def build_fft_scale(x, y, size):
W = []
pnet = ll.InputLayer((None, 3, 101, 101), input_var=None)
pnet = ll.Conv2DLayer(pnet, 64, (3, 3), pad='same', nonlinearity=None)
pnet = ll.NonlinearityLayer(ll.BatchNormLayer(pnet))
pnet = ll.Pool2DLayer(pnet, (3, 3), (2, 2))
pnet = ll.Conv2DLayer(pnet, 64, (3, 3), pad='same', nonlinearity=None)
pnet = ll.NonlinearityLayer(
ll.BatchNormLayer(pnet),
nonlinearity=l.nonlinearities.LeakyRectify(0.1))
pnet = ll.Conv2DLayer(pnet, 32, (3, 3), pad='same', nonlinearity=None)
pnet = ll.BatchNormLayer(pnet)
x_p, y_p = ll.get_output(pnet, x), ll.get_output(pnet, y)
z_p = Customfftlayer(x_p, y_p)
net = ll.InputLayer((None, 64, 50, 50), input_var=z_p)
net = ll.BatchNormLayer(net)
net = ll.NonlinearityLayer(
ll.BatchNormLayer(
ll.Conv2DLayer(net, 64, (5, 5), pad='same', nonlinearity=None)))
net = ll.Pool2DLayer(net, (2, 2), mode='average_inc_pad')
net = ll.NonlinearityLayer(
ll.BatchNormLayer(
ll.Conv2DLayer(net, 64, (5, 5), pad='same', nonlinearity=None)))
net = ll.BatchNormLayer(ll.Conv2DLayer(net, 10, (1, 1), nonlinearity=None))
# return scale different: x_new/x_lod-1
p_scale = ll.get_output(net)
#p_scale = theano.gradient.disconnected_grad(p_scale)
net_scale = ll.InputLayer((None, 10, 25, 25), p_scale)
net_scale = ll.DenseLayer(net_scale,
100,
b=None,
nonlinearity=l.nonlinearities.tanh)
W.append(net_scale.get_params(regularizable=True)[0])
net_scale = ll.DenseLayer(net_scale, 2, b=None, nonlinearity=None)
# return heatmap with 2 times upsample of size
net_heat = ll.DenseLayer(net,
500,
b=None,
nonlinearity=l.nonlinearities.tanh)
W.append(net_heat.get_params(regularizable=True)[0])
net_heat = ll.DenseLayer(net, size**2, b=None, nonlinearity=None)
W.append(net_heat.get_params(regularizable=True)[0])
net_heat = ll.BatchNormLayer(net_heat)
net_heat = ll.ReshapeLayer(net_heat, ([0], 1, size, size))
net_heat = ll.Deconv2DLayer(net_heat,
64, (5, 5), (2, 2),
b=None,
crop='same',
nonlinearity=None)
net_heat = ll.BatchNormLayer(net_heat)
net_heat = ll.Conv2DLayer(net_heat,
1, (3, 3),
b=None,
pad='same',
nonlinearity=None)
W.append(net_heat.get_params(regularizable=True)[0])
return pnet, net_scale, net_heat, W
def _build(self):
layer = layers.InputLayer(shape=(None, 3, 224, 224), input_var=self.X)
layer = Conv2DDNNLayer(layer, num_filters=64, filter_size=(3, 3), pad=1, flip_filters=False)
layer = Conv2DDNNLayer(layer, num_filters=64, filter_size=(3, 3), pad=1, flip_filters=False)
layer = layers.Pool2DLayer(layer, pool_size=(2, 2), mode='max')
layer = Conv2DDNNLayer(layer, num_filters=128, filter_size=(3, 3), pad=1, flip_filters=False)
layer = Conv2DDNNLayer(layer, num_filters=128, filter_size=(3, 3), pad=1, flip_filters=False)
layer = layers.Pool2DLayer(layer, pool_size=(2, 2), mode='max')
layer = Conv2DDNNLayer(layer, num_filters=256, filter_size=(3, 3), pad=1, flip_filters=False)
layer = Conv2DDNNLayer(layer, num_filters=256, filter_size=(3, 3), pad=1, flip_filters=False)
layer = Conv2DDNNLayer(layer, num_filters=256, filter_size=(3, 3), pad=1, flip_filters=False)
layer = layers.Pool2DLayer(layer, pool_size=(2, 2), mode='max')
layer = Conv2DDNNLayer(layer, num_filters=512, filter_size=(3, 3), pad=1, flip_filters=False)
layer = Conv2DDNNLayer(layer, num_filters=512, filter_size=(3, 3), pad=1, flip_filters=False)
layer = Conv2DDNNLayer(layer, num_filters=512, filter_size=(3, 3), pad=1, flip_filters=False)
layer = layers.Pool2DLayer(layer, pool_size=(2, 2), mode='max')
layer = Conv2DDNNLayer(layer, num_filters=512, filter_size=(3, 3), pad=1, flip_filters=False)
layer = Conv2DDNNLayer(layer, num_filters=512, filter_size=(3, 3), pad=1, flip_filters=False)
layer = Conv2DDNNLayer(layer, num_filters=512, filter_size=(3, 3), pad=1, flip_filters=False)
layer = layers.Pool2DLayer(layer, pool_size=(2, 2), mode='max')
layer = layers.DenseLayer(layer, num_units=4096)
layer = layers.DropoutLayer(layer, p=0.5)
layer = layers.DenseLayer(layer, num_units=4096)
layer = layers.DropoutLayer(layer, p=0.5)
layer = layers.DenseLayer(layer, num_units=1000)
layer = layers.NonlinearityLayer(layer, nonlinearity=nonlinearities.softmax)
return layer
def build_siamese(layer):
""""""
smx = nonlinearities.softmax
lnr = nonlinearities.linear
layers = L.get_all_layers(layer)
nl = filter(
lambda l: hasattr(l, 'nonlinearity') and (
(l.nonlinearity != smx) and (l.nonlinearity != lnr)),
layers)[0].nonlinearity
if len(layers[0].output_shape) == 3:
Xl = T.tensor3('left')
Xr = T.tensor3('right')
elif len(layers[0].output_shape) == 4:
Xl = T.tensor4('left')
Xr = T.tensor4('right')
Ol = L.get_output(layer, inputs=Xl)
# Ol_vl = L.get_output(layer, inputs=Xl, deterministic=True)
Or = L.get_output(layer, inputs=Xr)
O = T.concatenate([Ol, Or], axis=-1)
layer = L.InputLayer((None, layer.output_shape[-1] * 2), input_var=O)
layer = L.DenseLayer(layer, 128, nonlinearity=None, name='hc1')
layer = L.BatchNormLayer(layer)
layer = L.NonlinearityLayer(layer, nonlinearity=nl)
layer = L.DenseLayer(layer, 2, nonlinearity=smx)
return layer, (Xl, Xr)
def ptb_lstm(input_var, vocabulary_size, hidden_size, seq_len, num_layers,
dropout, batch_size):
l_input = L.InputLayer(shape=(batch_size, seq_len), input_var=input_var)
l_embed = L.EmbeddingLayer(l_input,
vocabulary_size,
hidden_size,
W=init.Uniform(1.0))
l_lstms = []
for i in range(num_layers):
l_lstm = L.LSTMLayer(l_embed if i == 0 else l_lstms[-1],
hidden_size,
ingate=L.Gate(W_in=init.GlorotUniform(),
W_hid=init.Orthogonal()),
forgetgate=L.Gate(W_in=init.GlorotUniform(),
W_hid=init.Orthogonal(),
b=init.Constant(1.0)),
cell=L.Gate(
W_in=init.GlorotUniform(),
W_hid=init.Orthogonal(),
W_cell=None,
nonlinearity=lasagne.nonlinearities.tanh),
outgate=L.Gate(W_in=init.GlorotUniform(),
W_hid=init.Orthogonal()))
l_lstms.append(l_lstm)
l_drop = L.DropoutLayer(l_lstms[-1], dropout)
l_out = L.DenseLayer(l_drop, num_units=vocabulary_size, num_leading_axes=2)
l_out = L.ReshapeLayer(
l_out,
(l_out.output_shape[0] * l_out.output_shape[1], l_out.output_shape[2]))
l_out = L.NonlinearityLayer(l_out,
nonlinearity=lasagne.nonlinearities.softmax)
return l_out
def build_TOY(x, y):
z_p = T.concatenate((x, y), axis=1)
net = ll.InputLayer((None, 2, 100, 100), input_var=z_p)
net = ll.BatchNormLayer(net)
net = ll.NonlinearityLayer(
ll.BatchNormLayer(
ll.Conv2DLayer(net, 64, (5, 5), pad='same', nonlinearity=None)))
net = ll.Pool2DLayer(net, (2, 2), mode='average_inc_pad')
net = ll.NonlinearityLayer(
ll.BatchNormLayer(
ll.Conv2DLayer(net, 64, (5, 5), pad='same', nonlinearity=None)))
net = ll.Pool2DLayer(net, (2, 2), mode='average_inc_pad')
net = ll.BatchNormLayer(ll.Conv2DLayer(net, 10, (1, 1), nonlinearity=None))
net = ll.DenseLayer(net, 625, b=None, nonlinearity=None)
net = ll.ReshapeLayer(net, ([0], 1, 25, 25))
return net
def build_correlation_fft(x, y, size):
pnet = ll.InputLayer((None, 3, 101, 101), input_var=None)
pnet = ll.BatchNormLayer(pnet)
pnet = ll.Conv2DLayer(pnet, 64, (3, 3), pad='same', nonlinearity=None)
pnet = ll.NonlinearityLayer(
ll.BatchNormLayer(pnet),
nonlinearity=l.nonlinearities.LeakyRectify(0.1))
pnet = ll.Pool2DLayer(pnet, (3, 3), stride=(2, 2))
pnet = ll.Conv2DLayer(pnet, 64, (3, 3), pad='same', nonlinearity=None)
pnet = ll.NonlinearityLayer(
ll.BatchNormLayer(pnet),
nonlinearity=l.nonlinearities.LeakyRectify(0.1))
pnet = ll.Conv2DLayer(pnet, 32, (3, 3), pad='same', nonlinearity=None)
pnet = ll.BatchNormLayer(pnet)
x_p, y_p = ll.get_output(pnet, x), ll.get_output(pnet, y)
x_p, y_p = fft.rfft(x_p, 'ortho'), fft.rfft(y_p, 'ortho')
XX, XY = T.zeros_like(x_p), T.zeros_like(y_p)
XX = T.set_subtensor(
XX[:, :, :, :, 0], x_p[:, :, :, :, 0] * x_p[:, :, :, :, 0] +
x_p[:, :, :, :, 1] * x_p[:, :, :, :, 1])
XY = T.set_subtensor(
XY[:, :, :, :, 0], x_p[:, :, :, :, 0] * y_p[:, :, :, :, 0] +
x_p[:, :, :, :, 1] * y_p[:, :, :, :, 1])
XY = T.set_subtensor(
XY[:, :, :, :, 1], x_p[:, :, :, :, 0] * y_p[:, :, :, :, 1] -
x_p[:, :, :, :, 1] * y_p[:, :, :, :, 0])
xx = fft.irfft(XX, 'ortho')
xy = fft.irfft(XY, 'ortho')
z_p = T.concatenate((xx, xy), axis=1)
z_p *= T.constant(hanningwindow(50))
net = ll.InputLayer((None, 64, 50, 50), input_var=z_p)
net = ll.BatchNormLayer(net)
net = ll.NonlinearityLayer(
ll.BatchNormLayer(
ll.Conv2DLayer(net, 64, (5, 5), pad='same', nonlinearity=None)))
net = ll.Pool2DLayer(net, (2, 2), mode='average_inc_pad')
net = ll.NonlinearityLayer(
ll.BatchNormLayer(
ll.Conv2DLayer(net, 64, (5, 5), pad='same', nonlinearity=None)))
net = ll.BatchNormLayer(ll.Conv2DLayer(net, 10, (1, 1), nonlinearity=None))
net = ll.DenseLayer(net, size**2, b=None, nonlinearity=None)
net = ll.ReshapeLayer(net, ([0], 1, size, size))
return pnet, net
def conv_bn_rectify(net, num_filters):
net = ConvLayer(net,
int(num_filters),
3,
W=init.Normal(),
pad=1,
nonlinearity=None)
net = BatchNormLayer(net, epsilon=1e-3)
net = ll.NonlinearityLayer(net)
return net
def conv_bn_rectify(net, num_filters):
net = layers.Conv2DVarDropOutARD(net,
int(num_filters),
3,
W=init.Normal(),
pad=1,
nonlinearity=nl.linear)
net = BatchNormLayer(net, epsilon=1e-3)
net = ll.NonlinearityLayer(net)
return net
def batch_norm(layer, **kwargs):
"""
Apply batch normalization to an existing layer. This is a convenience
function modifying an existing layer to include batch normalization: It
will steal the layer's nonlinearity if there is one (effectively
introducing the normalization right before the nonlinearity), remove
the layer's bias if there is one (because it would be redundant), and add
a :class:`BatchNormLayer` and :class:`NonlinearityLayer` on top.
Parameters
----------
layer : A :class:`Layer` instance
The layer to apply the normalization to; note that it will be
irreversibly modified as specified above
**kwargs
Any additional keyword arguments are passed on to the
:class:`BatchNormLayer` constructor.
Returns
-------
BatchNormLayer or NonlinearityLayer instance
A batch normalization layer stacked on the given modified `layer`, or
a nonlinearity layer stacked on top of both if `layer` was nonlinear.
Examples
--------
Just wrap any layer into a :func:`batch_norm` call on creating it:
>>> from lasagne.layers import InputLayer, DenseLayer, batch_norm
>>> from lasagne.nonlinearities import tanh
>>> l1 = InputLayer((64, 768))
>>> l2 = batch_norm(DenseLayer(l1, num_units=500, nonlinearity=tanh))
This introduces batch normalization right before its nonlinearity:
>>> from lasagne.layers import get_all_layers
>>> [l.__class__.__name__ for l in get_all_layers(l2)]
['InputLayer', 'DenseLayer', 'BatchNormLayer', 'NonlinearityLayer']
"""
nonlinearity = getattr(layer, 'nonlinearity', None)
if nonlinearity is not None:
layer.nonlinearity = lasagne.nonlinearities.identity
if hasattr(layer, 'b') and layer.b is not None:
del layer.params[layer.b]
layer.b = None
layer = BatchNormLayer(layer, **kwargs)
if nonlinearity is not None:
layer = L.NonlinearityLayer(layer, nonlinearity)
return layer
def net_vgglike(k, input_shape, nclass):
input_x, target_y, Winit = T.tensor4("input"), T.vector(
"target", dtype='int32'), init.Normal()
net = ll.InputLayer(input_shape, input_x)
net = conv_bn_rectify(net, 64 * k)
net = ll.DropoutLayer(net, 0.3)
net = conv_bn_rectify(net, 64 * k)
net = MaxPool2DLayer(net, 2, 2)
net = conv_bn_rectify(net, 128 * k)
net = ll.DropoutLayer(net, 0.4)
net = conv_bn_rectify(net, 128 * k)
net = MaxPool2DLayer(net, 2, 2)
net = conv_bn_rectify(net, 256 * k)
net = ll.DropoutLayer(net, 0.4)
net = conv_bn_rectify(net, 256 * k)
net = ll.DropoutLayer(net, 0.4)
net = conv_bn_rectify(net, 256 * k)
net = MaxPool2DLayer(net, 2, 2)
net = conv_bn_rectify(net, 512 * k)
net = ll.DropoutLayer(net, 0.4)
net = conv_bn_rectify(net, 512 * k)
net = ll.DropoutLayer(net, 0.4)
net = conv_bn_rectify(net, 512 * k)
net = MaxPool2DLayer(net, 2, 2)
net = conv_bn_rectify(net, 512 * k)
net = ll.DropoutLayer(net, 0.4)
net = conv_bn_rectify(net, 512 * k)
net = ll.DropoutLayer(net, 0.4)
net = conv_bn_rectify(net, 512 * k)
net = MaxPool2DLayer(net, 2, 2)
net = ll.DenseLayer(net,
int(512 * k),
W=init.Normal(),
nonlinearity=nl.rectify)
net = BatchNormLayer(net, epsilon=1e-3)
net = ll.NonlinearityLayer(net)
net = ll.DropoutLayer(net, 0.5)
net = ll.DenseLayer(net, nclass, W=init.Normal(), nonlinearity=nl.softmax)
return net, input_x, target_y, k
def build_transition_down(incoming,
reduction,
p=0.1,
W_init=lasagne.init.GlorotUniform(),
b_init=None):
""""Builds a transition in the DenseNet model.
Transitions consist of the sequence: Batch Normalization, 1x1 Convolution,
2x2 Average Pooling. The channels can be compressed by specifying
0 < m <= 1, where num_channels = channels * m.
"""
num_filters = int(incoming.output_shape[1] * reduction)
network = nn.BatchNormLayer(incoming)
network = nn.NonlinearityLayer(network, lasagne.nonlinearities.rectify)
network = nn.Conv2DLayer(network, num_filters, 1, W=W_init, b=b_init)
if p > 0:
network = nn.DropoutLayer(network, p=p)
return nn.Pool2DLayer(network, 2, 2, mode='max')
def _build(self):
layer = layers.InputLayer(shape=(None, 3, 32, 32), input_var=self.X)
layer = nin(layer,
conv_filters=192,
filter_size=(5, 5),
pad=2,
cccp1_filters=160,
cccp2_filters=96)
layer = layers.Pool2DLayer(layer,
pool_size=(3, 3),
stride=2,
pad=(0, 0),
ignore_border=False,
mode='max')
layer = layers.DropoutLayer(layer, p=0.5)
layer = nin(layer,
conv_filters=192,
filter_size=(5, 5),
pad=2,
cccp1_filters=192,
cccp2_filters=192)
layer = layers.Pool2DLayer(layer,
pool_size=(3, 3),
stride=2,
ignore_border=False,
mode='average_exc_pad')
layer = layers.DropoutLayer(layer, p=0.5)
layer = nin(layer,
conv_filters=192,
filter_size=(3, 3),
pad=1,
cccp1_filters=192,
cccp2_filters=10)
layer = layers.Pool2DLayer(layer,
pool_size=(8, 8),
stride=1,
ignore_border=False,
mode='average_exc_pad')
layer = layers.flatten(layer, outdim=2)
layer = layers.NonlinearityLayer(layer,
nonlinearity=nonlinearities.softmax)
return layer
def _build_network(self, state_var, action_var):
"""Builds critic network:; inputs: (state, action), outputs: Q-val."""
# States -> Hidden
state_in = nn.InputLayer((None, ) + self.state_shape, state_var)
states = nn.DenseLayer(state_in, 30, W_init, b_init, relu)
states = nn.DenseLayer(states, 30, W_init, b_init, nonlinearity=None)
# Actions -> Hidden
action_in = nn.InputLayer((None, self.num_actions), action_var)
actions = nn.DenseLayer(action_in,
30,
W_init,
b=None,
nonlinearity=None)
# States_h + Actions_h -> Output
net = nn.ElemwiseSumLayer([states, actions])
net = nn.NonlinearityLayer(net, relu)
return nn.DenseLayer(net, 1, W_out, b_out, nonlinearity=None)
def conv2d(incoming, n_filters, filter_size, stride, pool_size, nonlinearity,
batch_norm, name, verbose, *args, **kwargs):
""""""
if stride is None:
stride = (1, 1)
layer = L.Conv2DLayer(incoming,
num_filters=n_filters,
filter_size=filter_size,
stride=stride,
pad='same',
nonlinearity=None,
name=name)
if batch_norm:
name += '.bn'
layer = L.BatchNormLayer(layer, name=name)
name += '.nonlin'
layer = L.NonlinearityLayer(layer, nonlinearity=nonlinearity)
return layer
def build_block(
incoming,
num_layers,
num_filters,
use_linear_skip=True,
filter_size=3,
p=0.1,
W_init=lasagne.init.GlorotUniform(),
b_init=None,
nonlinearity=lasagne.nonlinearities.rectify,
):
"""Builds a block in the DenseNet model."""
feature_maps = [incoming]
for i in xrange(num_layers):
if len(feature_maps) == 1:
network = incoming
else:
network = nn.ConcatLayer(feature_maps, axis=1)
network = nn.BatchNormLayer(network)
network = nn.NonlinearityLayer(network, nonlinearity)
network = nn.Conv2DLayer(network,
num_filters,
filter_size,
pad='same',
W=W_init,
b=b_init)
if p > 0:
network = nn.DropoutLayer(network, p=p)
feature_maps.append(network)
# Whether to return all connections (vanilla DenseNet), or to return only
# those feature maps created in the current block used in upscale path for
# semantic segmentation (100 layer tiramisu)
if use_linear_skip:
return nn.ConcatLayer(feature_maps, axis=1)
return nn.ConcatLayer(feature_maps[1:], axis=1)
def model(self, query_input, batch_size, query_vocab_size,
context_vocab_size, emb_dim_size):
l_input = L.InputLayer(shape=(batch_size, ), input_var=query_input)
l_embed_continuous = L.EmbeddingLayer(l_input,
input_size=query_vocab_size,
output_size=emb_dim_size)
l_values_discrete = L.EmbeddingLayer(l_input,
input_size=query_vocab_size,
output_size=emb_dim_size)
l_probabilities_discrete = L.NonlinearityLayer(
l_values_discrete, nonlinearity=lasagne.nonlinearities.softmax)
l_embed_discrete = StochasticLayer(l_probabilities_discrete,
estimator='MF')
l_merge = L.ElemwiseSumLayer([l_embed_continuous, l_embed_discrete])
l_out = L.DenseLayer(l_merge,
num_units=emb_dim_size,
nonlinearity=lasagne.nonlinearities.softmax)
l_merge_2 = L.ElemwiseMergeLayer([l_out, l_embed_discrete],
merge_function=T.mul)
l_final_out = L.DenseLayer(l_merge_2, num_units=context_vocab_size)
return l_values_discrete, l_final_out
def build_segmenter_simple_absurd_res():
sys.setrecursionlimit(1500)
inp = ll.InputLayer(shape=(None, 1, None, None), name='input')
n_layers = 64 # should get a 128 x 128 receptive field
layers = [inp]
for i in range(n_layers):
# every 2 layers, add a skip connection
layers.append(
ll.Conv2DLayer(layers[-1],
num_filters=8,
filter_size=(3, 3),
pad='same',
W=Orthogonal(),
nonlinearity=linear,
name='conv%d' % (i + 1)))
layers.append(ll.BatchNormLayer(layers[-1], name='bn%i' % (i + 1)))
if (i % 2 == 0) and (i != 0):
layers.append(
ll.ElemwiseSumLayer([
layers[-1], # prev layer
layers[-6],
] # 3 actual layers per block, skip the previous block
))
layers.append(ll.NonlinearityLayer(layers[-1], nonlinearity=rectify))
# our output layer is also convolutional, remember that our Y is going to be the same exact size as the
conv_final = ll.Conv2DLayer(layers[-1],
num_filters=2,
filter_size=(3, 3),
pad='same',
W=Orthogonal(),
name='conv_final',
nonlinearity=linear)
# we need to reshape it to be a (batch*n*m x 3), i.e. unroll s.t. the feature dimension is preserved
softmax = Softmax4D(conv_final, name='4dsoftmax')
return [softmax]
def sigmoid(_in, **kwargs):
return L.NonlinearityLayer(_in, nonlinearity=NL.sigmoid)
gen_layers = [
LL.InputLayer(shape=(args.batch_size, 10**2 + y_dim), input_var=Gen_input)
]
gen_layers.append(
LL.DenseLayer(gen_layers[-1],
num_units=128,
nonlinearity=lasagne.nonlinearities.rectify,
W=xavier_init([100 + y_dim, 128])))
gen_layers.append(
LL.DenseLayer(gen_layers[-1],
num_units=784,
W=xavier_init([128, 784]),
nonlinearity=lasagne.nonlinearities.identity))
gen_layers.append(
LL.NonlinearityLayer(gen_layers[-1],
nonlinearity=lasagne.nonlinearities.sigmoid))
####Discriminator D(X)######
'''
Composed of only two layers,
input is an image of dim=784 + number_of_classes, and returns logit value (not softmax probability) in the last layer
'''
dis_layers = [
LL.InputLayer(shape=(args.batch_size, 28**2 + y_dim), input_var=Dis_input)
]
dis_layers.append(
LL.DenseLayer(dis_layers[-1],
num_units=128,
nonlinearity=lasagne.nonlinearities.rectify,
W=xavier_init([784 + y_dim, 128])))
dis_layers.append(
def resblock(net_in, filters, kernel_size, stride=1, preactivated=True, block_id=1, name=''):
# Show input shape
#log.p(("\t\t" + name + " IN SHAPE:", l.get_output_shape(net_in)), new_line=False)
# Pre-activation
if block_id > 1:
net_pre = l.NonlinearityLayer(net_in, nonlinearity=nl.rectify)
else:
net_pre = net_in
# Pre-activated shortcut?
if preactivated:
net_in = net_pre
# Bottleneck Convolution
if stride > 1:
net_pre = l.batch_norm(l.Conv2DLayer(net_pre,
num_filters=l.get_output_shape(net_pre)[1],
filter_size=1,
pad='same',
stride=1,
nonlinearity=nl.rectify))
# First Convolution
net = l.batch_norm(l.Conv2DLayer(net_pre,
num_filters=l.get_output_shape(net_pre)[1],
filter_size=kernel_size,
pad='same',
stride=1,
nonlinearity=nl.rectify))
# Pooling layer
if stride > 1:
net = l.MaxPool2DLayer(net, pool_size=(stride, stride))
# Dropout Layer
net = l.DropoutLayer(net)
# Second Convolution
net = l.batch_norm(l.Conv2DLayer(net,
num_filters=filters,
filter_size=kernel_size,
pad='same',
stride=1,
nonlinearity=None))
# Shortcut Layer
if not l.get_output_shape(net) == l.get_output_shape(net_in):
# Average pooling
shortcut = l.Pool2DLayer(net_in, pool_size=(stride, stride), stride=stride, mode='average_exc_pad')
# Shortcut convolution
shortcut = l.batch_norm(l.Conv2DLayer(shortcut,
num_filters=filters,
filter_size=1,
pad='same',
stride=1,
nonlinearity=None))
else:
# Shortcut = input
shortcut = net_in
# Merge Layer
out = l.ElemwiseSumLayer([net, shortcut])
# Show output shape
#log.p(("OUT SHAPE:", l.get_output_shape(out), "LAYER:", len(l.get_all_layers(out)) - 1))
return out
def buildNet():
log.p('BUILDING BirdNET MODEL...', new_line=False)
# Input layer for images
net = l.InputLayer((None, cfg.IM_DIM, cfg.IM_SIZE[1], cfg.IM_SIZE[0]))
# Pre-processing stage
#log.p(("\tPRE-PROCESSING STAGE:"))
net = l.batch_norm(l.Conv2DLayer(net,
num_filters=int(FILTERS[0] * RESNET_K),
filter_size=(5, 5),
pad='same',
nonlinearity=nl.rectify))
#log.p(("\t\tFIRST CONV OUT SHAPE:", l.get_output_shape(net), "LAYER:", len(l.get_all_layers(net)) - 1))
# Max pooling
net = l.MaxPool2DLayer(net, pool_size=(1, 2))
#log.p(("\t\tPRE-MAXPOOL OUT SHAPE:", l.get_output_shape(net), "LAYER:", len(l.get_all_layers(net)) - 1))
# Residual Stacks
for i in range(1, len(FILTERS)):
#log.p(("\tRES STACK", i, ':'))
net = resblock(net,
filters=int(FILTERS[i] * RESNET_K),
kernel_size=KERNEL_SIZES[i],
stride=2,
preactivated=True,
block_id=i,
name='BLOCK ' + str(i) + '-1')
for j in range(1, RESNET_N):
net = resblock(net,
filters=int(FILTERS[i] * RESNET_K),
kernel_size=KERNEL_SIZES[i],
preactivated=False,
block_id=i+j,
name='BLOCK ' + str(i) + '-' + str(j + 1))
# Post Activation
net = l.batch_norm(net)
net = l.NonlinearityLayer(net, nonlinearity=nl.rectify)
# Classification branch
#log.p(("\tCLASS BRANCH:"))
net = classificationBranch(net, (4, 10))
#log.p(("\t\tBRANCH OUT SHAPE:", l.get_output_shape(net), "LAYER:", len(l.get_all_layers(net)) - 1))
# Pooling
net = l.GlobalPoolLayer(net, pool_function=logmeanexp)
#log.p(("\tGLOBAL POOLING SHAPE:", l.get_output_shape(net), "LAYER:", len(l.get_all_layers(net)) - 1))
# Sigmoid output
net = l.NonlinearityLayer(net, nonlinearity=nl.sigmoid)
#log.p(("\tFINAL NET OUT SHAPE:", l.get_output_shape(net), "LAYER:", len(l.get_all_layers(net))))
log.p("DONE!")
# Model stats
#log.p(("MODEL HAS", (sum(hasattr(layer, 'W') for layer in l.get_all_layers(net))), "WEIGHTED LAYERS"))
#log.p(("MODEL HAS", l.count_params(net), "PARAMS"))
return net
def build_resnet_model():
log.i('BUILDING RESNET MODEL...')
# Random Seed
lasagne_random.set_rng(cfg.getRandomState())
# Input layer for images
net = l.InputLayer((None, cfg.IM_DIM, cfg.IM_SIZE[1], cfg.IM_SIZE[0]))
# First Convolution
net = l.Conv2DLayer(net,
num_filters=cfg.FILTERS[0],
filter_size=cfg.KERNEL_SIZES[0],
pad='same',
W=initialization(cfg.NONLINEARITY),
nonlinearity=None)
log.i(("\tFIRST CONV OUT SHAPE:", l.get_output_shape(net), "LAYER:",
len(l.get_all_layers(net)) - 1))
# Residual Stacks
for i in range(0, len(cfg.FILTERS)):
net = resblock(net,
filters=cfg.FILTERS[i] * cfg.RESNET_K,
kernel_size=cfg.KERNEL_SIZES[i],
stride=2,
num_groups=cfg.NUM_OF_GROUPS[i])
for _ in range(1, cfg.RESNET_N):
net = resblock(net,
filters=cfg.FILTERS[i] * cfg.RESNET_K,
kernel_size=cfg.KERNEL_SIZES[i],
num_groups=cfg.NUM_OF_GROUPS[i],
preactivated=False)
log.i(("\tRES STACK", i + 1, "OUT SHAPE:", l.get_output_shape(net),
"LAYER:", len(l.get_all_layers(net)) - 1))
# Post Activation
net = batch_norm(net)
net = l.NonlinearityLayer(net, nonlinearity=nonlinearity(cfg.NONLINEARITY))
# Pooling
net = l.GlobalPoolLayer(net)
log.i(("\tFINAL POOLING SHAPE:", l.get_output_shape(net), "LAYER:",
len(l.get_all_layers(net)) - 1))
# Classification Layer
net = l.DenseLayer(net,
len(cfg.CLASSES),
nonlinearity=nonlinearity('identity'),
W=initialization('identity'))
net = l.NonlinearityLayer(net, nonlinearity=nonlinearity('softmax'))
log.i(("\tFINAL NET OUT SHAPE:", l.get_output_shape(net), "LAYER:",
len(l.get_all_layers(net))))
log.i("...DONE!")
# Model stats
log.i(("MODEL HAS",
(sum(hasattr(layer, 'W')
for layer in l.get_all_layers(net))), "WEIGHTED LAYERS"))
log.i(("MODEL HAS", l.count_params(net), "PARAMS"))
return net
def resblock(net_in,
filters,
kernel_size,
stride=1,
num_groups=1,
preactivated=True):
# Preactivation
net_pre = batch_norm(net_in)
net_pre = l.NonlinearityLayer(net_pre,
nonlinearity=nonlinearity(cfg.NONLINEARITY))
# Preactivated shortcut?
if preactivated:
net_sc = net_pre
else:
net_sc = net_in
# Stride size
if cfg.MAX_POOLING:
s = 1
else:
s = stride
# First Convolution (alwys has preactivated input)
net = batch_norm(
l.Conv2DLayer(net_pre,
num_filters=filters,
filter_size=kernel_size,
pad='same',
stride=s,
num_groups=num_groups,
W=initialization(cfg.NONLINEARITY),
nonlinearity=nonlinearity(cfg.NONLINEARITY)))
# Optional pooling layer
if cfg.MAX_POOLING and stride > 1:
net = l.MaxPool2DLayer(net, pool_size=stride)
# Dropout Layer (we support different types of dropout)
if cfg.DROPOUT_TYPE == 'channels' and cfg.DROPOUT > 0.0:
net = l.dropout_channels(net, p=cfg.DROPOUT)
elif cfg.DROPOUT_TYPE == 'location' and cfg.DROPOUT > 0.0:
net = l.dropout_location(net, p=cfg.DROPOUT)
elif cfg.DROPOUT > 0.0:
net = l.DropoutLayer(net, p=cfg.DROPOUT)
# Second Convolution
net = l.Conv2DLayer(net,
num_filters=filters,
filter_size=kernel_size,
pad='same',
stride=1,
num_groups=num_groups,
W=initialization(cfg.NONLINEARITY),
nonlinearity=None)
# Shortcut Layer
if not l.get_output_shape(net) == l.get_output_shape(net_sc):
shortcut = l.Conv2DLayer(net_sc,
num_filters=filters,
filter_size=1,
pad='same',
stride=s,
W=initialization(cfg.NONLINEARITY),
nonlinearity=None,
b=None)
# Optional pooling layer
if cfg.MAX_POOLING and stride > 1:
shortcut = l.MaxPool2DLayer(shortcut, pool_size=stride)
else:
shortcut = net_sc
# Merge Layer
out = l.ElemwiseSumLayer([net, shortcut])
return out
def __init__(self,
incomings,
vocab_size,
emb_size,
A=lasagne.init.Normal(std=0.1),
C=lasagne.init.Normal(std=0.1),
AT=lasagne.init.Normal(std=0.1),
CT=lasagne.init.Normal(std=0.1),
nonlin=lasagne.nonlinearities.softmax,
RN=0.,
**kwargs):
super(MemoryLayer, self).__init__(incomings, **kwargs)
self.vocab_size, self.emb_size = vocab_size, emb_size
self.nonlin = nonlin
self.RN = RN
# self.A, self.C, self.AT, self.CT = A, C, AT, CT
batch_size, c_count, c_length = self.input_shapes[0]
_, q_count, _ = self.input_shapes[2]
self.l_c_in = LL.InputLayer(shape=(batch_size, c_count, c_length))
self.l_c_in_pe = LL.InputLayer(shape=(batch_size, c_count, c_length,
self.emb_size))
self.l_u_in = LL.InputLayer(shape=(batch_size, q_count, self.emb_size))
self.l_c_A_enc = EncodingFullLayer((self.l_c_in, self.l_c_in_pe),
self.vocab_size, self.emb_size, A,
AT)
self.l_c_C_enc = EncodingFullLayer((self.l_c_in, self.l_c_in_pe),
self.vocab_size, self.emb_size, C,
CT)
self.A, self.C = self.l_c_A_enc.W, self.l_c_C_enc.W
self.AT, self.CT = self.l_c_A_enc.WT, self.l_c_C_enc.WT
if len(incomings
) == 4: # if there is also the probabilities over sentences
self.l_in_ac_prob = LL.InputLayer(shape=(batch_size, c_count,
emb_size))
self.l_c_A_enc_ = LL.ElemwiseMergeLayer(
(self.l_c_A_enc, self.l_in_ac_prob), merge_function=T.mul)
self.l_c_C_enc_ = LL.ElemwiseMergeLayer(
(self.l_c_C_enc, self.l_in_ac_prob), merge_function=T.mul)
self.l_u_in_tr = LL.DimshuffleLayer(self.l_u_in, pattern=(0, 2, 1))
if len(incomings) == 4:
self.l_p = BatchedDotLayer((self.l_c_A_enc_, self.l_u_in_tr))
else:
self.l_p = BatchedDotLayer((self.l_c_A_enc, self.l_u_in_tr))
if self.l_p.output_shape[2] == 1:
self.l_p = LL.FlattenLayer(self.l_p, outdim=2)
# self.l_p = LL.DimshuffleLayer(self.l_p, (0, 1))
if self.nonlin == 'MaxOut':
raise NotImplementedError
self.l_p = LL.NonlinearityLayer(self.l_p, nonlinearity=nonlin)
self.l_p = LL.DimshuffleLayer(self.l_p, (0, 1, 'x'))
# self.l_p = LL.ReshapeLayer(self.l_p, self.l_p.output_shape + (1,))
self.l_p = LL.ExpressionLayer(self.l_p,
lambda X: X.repeat(emb_size, 2),
output_shape='auto')
## self.l_p = RepeatDimLayer(self.l_p, emb_size, axis=2)
if len(incomings) == 4:
self.l_pc = LL.ElemwiseMergeLayer((self.l_p, self.l_c_C_enc_),
merge_function=T.mul)
else:
self.l_pc = LL.ElemwiseMergeLayer((self.l_p, self.l_c_C_enc),
merge_function=T.mul)
self.l_o = LL.ExpressionLayer(self.l_pc,
lambda X: X.sum(1),
output_shape='auto')
# self.l_o = SumLayer(self.l_pc, axis=1)
self.l_o = LL.DimshuffleLayer(self.l_o, pattern=(0, 'x', 1))
self.l_o_u = LL.ElemwiseMergeLayer((self.l_o, self.l_u_in),
merge_function=T.add)
params = LL.helper.get_all_params(self.l_o_u, trainable=True)
values = LL.helper.get_all_param_values(self.l_o_u, trainable=True)
for p, v in zip(params, values):
self.add_param(p, v.shape, name=p.name)
dropout = lambda p=0.1, rescale=True: lambda incoming: \
layers.DropoutLayer(incoming, p=p, rescale=rescale) if p is not None else incoming
batch_norm = lambda axes='auto': lambda incoming: layers.BatchNormLayer(
incoming, axes=axes)
class Select(object):
def __getitem__(self, item):
return lambda incomings: incomings[item]
select = Select()
take = select
nonlinearity = lambda f=None: lambda incoming: layers.NonlinearityLayer(
incoming, (nonlinearities.LeakyRectify(0.05) if f is None else f))
elementwise = lambda f=T.add: lambda incomings: layers.ElemwiseMergeLayer(
incomings, f)
elementwise_sum = lambda: lambda incomings: layers.ElemwiseMergeLayer(
incomings, T.add)
elementwise_mean = lambda: lambda incomings: \
nonlinearity(f=lambda x: x / len(incomings))(layers.ElemwiseMergeLayer(incomings, T.add))
flatten = lambda outdim=2: lambda incoming: layers.FlattenLayer(incoming,
outdim=outdim)
feature_pool = lambda pool_size=4, axis=1, f=T.max: lambda incoming: \
layers.FeaturePoolLayer(incoming, pool_size=pool_size, axis=axis, pool_function=f)
def conv_nonl(data, num_filters, name, pad, use_bn=True):
res = conv(data, num_filters, name, pad=pad)
if (use_bn):
res = L.BatchNormLayer(res, name='bn_' + name)
res = L.NonlinearityLayer(res, rectify, name='relu_' + name)
return res
def build_deconv_network():
input_var = theano.tensor.tensor4('input_var')
net = {}
net['input'] = layers.InputLayer(shape=(None, 3, PS, PS),
input_var=input_var)
# Encoding part
net['conv1_1'] = layers.NonlinearityLayer(
batch_norm(
layers.Conv2DLayer(net['input'],
64,
filter_size=(3, 3),
stride=1,
pad=1)))
net['conv1_2'] = layers.NonlinearityLayer(
batch_norm(
layers.Conv2DLayer(net['conv1_1'],
64,
filter_size=(3, 3),
stride=1,
pad=1)))
net['pool1'] = layers.Pool2DLayer(net['conv1_2'],
pool_size=(2, 2),
stride=2,
mode='max')
net['conv2_1'] = layers.NonlinearityLayer(
batch_norm(
layers.Conv2DLayer(net['pool1'],
128,
filter_size=(3, 3),
stride=1,
pad=1)))
net['conv2_2'] = layers.NonlinearityLayer(
batch_norm(
layers.Conv2DLayer(net['conv2_1'],
128,
filter_size=(3, 3),
stride=1,
pad=1)))
net['pool2'] = layers.Pool2DLayer(net['conv2_2'],
pool_size=(2, 2),
stride=2,
mode='max')
net['conv3_1'] = layers.NonlinearityLayer(
batch_norm(
layers.Conv2DLayer(net['pool2'],
256,
filter_size=(3, 3),
stride=1,
pad=1)))
net['conv3_2'] = layers.NonlinearityLayer(
batch_norm(
layers.Conv2DLayer(net['conv3_1'],
256,
filter_size=(3, 3),
stride=1,
pad=1)))
net['conv3_3'] = layers.NonlinearityLayer(
batch_norm(
layers.Conv2DLayer(net['conv3_2'],
256,
filter_size=(3, 3),
stride=1,
pad=1)))
net['pool3'] = layers.Pool2DLayer(net['conv3_3'],
pool_size=(2, 2),
stride=2,
mode='max')
net['conv4_1'] = layers.NonlinearityLayer(
batch_norm(
layers.Conv2DLayer(net['pool3'],
512,
filter_size=(3, 3),
stride=1,
pad=1)))
net['conv4_2'] = layers.NonlinearityLayer(
batch_norm(
layers.Conv2DLayer(net['conv4_1'],
512,
filter_size=(3, 3),
stride=1,
pad=1)))
net['conv4_3'] = layers.NonlinearityLayer(
batch_norm(
layers.Conv2DLayer(net['conv4_2'],
512,
filter_size=(3, 3),
stride=1,
pad=1)))
net['pool4'] = layers.Pool2DLayer(net['conv4_3'],
pool_size=(2, 2),
stride=2,
mode='max')
net['conv5_1'] = layers.NonlinearityLayer(
batch_norm(
layers.Conv2DLayer(net['pool4'],
512,
filter_size=(3, 3),
stride=1,
pad=1)))
net['conv5_2'] = layers.NonlinearityLayer(
batch_norm(
layers.Conv2DLayer(net['conv5_1'],
512,
filter_size=(3, 3),
stride=1,
pad=1)))
net['conv5_3'] = layers.NonlinearityLayer(
batch_norm(
layers.Conv2DLayer(net['conv5_2'],
512,
filter_size=(3, 3),
stride=1,
pad=1)))
net['pool5'] = layers.Pool2DLayer(net['conv5_3'],
pool_size=(2, 2),
stride=2,
mode='max')
net['fc6'] = layers.NonlinearityLayer(
batch_norm(
layers.Conv2DLayer(net['pool5'],
4096,
filter_size=(7, 7),
stride=1,
pad='same')))
# fc7 is the encoding layer
net['fc7'] = layers.NonlinearityLayer(
batch_norm(
layers.Conv2DLayer(net['fc6'],
4096,
filter_size=(1, 1),
stride=1,
pad='same')))
# Decoding part
net['fc6_deconv'] = layers.NonlinearityLayer(
batch_norm(
layers.Deconv2DLayer(net['fc7'],
512,
filter_size=(7, 7),
stride=1,
crop='same')))
net['unpool5'] = layers.InverseLayer(net['fc6_deconv'], net['pool5'])
net['deconv5_1'] = layers.NonlinearityLayer(
batch_norm(
layers.Deconv2DLayer(net['unpool5'],
512,
filter_size=(3, 3),
stride=1,
crop='same')))
net['deconv5_2'] = layers.NonlinearityLayer(
batch_norm(
layers.Deconv2DLayer(net['deconv5_1'],
512,
filter_size=(3, 3),
stride=1,
crop='same')))
net['deconv5_3'] = layers.NonlinearityLayer(
batch_norm(
layers.Deconv2DLayer(net['deconv5_2'],
512,
filter_size=(3, 3),
stride=1,
crop='same')))
net['unpool4'] = layers.InverseLayer(net['deconv5_3'], net['pool4'])
net['deconv4_1'] = layers.NonlinearityLayer(
batch_norm(
layers.Deconv2DLayer(net['unpool4'],
512,
filter_size=(3, 3),
stride=1,
crop='same')))
net['deconv4_2'] = layers.NonlinearityLayer(
batch_norm(
layers.Deconv2DLayer(net['deconv4_1'],
512,
filter_size=(3, 3),
stride=1,
crop='same')))
net['deconv4_3'] = layers.NonlinearityLayer(
batch_norm(
layers.Deconv2DLayer(net['deconv4_2'],
256,
filter_size=(3, 3),
stride=1,
crop='same')))
net['unpool3'] = layers.InverseLayer(net['deconv4_3'], net['pool3'])
net['deconv3_1'] = layers.NonlinearityLayer(
batch_norm(
layers.Deconv2DLayer(net['unpool3'],
256,
filter_size=(3, 3),
stride=1,
crop='same')))
net['deconv3_2'] = layers.NonlinearityLayer(
batch_norm(
layers.Deconv2DLayer(net['deconv3_1'],
256,
filter_size=(3, 3),
stride=1,
crop='same')))
net['deconv3_3'] = layers.NonlinearityLayer(
batch_norm(
layers.Deconv2DLayer(net['deconv3_2'],
128,
filter_size=(3, 3),
stride=1,
crop='same')))
net['unpool2'] = layers.InverseLayer(net['deconv3_3'], net['pool2'])
net['deconv2_1'] = layers.NonlinearityLayer(
batch_norm(
layers.Deconv2DLayer(net['unpool2'],
128,
filter_size=(3, 3),
stride=1,
crop='same')))
net['deconv2_2'] = layers.NonlinearityLayer(
batch_norm(
layers.Deconv2DLayer(net['deconv2_1'],
64,
filter_size=(3, 3),
stride=1,
crop='same')))
net['unpool1'] = layers.InverseLayer(net['deconv2_2'], net['pool1'])
net['deconv1_1'] = layers.NonlinearityLayer(
batch_norm(
layers.Deconv2DLayer(net['unpool1'],
64,
filter_size=(3, 3),
stride=1,
crop='same')))
net['deconv1_2'] = layers.NonlinearityLayer(
batch_norm(
layers.Deconv2DLayer(net['deconv1_1'],
64,
filter_size=(3, 3),
stride=1,
crop='same')))
# Segmentation layer
net['seg_score'] = layers.Deconv2DLayer(
net['deconv1_2'],
1,
filter_size=(1, 1),
stride=1,
crop='same',
nonlinearity=lasagne.nonlinearities.sigmoid)
network = ReshapeLayer(net['seg_score'], ([0], -1))
output_var = lasagne.layers.get_output(network)
all_param = lasagne.layers.get_all_params(network, trainable=True)
return network, input_var, output_var, all_param
inceptionModule(disc_layers[-1], [64, 112, 144, 288, 32, 64]))
disc_layers.extend(
inceptionModule(disc_layers[-1], [128, 256, 160, 320, 32, 128]))
disc_layers.append(
ll.MaxPool2DLayer(disc_layers[-1],
pool_size=3,
stride=2,
ignore_border=False))
disc_layers.extend(
inceptionModule(disc_layers[-1], [128, 256, 160, 320, 32, 128]))
disc_layers.extend(
inceptionModule(disc_layers[-1], [128, 384, 192, 384, 48, 128]))
disc_layers.append(ll.GlobalPoolLayer(disc_layers[-1]))
disc_layers.append(
ll.DenseLayer(disc_layers[-1], num_units=1000, nonlinearity=linear))
disc_layers.append(ll.NonlinearityLayer(disc_layers[-1], nonlinearity=softmax))
disc_params = ll.get_all_params(disc_layers, trainable=True)
print("DISCRIMINATOR CREATED")
# costs
labels = T.ivector()
x_lab = T.tensor4()
temp = ll.get_output(disc_layers[-1], x_lab, deterministic=False, init=True)
init_updates = [u for l in disc_layers for u in getattr(l, 'init_updates', [])]
output_before_softmax_lab = ll.get_output(disc_layers[-1],
x_lab,
deterministic=False)
l_lab = output_before_softmax_lab[T.arange(args.batch_size), labels]
def __init__(
self,
image_shape,
filter_shape,
num_class,
conv_type,
kernel_size,
kernel_pool_size,
dropout_rate,
):
"""
"""
self.filter_shape = filter_shape
self.n_visible = numpy.prod(image_shape)
self.n_layers = len(filter_shape)
self.rng = RandomStreams(123)
self.x = T.matrix()
self.y = T.ivector()
self.conv_layers = []
NoiseLayer = layers.DropoutLayer
dropout_rate = float(dropout_rate)
self.l_input = layers.InputLayer((None, self.n_visible), self.x)
this_layer = layers.ReshapeLayer(self.l_input, ([0], ) + image_shape)
for l in range(self.n_layers):
activation = lasagne.nonlinearities.rectify
if len(filter_shape[l]) == 3:
if conv_type == 'double' and filter_shape[l][1] > kernel_size:
this_layer = DoubleConvLayer(
this_layer,
filter_shape[l][0],
filter_shape[l][1:],
pad='same',
nonlinearity=activation,
kernel_size=kernel_size,
kernel_pool_size=kernel_pool_size)
this_layer = layers.batch_norm(this_layer)
elif conv_type == 'maxout':
this_layer = layers.Conv2DLayer(this_layer,
filter_shape[l][0],
filter_shape[l][1:],
b=None,
pad='same',
nonlinearity=None)
this_layer = layers.FeaturePoolLayer(
this_layer, pool_size=kernel_pool_size**2)
this_layer = layers.BatchNormLayer(this_layer)
this_layer = layers.NonlinearityLayer(
this_layer, activation)
elif conv_type == 'cyclic':
this_layers = []
this_layers.append(
layers.Conv2DLayer(this_layer,
filter_shape[l][0],
filter_shape[l][1:],
b=None,
pad='same',
nonlinearity=None))
for _ in range(3):
W = this_layers[-1].W.dimshuffle(0, 1, 3,
2)[:, :, :, ::-1]
this_layers.append(
layers.Conv2DLayer(this_layer,
filter_shape[l][0],
filter_shape[l][1:],
W=W,
b=None,
pad='same',
nonlinearity=None))
this_layer = layers.ElemwiseMergeLayer(
this_layers, T.maximum)
this_layer = layers.BatchNormLayer(this_layer)
this_layer = layers.NonlinearityLayer(
this_layer, activation)
elif conv_type == 'standard' \
or (conv_type == 'double' and filter_shape[l][1] <= kernel_size):
this_layer = layers.Conv2DLayer(this_layer,
filter_shape[l][0],
filter_shape[l][1:],
pad='same',
nonlinearity=activation)
this_layer = layers.batch_norm(this_layer)
else:
raise NotImplementedError
self.conv_layers.append(this_layer)
elif len(filter_shape[l]) == 2:
this_layer = layers.MaxPool2DLayer(this_layer, filter_shape[l])
this_layer = NoiseLayer(this_layer, dropout_rate)
elif len(filter_shape[l]) == 1:
raise NotImplementedError
self.top_conv_layer = this_layer
this_layer = layers.GlobalPoolLayer(this_layer, T.mean)
self.clf_layer = layers.DenseLayer(this_layer,
num_class,
W=lasagne.init.Constant(0.),
nonlinearity=T.nnet.softmax)
self.params = layers.get_all_params(self.clf_layer, trainable=True)
self.params_all = layers.get_all_params(self.clf_layer)
def tanh(_in, **kwargs):
return L.NonlinearityLayer(_in, nonlinearity=NL.tanh)
