=3, stride=2, mode='max')
net['conv_3'] = bn_conv(net['pool'], num_filters=80, filter_size=1)
net['conv_4'] = bn_conv(net['conv_3'], num_filters=192, filter_size=3)
net['pool_1'] = Pool2DLayer(net['conv_4'],
pool_size=3,
stride=2,
mode='max')
net['mixed/join'] = inceptionA(net['pool_1'],
nfilt=((64, ), (48, 64), (64, 96, 96),
(32, )))
net['mixed_1/join'] = inceptionA(net['mixed/join'],
nfilt=((64, ), (48, 64), (64, 96, 96),
(64, )))
net['mixed_2/join'] = inceptionA(net['mixed_1/join'],
nfilt=((64, ), (48, 64), (64, 96, 96),
(64, )))
net['mixed_3/join'] = inceptionB(net['mixed_2/join'],
nfilt=((384, ), (64, 96, 96)))
net['mixed_4/join'] = inceptionC(net['mixed_3/join'],
nfilt=((192, ), (128, 128, 192),
(128, 128, 128, 128,
192), (192, )))
net['mixed_5/join'] = inceptionC(net['mixed_4/join'],
nfilt=((192, ), (160, 160, 192),
(160, 160, 160, 160,
192), (192, )))
net['mixed_6/join'] = inceptionC(net['mixed_5/join'],
nfilt=((192, ), (160, 160, 192),
(160, 160, 160, 160,
192), (192, )))
net['mixed_7/join'] = inceptionC(net['mixed_6/join'],
nfilt=((192, ), (192, 192, 192),
(192, 192, 192, 192,
192), (192, )))
net['mixed_8/join'] = inceptionD(net['mixed_7/join'],
nfilt=((192, 320), (192, 192, 192, 192)))
net['mixed_9/join'] = inceptionE(net['mixed_8/join'],
nfilt=((320, ), (384, 384, 384),
(448, 384, 384, 384), (192, )),
pool_mode='average_exc_pad')
net['mixed_10/join'] = inceptionE(net['mixed_9/join'],
nfilt=((320, ), (384, 384, 384),
(448, 384, 384, 384), (192, )),
pool_mode='max')
if pooling == "max":
net['pool3'] = GlobalPoolLayer(net['mixed_10/join'],
pool_function=theano.tensor.max)
elif pooling in ("avg", "mean"):
net['pool3'] = GlobalPoolLayer(net['mixed_10/join'],
pool_function=theano.tensor.mean)
else:
raise "pooling not specified"
net['softmax'] = DenseLayer(net['pool3'],
num_units=1008,
nonlinearity=softmax)
return net
random_state = np.random.RandomState(1999)
n_epochs = 200
# theano land tensor4 for 4 dimensions
input_var = tensor.tensor4('X')
target_var = tensor.tensor4('y')
outchan = y_train.shape[0]
inchan = X_train.shape[0]
width = X_train.shape[1]
height = X_train.shape[2]
input_var.tag.test_value = X_mb
target_var.tag.test_value = y_mb
# setting up theano - use None to indicate that dimension may change
coarse_input = InputLayer((minibatch_size, inchan, width, height),
input_var=input_var)
# choose number of filters and filter size
coarse_conv1 = Conv2DLayer(coarse_input,
num_filters=32,
filter_size=(5, 5),
nonlinearity=rectify,
W=GlorotUniform(),
pad=(2, 2))
coarse_pool1 = MaxPool2DLayer(coarse_conv1, pool_size=(2, 2))
coarse_conv2 = Conv2DLayer(coarse_pool1,
num_filters=64,
filter_size=(3, 3),
nonlinearity=rectify,
W=GlorotUniform(),
def main():
# import VGG model
net = {}
net['input'] = InputLayer((None, 3, 224, 224))
net['conv1_1'] = ConvLayer(net['input'], 64, 3, pad=1, flip_filters=False)
net['conv1_2'] = ConvLayer(net['conv1_1'], 64, 3, pad=1, flip_filters=False)
net['pool1'] = PoolLayer(net['conv1_2'], 2)
net['conv2_1'] = ConvLayer(net['pool1'], 128, 3, pad=1, flip_filters=False)
net['conv2_2'] = ConvLayer(net['conv2_1'], 128, 3, pad=1, flip_filters=False)
net['pool2'] = PoolLayer(net['conv2_2'], 2)
net['conv3_1'] = ConvLayer(net['pool2'], 256, 3, pad=1, flip_filters=False)
net['conv3_2'] = ConvLayer(net['conv3_1'], 256, 3, pad=1, flip_filters=False)
net['conv3_3'] = ConvLayer(net['conv3_2'], 256, 3, pad=1, flip_filters=False)
net['pool3'] = PoolLayer(net['conv3_3'], 2)
net['conv4_1'] = ConvLayer(net['pool3'], 512, 3, pad=1, flip_filters=False)
net['conv4_2'] = ConvLayer(net['conv4_1'], 512, 3, pad=1, flip_filters=False)
net['conv4_3'] = ConvLayer(net['conv4_2'], 512, 3, pad=1, flip_filters=False)
net['pool4'] = PoolLayer(net['conv4_3'], 2)
net['conv5_1'] = ConvLayer(net['pool4'], 512, 3, pad=1, flip_filters=False)
net['conv5_2'] = ConvLayer(net['conv5_1'], 512, 3, pad=1, flip_filters=False)
net['conv5_3'] = ConvLayer(net['conv5_2'], 512, 3, pad=1, flip_filters=False)
net['pool5'] = PoolLayer(net['conv5_3'], 2)
net['fc6'] = DenseLayer(net['pool5'], num_units=4096)
net['fc7'] = DenseLayer(net['fc6'], num_units=4096)
net['fc8'] = DenseLayer(net['fc7'], num_units=1000, nonlinearity=None)
net['prob'] = NonlinearityLayer(net['fc8'], softmax)
# loading one image and compute its saliency map
lasagne.layers.set_all_param_values(net['prob'], weights)
url = 'http://farm5.static.flickr.com/4064/4334173592_145856d89b.jpg'
img_original, img = prepare_image(url)
saliency_fn = compile_saliency_function(net)
saliency, max_class = saliency_fn(img)
show_images(img_original, saliency, max_class, "default gradient")
relu = lasagne.nonlinearities.rectify
relu_layers = [layer for layer in lasagne.layers.get_all_layers(net['prob'])
if getattr(layer, 'nonlinearity', None) is relu]
# Guided Backpropgation
modded_relu = GuidedBackprop(relu) # important: only instantiate this once!
for layer in relu_layers:
layer.nonlinearity = modded_relu
saliency_fn = compile_saliency_function(net)
saliency, max_class = saliency_fn(img)
show_images(img_original, saliency, max_class, "guided backprop")
# using Zeiler Backpropagation
modded_relu = ZeilerBackprop(relu)
for layer in relu_layers:
layer.nonlinearity = modded_relu
saliency_fn = compile_saliency_function(net)
saliency, max_class = saliency_fn(img)
show_images(img_original, saliency, max_class, "deconvnet")
def create_model(ae,
diff_ae,
input_shape,
input_var,
mask_shape,
mask_var,
diff_shape,
diff_var,
lstm_size=250,
win=T.iscalar('theta)'),
output_classes=26,
fusiontype='concat',
w_init_fn=las.init.Orthogonal(),
use_peepholes=True):
bn_weights, bn_biases, bn_shapes, bn_nonlinearities = ae
diff_weights, diff_biases, diff_shapes, diff_nonlinearities = diff_ae
gate_parameters = Gate(W_in=w_init_fn,
W_hid=w_init_fn,
b=las.init.Constant(0.))
cell_parameters = Gate(
W_in=w_init_fn,
W_hid=w_init_fn,
# Setting W_cell to None denotes that no cell connection will be used.
W_cell=None,
b=las.init.Constant(0.),
# By convention, the cell nonlinearity is tanh in an LSTM.
nonlinearity=tanh)
l_raw = InputLayer(input_shape, input_var, 'raw_im')
l_mask = InputLayer(mask_shape, mask_var, 'mask')
l_diff = InputLayer(diff_shape, diff_var, 'diff_im')
symbolic_batchsize_raw = l_raw.input_var.shape[0]
symbolic_seqlen_raw = l_raw.input_var.shape[1]
symbolic_batchsize_diff = l_diff.input_var.shape[0]
symbolic_seqlen_diff = l_diff.input_var.shape[1]
l_reshape1_raw = ReshapeLayer(l_raw, (-1, input_shape[-1]),
name='reshape1_raw')
l_encoder_raw = create_pretrained_encoder(
l_reshape1_raw, bn_weights, bn_biases, bn_shapes, bn_nonlinearities,
['fc1_raw', 'fc2_raw', 'fc3_raw', 'bottleneck_raw'])
raw_len = las.layers.get_output_shape(l_encoder_raw)[-1]
l_reshape2_raw = ReshapeLayer(
l_encoder_raw, (symbolic_batchsize_raw, symbolic_seqlen_raw, raw_len),
name='reshape2_raw')
l_delta_raw = DeltaLayer(l_reshape2_raw, win, name='delta_raw')
# diff images
l_reshape1_diff = ReshapeLayer(l_diff, (-1, diff_shape[-1]),
name='reshape1_diff')
l_encoder_diff = create_pretrained_encoder(
l_reshape1_diff, diff_weights, diff_biases, diff_shapes,
diff_nonlinearities,
['fc1_diff', 'fc2_diff', 'fc3_diff', 'bottleneck_diff'])
diff_len = las.layers.get_output_shape(l_encoder_diff)[-1]
l_reshape2_diff = ReshapeLayer(
l_encoder_diff,
(symbolic_batchsize_diff, symbolic_seqlen_diff, diff_len),
name='reshape2_diff')
l_delta_diff = DeltaLayer(l_reshape2_diff, win, name='delta_diff')
l_lstm_raw = LSTMLayer(
l_delta_raw,
int(lstm_size),
peepholes=use_peepholes,
# We need to specify a separate input for masks
mask_input=l_mask,
# Here, we supply the gate parameters for each gate
ingate=gate_parameters,
forgetgate=gate_parameters,
cell=cell_parameters,
outgate=gate_parameters,
# We'll learn the initialization and use gradient clipping
learn_init=True,
grad_clipping=5.,
name='lstm_raw')
l_lstm_diff = LSTMLayer(
l_delta_diff,
lstm_size,
peepholes=use_peepholes,
# We need to specify a separate input for masks
mask_input=l_mask,
# Here, we supply the gate parameters for each gate
ingate=gate_parameters,
forgetgate=gate_parameters,
cell=cell_parameters,
outgate=gate_parameters,
# We'll learn the initialization and use gradient clipping
learn_init=True,
grad_clipping=5.,
name='lstm_diff')
# We'll combine the forward and backward layer output by summing.
# Merge layers take in lists of layers to merge as input.
if fusiontype == 'adasum':
l_fuse = AdaptiveElemwiseSumLayer([l_lstm_raw, l_lstm_diff],
name='adasum1')
elif fusiontype == 'sum':
l_fuse = ElemwiseSumLayer([l_lstm_raw, l_lstm_diff], name='sum1')
elif fusiontype == 'concat':
l_fuse = ConcatLayer([l_lstm_raw, l_lstm_diff], axis=-1, name='concat')
f_lstm_agg = create_lstm(l_fuse, l_mask, lstm_size, cell_parameters,
gate_parameters, 'lstm_agg')
# reshape to (num_examples * seq_len, lstm_size)
l_reshape3 = ReshapeLayer(f_lstm_agg, (-1, lstm_size))
# Now, we can apply feed-forward layers as usual.
# We want the network to predict a classification for the sequence,
# so we'll use a the number of classes.
l_softmax = DenseLayer(l_reshape3,
num_units=output_classes,
nonlinearity=las.nonlinearities.softmax,
name='softmax')
l_out = ReshapeLayer(l_softmax, (-1, symbolic_seqlen_raw, output_classes),
name='output')
return l_out, l_fuse
def test_gru_grad_clipping():
# test that you can set grad_clip variable
x = T.tensor3()
l_rec = GRULayer(InputLayer((2, 2, 3)), 5, grad_clipping=1)
output = lasagne.layers.get_output(l_rec, x)
def test_lstm_hid_init_layer_eval():
# Test `hid_init` as a `Layer` with some dummy input. Compare the output of
# a network with a `Layer` as input to `hid_init` to a network with a
# `np.array` as input to `hid_init`
n_units = 7
n_test_cases = 2
in_shp = (n_test_cases, 2, 3)
in_h_shp = (1, n_units)
in_cell_shp = (1, n_units)
# dummy inputs
X_test = np.ones(in_shp, dtype=theano.config.floatX)
Xh_test = np.ones(in_h_shp, dtype=theano.config.floatX)
Xc_test = np.ones(in_cell_shp, dtype=theano.config.floatX)
Xh_test_batch = np.tile(Xh_test, (n_test_cases, 1))
Xc_test_batch = np.tile(Xc_test, (n_test_cases, 1))
# network with `Layer` initializer for hid_init
l_inp = InputLayer(in_shp)
l_inp_h = InputLayer(in_h_shp)
l_inp_cell = InputLayer(in_cell_shp)
l_rec_inp_layer = LSTMLayer(l_inp,
n_units,
hid_init=l_inp_h,
cell_init=l_inp_cell,
nonlinearity=None)
# network with `np.array` initializer for hid_init
l_rec_nparray = LSTMLayer(l_inp,
n_units,
hid_init=Xh_test,
cell_init=Xc_test,
nonlinearity=None)
# copy network parameters from l_rec_inp_layer to l_rec_nparray
l_il_param = dict([(p.name, p) for p in l_rec_inp_layer.get_params()])
l_rn_param = dict([(p.name, p) for p in l_rec_nparray.get_params()])
for k, v in l_rn_param.items():
if k in l_il_param:
v.set_value(l_il_param[k].get_value())
# build the theano functions
X = T.tensor3()
Xh = T.matrix()
Xc = T.matrix()
output_inp_layer = lasagne.layers.get_output(l_rec_inp_layer, {
l_inp: X,
l_inp_h: Xh,
l_inp_cell: Xc
})
output_nparray = lasagne.layers.get_output(l_rec_nparray, {l_inp: X})
# test both nets with dummy input
output_val_inp_layer = output_inp_layer.eval({
X: X_test,
Xh: Xh_test_batch,
Xc: Xc_test_batch
})
output_val_nparray = output_nparray.eval({X: X_test})
# check output given `Layer` is the same as with `np.array`
assert np.allclose(output_val_inp_layer, output_val_nparray)
def build_network(batch_size, z_shape, img_height, img_width,
conv_nonlinearity, dense_nonlinearity):
# Draws heavy inspiration from ResNet
num_filters = 32
filter_shape = (5, 5)
l_in = InputLayer((batch_size, 1, z_shape))
dense_nonlinearity = lasagne.nonlinearities.rectify
conv_nonlinearity = lasagne.nonlinearities.rectify
config = {
'conv_1_repeats': 0,
'conv_2_repeats': 0,
'conv_3_repeats': 0,
'conv_4_repeats': 0
}
#####################
### Decoding half ###
#####################
h_test = 2
w_test = 5
dec_2_size = h_test * w_test * num_filters * 8
l_hid_dec_2 = batch_norm(
DenseLayer(l_in, dec_2_size, nonlinearity=dense_nonlinearity))
l_dec_reshape = ReshapeLayer(
l_hid_dec_2,
[batch_size, dec_2_size / h_test / w_test, h_test, w_test])
conv_1 = batch_norm(
TransposedConv2DLayer(l_dec_reshape,
num_filters * 8,
filter_shape,
nonlinearity=conv_nonlinearity,
untie_biases=True))
for _ in range(config['conv_1_repeats']):
conv_1 = batch_norm(
TransposedConv2DLayer(conv_1,
num_filters * 8,
filter_shape,
nonlinearity=conv_nonlinearity,
untie_biases=True,
crop='same'))
conv_2 = batch_norm(
TransposedConv2DLayer(conv_1,
num_filters * 4,
filter_shape,
nonlinearity=conv_nonlinearity,
untie_biases=True,
stride=(2, 2),
crop='same'))
for _ in range(config['conv_2_repeats']):
conv_2 = batch_norm(
TransposedConv2DLayer(conv_2,
num_filters * 4,
filter_shape,
nonlinearity=conv_nonlinearity,
untie_biases=True,
crop='same'))
conv_3 = batch_norm(
TransposedConv2DLayer(conv_2,
num_filters * 2,
filter_shape,
nonlinearity=conv_nonlinearity,
untie_biases=True,
stride=(2, 2),
crop='same'))
for _ in range(config['conv_3_repeats']):
conv_3 = batch_norm(
TransposedConv2DLayer(conv_3,
num_filters * 2,
filter_shape,
nonlinearity=conv_nonlinearity,
untie_biases=True,
crop='same'))
conv_4 = batch_norm(
TransposedConv2DLayer(conv_3,
num_filters,
filter_shape,
nonlinearity=conv_nonlinearity,
untie_biases=True,
stride=(2, 2),
crop='same'))
for _ in range(config['conv_4_repeats']):
conv_4 = batch_norm(
TransposedConv2DLayer(conv_4,
num_filters,
filter_shape,
nonlinearity=conv_nonlinearity,
untie_biases=True,
crop='same'))
l_out = batch_norm(
TransposedConv2DLayer(conv_4,
1,
filter_shape,
nonlinearity=lasagne.nonlinearities.sigmoid,
untie_biases=True,
crop='same'))
return l_in, l_out
def build_triamese_alpha(inputlist, imgh=50, imgw=50,
convpool1dict=None, convpool2dict=None,
convpooldictlist=None, nhidden=None,
dropoutp=None, noutputs=11,
depth=1
):
"""
'triamese' (one branch for each view, feeding a fully-connected network),
model using two layers of convolutions and pooling.
"""
# Input layer
input_var_x, input_var_u, input_var_v = \
inputlist[0], inputlist[1], inputlist[2]
tshape = (None, depth, imgw, imgh)
l_in1_x = InputLayer(shape=tshape, input_var=input_var_x)
l_in1_u = InputLayer(shape=tshape, input_var=input_var_u)
l_in1_v = InputLayer(shape=tshape, input_var=input_var_v)
if convpool1dict is None:
convpool1dict = {}
convpool1dict['nfilters'] = 32
convpool1dict['filter_size'] = (3, 3)
convpool1dict['pool_size'] = (2, 2)
logger.info("Convpool1 params: {}".format(convpool1dict))
if convpool2dict is None:
convpool2dict = {}
convpool2dict['nfilters'] = 32
convpool2dict['filter_size'] = (3, 3)
convpool2dict['pool_size'] = (2, 2)
logger.info("Convpool2 params: {}".format(convpool2dict))
logger.info("Network: one dense layer per column...")
def make_branch(input_layer,
num_filters1, filter_size1, pool_size1,
num_filters2, filter_size2, pool_size2):
"""
see: http://lasagne.readthedocs.org/en/latest/modules/layers.html
"""
convlayer1 = Conv2DLayer(input_layer, num_filters=num_filters1,
filter_size=filter_size1,
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.GlorotUniform())
maxpoollayer1 = MaxPool2DLayer(convlayer1, pool_size=pool_size1)
convlayer2 = Conv2DLayer(maxpoollayer1, num_filters=num_filters2,
filter_size=filter_size1,
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.GlorotUniform())
maxpoollayer2 = MaxPool2DLayer(convlayer2, pool_size=pool_size2)
dense1 = DenseLayer(
dropout(maxpoollayer2, p=.5),
num_units=256,
nonlinearity=lasagne.nonlinearities.rectify)
return dense1
l_branch_x = make_branch(l_in1_x,
convpool1dict['nfilters'],
convpool1dict['filter_size'],
convpool1dict['pool_size'],
convpool2dict['nfilters'],
convpool2dict['filter_size'],
convpool2dict['pool_size'])
l_branch_u = make_branch(l_in1_u,
convpool1dict['nfilters'],
convpool1dict['filter_size'],
convpool1dict['pool_size'],
convpool2dict['nfilters'],
convpool2dict['filter_size'],
convpool2dict['pool_size'])
l_branch_v = make_branch(l_in1_v,
convpool1dict['nfilters'],
convpool1dict['filter_size'],
convpool1dict['pool_size'],
convpool2dict['nfilters'],
convpool2dict['filter_size'],
convpool2dict['pool_size'])
# Concatenate the parallel inputs
l_concat = ConcatLayer((l_branch_x, l_branch_u, l_branch_v))
logger.info("Network: Concat all three columns...")
# And, finally, the noutputs-unit output layer
outp = DenseLayer(
l_concat,
num_units=noutputs,
nonlinearity=lasagne.nonlinearities.softmax
)
logger.info("Network: Softmax classification layer.")
logger.info("n-parameters: {}".format(lasagne.layers.count_params(outp)))
return outp
def build_resnet():
net = {}
net['input'] = InputLayer((None, 3, 224, 224))
sub_net, parent_layer_name = build_simple_block(
net['input'], ['conv1', 'bn_conv1', 'conv1_relu'],
64,
7,
2,
3,
use_bias=True)
net.update(sub_net)
net['pool1'] = PoolLayer(net[parent_layer_name],
pool_size=3,
stride=2,
pad=0,
mode='max',
ignore_border=False)
block_size = list('abc')
parent_layer_name = 'pool1'
for c in block_size:
if c == 'a':
sub_net, parent_layer_name = build_residual_block(
net[parent_layer_name], 1, 1, True, 4, ix='2%s' % c)
else:
sub_net, parent_layer_name = build_residual_block(
net[parent_layer_name], 1.0 / 4, 1, False, 4, ix='2%s' % c)
net.update(sub_net)
block_size = list('abcd')
for c in block_size:
if c == 'a':
sub_net, parent_layer_name = build_residual_block(
net[parent_layer_name],
1.0 / 2,
1.0 / 2,
True,
4,
ix='3%s' % c)
else:
sub_net, parent_layer_name = build_residual_block(
net[parent_layer_name], 1.0 / 4, 1, False, 4, ix='3%s' % c)
net.update(sub_net)
block_size = list('abcdef')
for c in block_size:
if c == 'a':
sub_net, parent_layer_name = build_residual_block(
net[parent_layer_name],
1.0 / 2,
1.0 / 2,
True,
4,
ix='4%s' % c)
else:
sub_net, parent_layer_name = build_residual_block(
net[parent_layer_name], 1.0 / 4, 1, False, 4, ix='4%s' % c)
net.update(sub_net)
block_size = list('abc')
for c in block_size:
if c == 'a':
sub_net, parent_layer_name = build_residual_block(
net[parent_layer_name],
1.0 / 2,
1.0 / 2,
True,
4,
ix='5%s' % c)
else:
sub_net, parent_layer_name = build_residual_block(
net[parent_layer_name], 1.0 / 4, 1, False, 4, ix='5%s' % c)
net.update(sub_net)
net['pool5'] = PoolLayer(net[parent_layer_name],
pool_size=7,
stride=1,
pad=0,
mode='average_exc_pad',
ignore_border=False)
net['fc1000'] = DenseLayer(net['pool5'], num_units=1000, nonlinearity=None)
net['prob'] = NonlinearityLayer(net['fc1000'], nonlinearity=softmax)
return net