def build_cnn(input_var, pretrained_model):
#pretrained layers from vgg16
conv1_1 = pretrained_model['conv1_1']
conv1_2 = pretrained_model['conv1_2']
#new layers
network = InputLayer(shape=(None, 3, 48, 48), input_var=input_var)
network = ConvLayer(network,
64,
3,
pad=1,
flip_filters=False,
W=conv1_1.W.get_value(),
b=conv1_1.b.get_value())
network = ConvLayer(network,
64,
3,
pad=1,
flip_filters=False,
W=conv1_2.W.get_value(),
b=conv1_2.b.get_value())
network = MaxPoolLayer(network, pool_size=(2, 2))
network = DenseLayer(dropout(network, p=.5),
num_units=256,
nonlinearity=lasagne.nonlinearities.rectify)
network = DenseLayer(dropout(network, p=.5),
num_units=7,
nonlinearity=lasagne.nonlinearities.softmax)
return network
def build_cnn(config, use_noise=True, use_bn=True):
# NOTE: Neither Conv2DDNNLayer nor Conv2DMMLayer will not work
# with T.Rop operation, which used for the Fisher-vector product.
l_input = L.InputLayer((None, 1, config['height'], config['width']))
l_out = L.Conv2DLayer(l_input,
num_filters=config['cnn_f1'], filter_size=(6,6), stride=2,
nonlinearity=relu, W=LI.HeUniform('relu'), b=LI.Constant(0.)
)
# https://arxiv.org/pdf/1602.01407v2.pdf
# QUOTE: KFC-pre and BN can be combined synergistically.
if use_bn: l_out = L.batch_norm(l_out, beta=None, gamma=None)
l_out = L.Conv2DLayer(l_out,
num_filters=config['cnn_f2'], filter_size=(4,4), stride=2,
nonlinearity=relu, W=LI.HeUniform('relu'), b=LI.Constant(0.)
)
if use_bn: l_out = L.batch_norm(l_out, beta=None, gamma=None)
if use_noise: l_out = L.dropout(l_out)
l_out = L.Conv2DLayer(l_out,
num_filters=config['cnn_f3'], filter_size=(4,4), stride=2,
nonlinearity=relu, W=LI.HeUniform('relu'), b=LI.Constant(0.)
)
if use_bn: l_out = L.batch_norm(l_out, beta=None, gamma=None)
if use_noise: l_out = L.dropout(l_out)
return l_input, l_out
def build_network(self, ra_input_var, mc_input_var):
print('Building raw dnn with parameters:')
pp = pprint.PrettyPrinter(indent=4)
pp.pprint(self.net_opts)
ra_network_1 = layers.InputLayer((None, 1, 3969), ra_input_var)
ra_network_1 = self.set_conv_layer(ra_network_1, 'ra_conv_1', dropout=False, pad='same')
ra_network_1 = self.set_pool_layer(ra_network_1, 'ra_pool_1')
ra_network_1 = self.set_conv_layer(ra_network_1, 'ra_conv_2', pad='same')
ra_network_1 = self.set_pool_layer(ra_network_1, 'ra_pool_2')
ra_network_1 = self.set_conv_layer(ra_network_1, 'ra_conv_3', pad='same')
ra_network_1 = self.set_pool_layer(ra_network_1, 'ra_pool_3')
ra_network_1 = self.set_conv_layer(ra_network_1, 'ra_conv_4', pad='same')
ra_network_1 = self.set_pool_layer(ra_network_1, 'ra_pool_4')
concat_list = [ra_network_1]
mc_input = layers.InputLayer((None, 2, MC_LENGTH), mc_input_var)
concat_list.append(mc_input)
network = layers.ConcatLayer(concat_list, axis=1, cropping=[None, None, 'center'])
network = layers.BatchNormLayer(network)
for n in self.net_opts['layer_list']:
network = layers.DenseLayer(layers.dropout(network, p=self.net_opts['dropout_p']),
n,
nonlinearity=lasagne.nonlinearities.rectify)
network = layers.DenseLayer(layers.dropout(network, p=self.net_opts['dropout_p']),
self.net_opts['num_class'],
nonlinearity=lasagne.nonlinearities.softmax)
# print(layers.get_output_shape(network))
self.network = network
return self.network
def buildControlNN(fingerprint_vals, fingerprint_dim, output_dim, final_layer_type,
dropout_prob=0.0, neural_net=[]):
network_vals = {}
l_in = InputLayer(shape=(None,fingerprint_dim), input_var=fingerprint_vals)
#do dropout
network_vals['drop0'] = dropout(l_in,p=dropout_prob)
#run through the layers I have
for layerNum,hiddenUnits in enumerate(neural_net):
oldLayerNum = layerNum
currLayerNum = layerNum + 1
network_vals['dense'+str(currLayerNum)] = DenseLayer(network_vals['drop'+str(oldLayerNum)], \
hiddenUnits,nonlinearity=lasagne.nonlinearities.rectify)
network_vals['drop'+str(currLayerNum)] = dropout(network_vals['dense'+str(currLayerNum)],p=dropout_prob)
if neural_net == []:
network_vals['final_out'] = l_in
else:
network_vals['final_out'] = network_vals['dense'+str(currLayerNum)]
#finally, project it into the dimensionality we want
network_vals['output'] = DenseLayer(network_vals['final_out'],\
num_units=output_dim, nonlinearity=final_layer_type)
return network_vals
def architecture(input_var, input_shape):
network_trained = {}
network_trained['input'] = InputLayer(input_shape, input_var)
kwargs = dict(nonlinearity=lasagne.nonlinearities.leaky_rectify,
W=lasagne.init.Orthogonal())
network_trained['conv1'] = Conv2DLayer(network_trained['input'], 64, 3,
**kwargs)
network_trained['conv2'] = Conv2DLayer(network_trained['conv1'], 32, 3,
**kwargs)
network_trained['mp3'] = MaxPool2DLayer(network_trained['conv2'], 3)
network_trained['conv4'] = Conv2DLayer(network_trained['mp3'], 128, 3,
**kwargs)
network_trained['conv5'] = Conv2DLayer(network_trained['conv4'], 64, 3,
**kwargs)
network_trained['mp6'] = MaxPool2DLayer(network_trained['conv5'], 3)
network_trained['fc7'] = DenseLayer(dropout(network_trained['mp6'], 0.5),
256, **kwargs)
network_trained['fc8'] = DenseLayer(dropout(network_trained['fc7'], 0.5),
64, **kwargs)
network_trained['fc9'] = DenseLayer(
dropout(network_trained['fc8'], 0.5),
1,
nonlinearity=lasagne.nonlinearities.sigmoid,
W=lasagne.init.Orthogonal())
return network_trained
def build_network(self, mfcc_input_var):
print('Building cnn with parameters:')
pp = pprint.PrettyPrinter(indent=4)
pp.pprint(self.net_opts)
mfcc_network = layers.InputLayer((None, 130, MC_LENGTH), mfcc_input_var)
mfcc_network = layers.BatchNormLayer(mfcc_network)
mfcc_network = self.set_conv_layer(mfcc_network, 'conv_1', bnorm=False)
mfcc_network = self.set_pool_layer(mfcc_network, 'pool_1')
mfcc_network = self.set_conv_layer(mfcc_network, 'conv_2', bnorm=False)
mfcc_network = self.set_pool_layer(mfcc_network, 'pool_2')
for n in self.net_opts['layer_list']:
# mfcc_network = layers.batch_norm(layers.DenseLayer(layers.dropout(mfcc_network, p=self.net_opts['dropout_p']),
# n,
# nonlinearity=lasagne.nonlinearities.rectify)
# )
mfcc_network = layers.DenseLayer(layers.dropout(mfcc_network, p=self.net_opts['dropout_p']),
n,
nonlinearity=lasagne.nonlinearities.rectify)
# mfcc_network = layers.BatchNormLayer(mfcc_network)
mfcc_network = layers.DenseLayer(layers.dropout(mfcc_network, p=self.net_opts['dropout_p']),
self.net_opts['num_class'],
nonlinearity=lasagne.nonlinearities.softmax)
self.network = mfcc_network
return self.network
def build_cascade(input_var, nb_classes, n_chanels=1, input_size=20, reshaped_input_size=20, activity=softmax):
"""
Builds the complete network with 1D-conv1d layer to integrate time from sequences of EEG images.
:param input_vars: list of EEG images (one image per time window)
:param nb_classes: number of classes
:return: a pointer to the output of last layer
"""
# Input layer
network = InputLayer(shape=(None, 1, input_size), input_var=input_var)
network = ReshapeLayer(network, (([0], n_chanels, reshaped_input_size)))
network = Conv1DLayer(network, 1024, 5)
network = MaxPool1DLayer(network, 2)
network = DimshuffleLayer(network, (0, 2, 1))
network = LSTMLayer(network, num_units=256, grad_clipping=100, nonlinearity=tanh)
network = SliceLayer(network, -1, 1)
# A fully-connected layer of 512 units with 50% dropout on its inputs:
network = DenseLayer(dropout(network, p=.5),
num_units=64, nonlinearity=rectify)
# And, finally, the output layer with 50% dropout on its inputs:
network = DenseLayer(dropout(network, p=.5),
num_units=nb_classes, nonlinearity=activity)
return network
def build_model(timesteps, pX, pY):
net = OrderedDict()
net['input'] = InputLayer(shape=(None, timesteps, pX, pY))
net['conv1_1'] = batch_norm(
ConvLayer(net['input'],
num_filters=32,
filter_size=(3, 3),
nonlinearity=ReLU,
pad='same'))
net['conv1_2'] = batch_norm(
ConvLayer(net['conv1_1'],
num_filters=32,
filter_size=(3, 3),
nonlinearity=ReLU,
pad='same'))
net['pool1'] = PoolLayer(net['conv1_2'], pool_size=(2, 2))
net['conv2_1'] = batch_norm(
ConvLayer(net['pool1'],
num_filters=64,
filter_size=(3, 3),
nonlinearity=ReLU,
pad='same'))
net['conv2_2'] = batch_norm(
ConvLayer(net['conv2_1'],
num_filters=64,
filter_size=(3, 3),
nonlinearity=ReLU,
pad='same'))
net['pool2'] = PoolLayer(net['conv2_2'], pool_size=(2, 2))
net['conv3_1'] = batch_norm(
ConvLayer(net['pool2'],
num_filters=128,
filter_size=(3, 3),
nonlinearity=ReLU,
pad='same'))
net['conv3_2'] = batch_norm(
ConvLayer(net['conv3_1'],
num_filters=128,
filter_size=(3, 3),
nonlinearity=ReLU,
pad='same'))
net['pool3'] = PoolLayer(net['conv3_2'], pool_size=(2, 2))
net['fcLayer1'] = batch_norm(
DenseLayer(dropout(net['pool3'], p=0.5),
num_units=512,
nonlinearity=ReLU))
net['output'] = DenseLayer(dropout(net['fcLayer1'], p=0.5),
num_units=2,
nonlinearity=softmax)
return net
def build_network(self, ra_input_var, mc_input_var):
print('Building raw dnn with parameters:')
pp = pprint.PrettyPrinter(indent=4)
pp.pprint(self.net_opts)
ra_network_1 = layers.InputLayer((None, 1, 3969), ra_input_var)
ra_network_1 = self.set_conv_layer(ra_network_1, 'ra_conv_1', dropout=False, pad='same')
ra_network_1 = self.set_pool_layer(ra_network_1, 'ra_pool_1')
ra_network_1 = self.set_conv_layer(ra_network_1, 'ra_conv_2', pad='same')
ra_network_1 = self.set_pool_layer(ra_network_1, 'ra_pool_2')
ra_network_1 = self.set_conv_layer(ra_network_1, 'ra_conv_3', pad='same')
ra_network_1 = self.set_pool_layer(ra_network_1, 'ra_pool_3')
ra_network_1 = self.set_conv_layer(ra_network_1, 'ra_conv_4', pad='same')
ra_network_1 = self.set_pool_layer(ra_network_1, 'ra_pool_4')
concat_list = [ra_network_1]
mc_input = layers.InputLayer((None, 2, MC_LENGTH), mc_input_var)
concat_list.append(mc_input)
network = layers.ConcatLayer(concat_list, axis=1, cropping=[None, None, 'center'])
network = layers.BatchNormLayer(network)
for n in self.net_opts['layer_list']:
network = layers.DenseLayer(layers.dropout(network, p=self.net_opts['dropout_p']),
n,
nonlinearity=lasagne.nonlinearities.rectify)
network = layers.DenseLayer(layers.dropout(network, p=self.net_opts['dropout_p']),
self.net_opts['num_class'],
nonlinearity=lasagne.nonlinearities.softmax)
# print(layers.get_output_shape(network))
self.network = network
return self.network
def build_network(self, mfcc_input_var):
print('Building cnn with parameters:')
pp = pprint.PrettyPrinter(indent=4)
pp.pprint(self.net_opts)
mfcc_network = layers.InputLayer((None, 130, MC_LENGTH), mfcc_input_var)
mfcc_network = layers.BatchNormLayer(mfcc_network)
mfcc_network = self.set_conv_layer(mfcc_network, 'conv_1', bnorm=False)
mfcc_network = self.set_pool_layer(mfcc_network, 'pool_1')
mfcc_network = self.set_conv_layer(mfcc_network, 'conv_2', bnorm=False)
mfcc_network = self.set_pool_layer(mfcc_network, 'pool_2')
for n in self.net_opts['layer_list']:
# mfcc_network = layers.batch_norm(layers.DenseLayer(layers.dropout(mfcc_network, p=self.net_opts['dropout_p']),
# n,
# nonlinearity=lasagne.nonlinearities.rectify)
# )
mfcc_network = layers.DenseLayer(layers.dropout(mfcc_network, p=self.net_opts['dropout_p']),
n,
nonlinearity=lasagne.nonlinearities.rectify)
# mfcc_network = layers.BatchNormLayer(mfcc_network)
mfcc_network = layers.DenseLayer(layers.dropout(mfcc_network, p=self.net_opts['dropout_p']),
self.net_opts['num_class'],
nonlinearity=lasagne.nonlinearities.softmax)
self.network = mfcc_network
return self.network
def __init__(
self,
n_words,
dim_emb,
num_units,
n_classes,
w_emb=None,
dropout=0.2,
use_final=False,
lr=0.001,
pretrain=None,
):
self.n_words = n_words
self.dim_emb = dim_emb
self.num_units = num_units
self.n_classes = n_classes
self.lr = lr
if w_emb is None:
w_emb = init.Normal()
self.l_x = layers.InputLayer((None, None))
self.l_m = layers.InputLayer((None, None))
self.l_emb = layers.EmbeddingLayer(self.l_x, n_words, dim_emb, W=w_emb)
self.l_ebd = self.l_emb
if dropout:
self.l_emb = layers.dropout(self.l_emb, dropout)
if use_final:
self.l_enc = layers.LSTMLayer(self.l_emb,
num_units,
mask_input=self.l_m,
only_return_final=True,
grad_clipping=10.0,
gradient_steps=400)
self.l_rnn = self.l_enc
else:
self.l_enc = layers.LSTMLayer(self.l_emb,
num_units,
mask_input=self.l_m,
only_return_final=False,
grad_clipping=10.0,
gradient_steps=400)
self.l_rnn = self.l_enc
self.l_enc = MeanLayer(self.l_enc, self.l_m)
if dropout:
self.l_enc = layers.dropout(self.l_enc, dropout)
self.l_y = layers.DenseLayer(self.l_enc,
n_classes,
nonlinearity=nonlinearities.softmax)
if pretrain:
self.load_pretrain(pretrain)
def fcn_transfer(params):
""""""
assert 'inputs' in params
layers = L.InputLayer((None, 256 * len(params['inputs'])))
layers = L.dropout(layers)
layers = L.DenseLayer(layers, 1024, nonlinearity=nl.elu)
layers = L.dropout(layers)
layers = L.DenseLayer(layers, 16, nonlinearity=nl.softmax)
return layers
def build_network(input_var, image_size=28, output_dim=10):
nonlin = lasagne.nonlinearities.rectify
W_init = lasagne.init.GlorotUniform()
b_init = lasagne.init.Constant(0.)
input_shape = (None, 1, image_size, image_size)
network = nn.InputLayer(input_shape, input_var)
network = nn.Conv2DLayer(network,
num_filters=64,
filter_size=(3, 3),
nonlinearity=nonlin,
W=W_init,
b=b_init)
network = nn.Conv2DLayer(network,
num_filters=64,
filter_size=(3, 3),
nonlinearity=nonlin,
W=W_init,
b=b_init)
network = nn.MaxPool2DLayer(network, pool_size=(2, 2))
network = nn.Conv2DLayer(network,
num_filters=128,
filter_size=(3, 3),
W=W_init,
b=b_init,
nonlinearity=nonlin)
network = nn.Conv2DLayer(network,
num_filters=128,
filter_size=(3, 3),
W=W_init,
b=b_init,
nonlinearity=nonlin)
network = nn.MaxPool2DLayer(network, pool_size=(2, 2))
network = nn.dropout(network, p=0.5)
network = nn.DenseLayer(network,
num_units=256,
W=W_init,
b=b_init,
nonlinearity=nonlin)
network = nn.dropout(network, p=0.5)
network = nn.DenseLayer(network,
num_units=output_dim,
W=W_init,
b=b_init,
nonlinearity=None)
return network
def build_network(self, ra_input_var, mc_input_var):
print('Building raw network with parameters:')
pp = pprint.PrettyPrinter(indent=4)
pp.pprint(self.net_opts)
ra_network_1 = layers.InputLayer((None, 1, None), ra_input_var)
ra_network_1 = self.set_conv_layer(ra_network_1,
'ra_conv_1',
dropout=False,
pad='same')
ra_network_1 = self.set_pool_layer(ra_network_1, 'ra_pool_1')
ra_network_1 = self.set_conv_layer(ra_network_1,
'ra_conv_2',
pad='same')
ra_network_1 = self.set_pool_layer(ra_network_1, 'ra_pool_2')
ra_network_1 = self.set_conv_layer(ra_network_1,
'ra_conv_3',
pad='same')
ra_network_1 = self.set_pool_layer(ra_network_1, 'ra_pool_3')
ra_network_1 = self.set_conv_layer(ra_network_1,
'ra_conv_4',
pad='same')
ra_network_1 = self.set_pool_layer(ra_network_1, 'ra_pool_4')
concat_list = [ra_network_1]
mc_input = layers.InputLayer((None, 2, None), mc_input_var)
concat_list.append(mc_input)
network = layers.ConcatLayer(concat_list,
axis=1,
cropping=[None, None, 'center'])
network = self.set_conv_layer(network, 'conv_1')
network = self.set_pool_layer(network, 'pool_1')
network = self.set_conv_layer(network, 'conv_2')
network = self.set_pool_layer(network, 'pool_2')
network = self.set_conv_layer(network, 'conv_3')
network = layers.GlobalPoolLayer(
network, getattr(T, self.net_opts['global_pool_func']))
# print(layers.get_output_shape(network))
# network = layers.DenseLayer(layers.dropout(network, p=self.net_opts['dropout_p']),
# self.net_opts['dens_1'],
# nonlinearity=lasagne.nonlinearities.rectify)
network = layers.DenseLayer(
layers.dropout(network, p=self.net_opts['dropout_p']),
self.net_opts['dens_2'],
nonlinearity=lasagne.nonlinearities.rectify)
network = layers.DenseLayer(
layers.dropout(network, p=self.net_opts['dropout_p']),
self.net_opts['num_class'],
nonlinearity=lasagne.nonlinearities.softmax)
# print(layers.get_output_shape(network))
self.network = network
return self.network
def __init__(self, vocab, num_users):
self.vocab = vocab
self._user_id = T.ivector('user ids')
self._good_utterance = T.imatrix('utterance from user')
self._bad_utterance = T.imatrix('utterance not from user')
self.l_utt_enc = Enc(vocab)
self._user_inp = InputLayer((None, ),
input_var=self._user_id,
name='user ids layer')
self.l_user_emb = EmbeddingLayer(self._user_inp,
num_users,
DssmConfig.USER_EMB_SIZE,
name='user embedding')
self.l_user_semantic = DenseLayer(self.l_user_emb,
DssmConfig.SEMANTIC_SPACE_SIZE,
name='user representation')
self.l_user_semantic = dropout(self.l_user_semantic,
p=DssmConfig.DROPOUT_RATE)
self.l_utt_semantic = DenseLayer(self.l_utt_enc.output,
DssmConfig.SEMANTIC_SPACE_SIZE,
name='utterance representation')
self.l_utt_semantic = dropout(self.l_utt_semantic,
p=DssmConfig.DROPOUT_RATE)
self.user_semantic = get_output(self.l_user_semantic)
self.user_semantic_d = get_output(self.l_user_semantic,
deterministic=True)
self.good_utt_semantic = get_output(
self.l_utt_semantic,
inputs={self.l_utt_enc.l_in: self._good_utterance})
self.good_utt_semantic_d = get_output(
self.l_utt_semantic,
inputs={self.l_utt_enc.l_in: self._good_utterance},
deterministic=True)
self.bad_utt_semantic = get_output(
self.l_utt_semantic,
inputs={self.l_utt_enc.l_in: self._bad_utterance})
self.bad_utt_semantic_d = get_output(
self.l_utt_semantic,
inputs={self.l_utt_enc.l_in: self._bad_utterance},
deterministic=True)
self._build_loss_and_ops()
def build_network(input_var=None, input_shape=227):
nf = 32
n = lasagne.nonlinearities.tanh
W_init = lasagne.init.GlorotUniform()
net = nn.InputLayer((None, 3, None, None), input_var=input_var)
# Block 1
net = nn.Conv2DLayer(net, nf, 3, W=W_init, nonlinearity=n, pad='same')
net = nn.Conv2DLayer(net, nf, 3, W=W_init, nonlinearity=n, pad='same')
net = nn.MaxPool2DLayer(net, 2)
# Block 2
net = nn.Conv2DLayer(net, nf * 2, 3, W=W_init, nonlinearity=n, pad='same')
net = nn.Conv2DLayer(net, nf * 2, 3, W=W_init, nonlinearity=n, pad='same')
net = nn.SpatialPyramidPoolingLayer(net, [4, 2, 1],
implementation='kaiming')
net = nn.DenseLayer(net, 512, W=W_init, nonlinearity=n)
net = nn.dropout(net, p=0.5)
net = nn.DenseLayer(net, 128, W=W_init, nonlinearity=n)
return nn.DenseLayer(net, 1, W=W_init, nonlinearity=T.nnet.sigmoid)
def build_computation_graph(input_var, input_shape, dimensions, p_input=0.2, p_weight=0.5):
#dimension[-1][-1] is the last output size of last stacked layer a.k.a size of the image vector
input = InputLayer(shape=input_shape, input_var=input_var, name='input')
input = dropout(input, p=p_input, name='input_drop')
network, classification_branch, features = NecklaceNetwork(input, dimensions, LISTAWithDropout, True, True, True,
False, p_weight)
return network, classification_branch, features
def make_Nconvpool_1dense_branch(view, input_layer, cpdictlist, nhidden=256, dropoutp=0.5):
"""
see: http://lasagne.readthedocs.org/en/latest/modules/layers.html
loop through the `cpdictlist` for set of convolutional filter and pooling
parameter specifications; when through, add a dense layer with dropout
"""
net = {}
convname = ""
mpname = ""
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[mpname]
net[convname] = Conv2DLayer(
layer,
num_filters=cpdict["nfilters"],
filter_size=cpdict["filter_size"],
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.GlorotUniform(),
)
mpname = "maxpool-{}-{}".format(view, i)
net[mpname] = MaxPool2DLayer(net[convname], pool_size=cpdict["pool_size"])
logger.info("Convpool {}".format(mpname))
densename = "dense-{}".format(view)
net[densename] = DenseLayer(
dropout(net[mpname], p=dropoutp), num_units=nhidden, nonlinearity=lasagne.nonlinearities.rectify
)
logger.info("Dense {} with nhidden = {}, dropout = {}".format(densename, nhidden, dropoutp))
return net
def architecture(input_var, input_shape):
layer = InputLayer(input_shape, input_var)
kwargs = dict(nonlinearity=lasagne.nonlinearities.leaky_rectify,
W=lasagne.init.Orthogonal())
layer = Conv2DLayer(layer, 64, 3, **kwargs)
layer = Conv2DLayer(layer, 32, 3, **kwargs)
layer = MaxPool2DLayer(layer, 3)
layer = Conv2DLayer(layer, 128, 3, **kwargs)
layer = Conv2DLayer(layer, 64, 3, **kwargs)
layer = MaxPool2DLayer(layer, 3)
layer = DenseLayer(dropout(layer, 0.5), 256, **kwargs)
layer = DenseLayer(dropout(layer, 0.5), 64, **kwargs)
layer = DenseLayer(dropout(layer, 0.5), 1,
nonlinearity=lasagne.nonlinearities.sigmoid,
W=lasagne.init.Orthogonal())
return layer
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 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
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
def __init__(self, input_layer, output_dim, hidden_sizes,
hidden_act=nonlinearities.tanh,
output_act=nonlinearities.identity,
params=None,
batch_norm=False,
dropout=False):
out_layer = input_layer
param_idx = 0
for hidden_size in hidden_sizes:
w_args = {}
if params is not None:
w_args = dict(W=params[param_idx], b=params[param_idx+1])
out_layer = L.DenseLayer(out_layer, hidden_size, nonlinearity=hidden_act,
**w_args)
if batch_norm:
out_layer = L.batch_norm(out_layer)
if dropout:
out_layer = L.dropout(out_layer)
param_idx += 2
w_args = {}
if params is not None:
w_args = dict(W=params[param_idx], b=params[param_idx + 1])
out_layer = L.DenseLayer(out_layer, output_dim, nonlinearity=output_act,
**w_args)
self.out_layer = out_layer
self.output = L.get_output(self.out_layer)
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_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_network(self, mspec_input_var):
print('Building spec dnn with parameters:')
pp = pprint.PrettyPrinter(indent=4)
pp.pprint(self.net_opts)
mspec_network = layers.InputLayer((None, 130, MC_LENGTH), mspec_input_var)
mspec_network = layers.BatchNormLayer(mspec_network)
for n in self.net_opts['layer_list']:
mspec_network = layers.DenseLayer(layers.dropout(mspec_network, p=self.net_opts['dropout_p']),
n,
nonlinearity=lasagne.nonlinearities.rectify)
mspec_network = layers.DenseLayer(layers.dropout(mspec_network, p=self.net_opts['dropout_p']),
self.net_opts['num_class'],
nonlinearity=lasagne.nonlinearities.softmax)
self.network = mspec_network
return self.network
def build_network(self, mspec_input_var):
print('Building spec dnn with parameters:')
pp = pprint.PrettyPrinter(indent=4)
pp.pprint(self.net_opts)
mspec_network = layers.InputLayer((None, 130, MC_LENGTH), mspec_input_var)
mspec_network = layers.BatchNormLayer(mspec_network)
for n in self.net_opts['layer_list']:
mspec_network = layers.DenseLayer(layers.dropout(mspec_network, p=self.net_opts['dropout_p']),
n,
nonlinearity=lasagne.nonlinearities.rectify)
mspec_network = layers.DenseLayer(layers.dropout(mspec_network, p=self.net_opts['dropout_p']),
self.net_opts['num_class'],
nonlinearity=lasagne.nonlinearities.softmax)
self.network = mspec_network
return self.network
def output_block(net, config, non_lin, verbose=True):
"""
"""
# output setting
out_acts = []
for out_act in config.hyper_parameters.out_act:
exec('from lasagne.nonlinearities import {}'.format(out_act))
out_acts.append(eval(out_act))
n_outs = config.hyper_parameters.n_out
# Global Average Pooling
last_conv_block_name = next(reversed(net))
net['gap'] = L.GlobalPoolLayer(net[last_conv_block_name], name='gap')
net['gap.bn'] = L.BatchNormLayer(net['gap'], name='gap.bn')
n_features = net['gap.bn'].output_shape[-1]
# feature Layer
net['fc'] = L.dropout(L.batch_norm(
L.DenseLayer(net['gap.bn'],
num_units=n_features,
nonlinearity=non_lin,
name='fc')),
name='fc.bn.do')
# output (prediction)
# check whether the model if for MTL or STL
# target is passed as list, regardless whether
# it's MTL or STL (configuration checker checks it)
targets = config.target
out_layer_names = []
for target, n_out, out_act in zip(targets, n_outs, out_acts):
out_layer_names.append('out.{}'.format(target))
if target == 'self':
net[out_layer_names[-1]], inputs = build_siamese(net['fc'])
else:
net[out_layer_names[-1]] = L.DenseLayer(net['fc'],
num_units=n_out,
nonlinearity=out_act,
name=out_layer_names[-1])
inputs = [net['input'].input_var]
# make a concatation layer just for save/load purpose
net['IO'] = L.ConcatLayer([
L.FlattenLayer(net[target_layer_name])
if target == 'self' else net[target_layer_name]
for target_layer_name in out_layer_names
],
name='IO')
if verbose:
print(net['gap.bn'].output_shape)
print(net['fc'].output_shape)
for target in targets:
print(net['out.{}'.format(target)].output_shape)
return net, inputs
def build_triamese_inception(inputlist, imgh=50, imgw=50):
"""
'triamese' (one branch for each view, feeding a fully-connected network),
model using a slightly modified set of Google inception modules
"""
input_var_x, input_var_u, input_var_v = \
inputlist[0], inputlist[1], inputlist[2]
net = {}
# Input layer
tshape = (None, 1, 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)
# nfilters: (pool_proj, 1x1, 3x3_reduce, 3x3, 5x5_reduce, 5x5)
nfilters = [32, 64, 96, 128, 16, 32]
net.update(build_inception_module('inc_x1', net['input_x'], nfilters))
net.update(build_inception_module('inc_u1', net['input_u'], nfilters))
net.update(build_inception_module('inc_v1', net['input_v'], nfilters))
net['dense_x'] = DenseLayer(
dropout(flatten(net['inc_x1/output']), p=.5),
num_units=100, nonlinearity=lasagne.nonlinearities.rectify)
net['dense_u'] = DenseLayer(
dropout(flatten(net['inc_u1/output']), p=.5),
num_units=100, nonlinearity=lasagne.nonlinearities.rectify)
net['dense_v'] = DenseLayer(
dropout(flatten(net['inc_v1/output']), p=.5),
num_units=100, nonlinearity=lasagne.nonlinearities.rectify)
# Concatenate the parallel inputs
net['concat'] = ConcatLayer((net['dense_x'],
net['dense_u'],
net['dense_v']))
# And, finally, the 11-unit output layer with 50% dropout on its inputs:
net['output_prob'] = DenseLayer(
dropout(net['concat'], p=.5),
num_units=11,
nonlinearity=lasagne.nonlinearities.softmax)
logger.info("n-parameters: {}".format(
lasagne.layers.count_params(net['output_prob']))
)
return net['output_prob']
def build_ae(self):
input_layer = InputLayer(shape=(None, self.num_vars * self.num_channels), input_var=self.input_var)
self.hidden = DenseLayer(
dropout(input_layer, p=self.dropout_rate), num_units=self.nodes, nonlinearity=self.activation
)
self.network = DenseLayer(self.hidden, num_units=self.num_vars, W=self.hidden.W.T, nonlinearity=linear)
def build_triamese_inception(inputlist, imgh=50, imgw=50):
"""
'triamese' (one branch for each view, feeding a fully-connected network),
model using a slightly modified set of Google inception modules
"""
input_var_x, input_var_u, input_var_v = \
inputlist[0], inputlist[1], inputlist[2]
net = {}
# Input layer
tshape = (None, 1, 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)
# nfilters: (pool_proj, 1x1, 3x3_reduce, 3x3, 5x5_reduce, 5x5)
nfilters = [32, 64, 96, 128, 16, 32]
net.update(build_inception_module('inc_x1', net['input_x'], nfilters))
net.update(build_inception_module('inc_u1', net['input_u'], nfilters))
net.update(build_inception_module('inc_v1', net['input_v'], nfilters))
net['dense_x'] = DenseLayer(
dropout(flatten(net['inc_x1/output']), p=.5),
num_units=100, nonlinearity=lasagne.nonlinearities.rectify)
net['dense_u'] = DenseLayer(
dropout(flatten(net['inc_u1/output']), p=.5),
num_units=100, nonlinearity=lasagne.nonlinearities.rectify)
net['dense_v'] = DenseLayer(
dropout(flatten(net['inc_v1/output']), p=.5),
num_units=100, nonlinearity=lasagne.nonlinearities.rectify)
# Concatenate the parallel inputs
net['concat'] = ConcatLayer((net['dense_x'],
net['dense_u'],
net['dense_v']))
# And, finally, the 11-unit output layer with 50% dropout on its inputs:
net['output_prob'] = DenseLayer(
dropout(net['concat'], p=.5),
num_units=11,
nonlinearity=lasagne.nonlinearities.softmax)
print("n-parameters: ", lasagne.layers.count_params(net['output_prob']))
return net['output_prob']
def build_mix(input_var, nb_classes, n_chanels=1, input_size=20, reshaped_input_size=20, activity=softmax):
"""
Builds the complete network with 1D-conv1d layer to integrate time from sequences of EEG images.
:param input_vars: list of EEG images (one image per time window)
:param nb_classes: number of classes
:return: a pointer to the output of last layer
"""
# Input layer
input = InputLayer(shape=(None, 1, input_size), input_var=input_var)
input = ReshapeLayer(input, (([0], n_chanels, reshaped_input_size)))
conv1 = Conv1DLayer(input, 1024, 5)
pool1 = MaxPool1DLayer(conv1, 2)
conv2 = Conv1DLayer(pool1, 512, 5)
pool2 = MaxPool1DLayer(conv2, 2)
conv3 = Conv1DLayer(pool2, 256, 2)
conv_layer = FlattenLayer(conv3)
rnn_input = DimshuffleLayer(conv2, (0, 2, 1))
rnnpool = LSTMLayer(rnn_input, num_units=256, nonlinearity=tanh)
rnn_layer = SliceLayer(rnnpool, -1, 1)
network = ConcatLayer([conv_layer, rnn_layer])
# A fully-connected layer of 512 units with 50% dropout on its inputs:
network = DenseLayer(dropout(network, p=.5),
num_units=64, nonlinearity=rectify)
# And, finally, the output layer with 50% dropout on its inputs:
network = DenseLayer(dropout(network, p=.5),
num_units=nb_classes, nonlinearity=activity)
return network
def build_partitioned(input_var, nb_classes, n_chanels, input_size, reshaped_input_size, activity=softmax):
"""
Builds the complete network with 1D-conv1d layer to integrate time from sequences of EEG images.
:param input_vars: list of EEG images (one image per time window)
:param nb_classes: number of classes
:return: a pointer to the output of last layer
"""
# Input layer
input_layer = InputLayer(shape=(None, 1, input_size), input_var=input_var)
input_layer = ReshapeLayer(input_layer, (([0], n_chanels, reshaped_input_size)))
#slice for partition
input_layers = []
for ix in range(n_chanels):
input_layers.append(SliceLayer(input_layer, indices=slice(ix, ix+1), axis=1))
#dnn
networks = []
for input_layer in input_layers:
tmp = DenseLayer(dropout(input_layer, p=.2),
num_units=10, nonlinearity=rectify)
tmp = DenseLayer(dropout(tmp, p=.5),
num_units=3, nonlinearity=rectify)
'''
tmp = Conv1DLayer(input_layer, 8, 5)
tmp = MaxPool1DLayer(tmp, 2)
tmp = Conv1DLayer(tmp, 8, 5)
tmp = MaxPool1DLayer(tmp, 2)
'''
networks.append(tmp)
network = ConcatLayer(networks)
network = DenseLayer(dropout(network, p=.5),
num_units=nb_classes, nonlinearity=activity)
return network
def set_conv_layer(self, network, layer_name, dropout=True, pad=0, bnorm=False):
opts = self.net_opts[layer_name]
ll = layers.Conv1DLayer(
layers.dropout(network, p=self.net_opts['dropout_p']) if dropout else network,
num_filters=opts['num_filters'],
filter_size=opts['filter_size'],
stride=opts['stride'],
pad=pad,
name=layer_name
)
return layers.batch_norm(ll) if bnorm else ll
def set_conv_layer(self, network, layer_name, dropout=True, pad=0, bnorm=False):
opts = self.net_opts[layer_name]
ll = layers.Conv1DLayer(
layers.dropout(network, p=self.net_opts['dropout_p']) if dropout else network,
num_filters=opts['num_filters'],
filter_size=opts['filter_size'],
stride=opts['stride'],
pad=pad,
name=layer_name
)
return layers.batch_norm(ll) if bnorm else ll
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 create_nn(): ''' Returns the theano function - train,test Returns the 'KerasNet' Using default values of adam - learning_rate=0.001, beta1=0.9, beta2=0.999, epsilon=1e-08 Input to the NN is (batch_size,3,32,32) and corresponding classes it belong to (batch_size,) ''' l_in = InputLayer((batch_size,3,32,32)) l_in_bn = BatchNormLayer(l_in) conv1 = Conv2DLayer(l_in_bn,pad='same',num_filters=64,filter_size=(3,3),nonlinearity=lasagne.nonlinearities.rectify) #Bx64x32x32 conv1_1 = Conv2DLayer(conv1,pad='same',num_filters=64,filter_size=(3,3),nonlinearity=lasagne.nonlinearities.rectify) #Bx64x32x32 conv1_mp = MaxPool2DLayer(conv1_1,pool_size=(2,2)) #Bx64x16x16 conv1_do = dropout(conv1_mp,p=0.25) conv2 = Conv2DLayer(conv1_do,pad='same',num_filters=128,filter_size=(3,3),nonlinearity=lasagne.nonlinearities.rectify) #Bx128x16x16 conv2_1 = Conv2DLayer(conv2,pad='same',num_filters=128,filter_size=(3,3),nonlinearity=lasagne.nonlinearities.rectify) #Bx128x16x16 conv2_mp = MaxPool2DLayer(conv2_1,pool_size=(2,2)) #Bx128x8x8 conv2_do = dropout(conv2_mp,p=0.25) flat = flatten(conv2_do,2) #Bx8192 fc = DenseLayer(flat,num_units=512,nonlinearity=lasagne.nonlinearities.rectify) #Bx512 fc_do = dropout(fc, p=0.5) network = DenseLayer(fc_do, num_units=nb_classes, nonlinearity=lasagne.nonlinearities.softmax ) #Bxnb_classes net_output = lasagne.layers.get_output(network) true_output = T.matrix() all_params = lasagne.layers.get_all_params(network,trainable=True) loss = T.mean(lasagne.objectives.categorical_crossentropy(net_output,true_output)) updates = lasagne.updates.adam(loss,all_params) train = theano.function(inputs= [l_in.input_var,true_output] , outputs=[net_output,loss], updates = updates) test = theano.function(inputs= [l_in.input_var], outputs= [net_output]) return train,test,network
def build_discriminator(input_var=None):
#D_inp = T.tensor4('Ds')
D = l.InputLayer(shape=(None, 1, 28, 28), input_var=input_var)
D = l.Conv2DLayer(D,
num_filters=20,
filter_size=(5, 5),
nonlinearity=reLU,
W=lasagne.init.GlorotUniform())
#D = l.Conv2DLayer(D,1,filter_size=(2,2), stride=2, nonlinearity=reLU)
D = l.DropoutLayer(D, p=0.2)
D = l.Conv2DLayer(D,
num_filters=20,
filter_size=(5, 5),
nonlinearity=reLU,
W=lasagne.init.GlorotUniform())
#D = l.Conv2DLayer(D,1,filter_size=(2,2), stride=2, nonlinearity=reLU)
D = l.DropoutLayer(D, p=0.2)
D = l.DenseLayer(l.dropout(D, p=0.5), num_units=256, nonlinearity=reLU)
D = l.DenseLayer(l.dropout(D, p=0.5), num_units=1, nonlinearity=softmax)
#D1.params = D.params
return D
def __build_24_net__(self):
network = layers.InputLayer((None, 3, 24, 24), input_var=self.__input_var__)
network = layers.dropout(network, p=0.1)
network = layers.Conv2DLayer(network,num_filters=64,filter_size=(5,5),stride=1,nonlinearity=relu)
network = layers.batch_norm(network)
network = layers.MaxPool2DLayer(network, pool_size = (3,3),stride = 2)
network = layers.DropoutLayer(network,p=0.5)
network = layers.batch_norm(network)
network = layers.DenseLayer(network,num_units = 64,nonlinearity = relu)
network = layers.DropoutLayer(network,p=0.5)
network = layers.DenseLayer(network,num_units = 2, nonlinearity = softmax)
return network
def __init__(self,
insize,
vocoder,
mlpg_wins=[],
hiddensize=256,
nonlinearity=lasagne.nonlinearities.very_leaky_rectify,
nblayers=3,
bn_axes=None,
dropout_p=-1.0,
grad_clipping=50):
if bn_axes is None:
bn_axes = [] # Recurrent nets don't like batch norm [ref needed]
model.Model.__init__(self, insize, vocoder, hiddensize)
if len(bn_axes) > 0:
warnings.warn(
'ModelBLSTM: You are using bn_axes={}, but batch normalisation is supposed to make Recurrent Neural Networks (RNNS) unstable [ref. needed]'
.format(bn_axes))
l_hid = ll.InputLayer(shape=(None, None, insize),
input_var=self._input_values,
name='input_conditional')
for layi in xrange(nblayers):
layerstr = 'l' + str(1 + layi) + '_BLSTM{}'.format(hiddensize)
fwd = layer_LSTM(l_hid,
hiddensize,
nonlinearity=nonlinearity,
backwards=False,
grad_clipping=grad_clipping,
name=layerstr + '.fwd')
bck = layer_LSTM(l_hid,
hiddensize,
nonlinearity=nonlinearity,
backwards=True,
grad_clipping=grad_clipping,
name=layerstr + '.bck')
l_hid = ll.ConcatLayer((fwd, bck), axis=2)
# Add batch normalisation
if len(bn_axes) > 0: l_hid = ll.batch_norm(l_hid, axes=bn_axes)
# Add dropout (after batchnorm)
if dropout_p > 0.0: l_hid = ll.dropout(l_hid, p=dropout_p)
l_out = layer_final(l_hid, vocoder, mlpg_wins)
self.init_finish(
l_out
) # Has to be called at the end of the __init__ to print out the architecture, get the trainable params, etc.
def build_dnn(input_var, nb_classes, n_chanels=1, input_size=20, reshaped_input_size=20, activity=softmax):
"""
Builds the complete network with 1D-conv1d layer to integrate time from sequences of EEG images.
:param input_vars: list of EEG images (one image per time window)
:param nb_classes: number of classes
:return: a pointer to the output of last layer
"""
# Input layer
network = InputLayer(shape=(None, 1, input_size), input_var=input_var)
#network = ReshapeLayer(network, (([0], n_chanels, reshaped_input_size)))
network = DenseLayer(dropout(network, p=.2),
num_units=160, nonlinearity=rectify)
network = DenseLayer(dropout(network, p=.5),
num_units=60, nonlinearity=rectify)
network = DenseLayer(dropout(network, p=.5),
num_units=nb_classes, nonlinearity=activity)
return network
def lenet5(inputs_shape,
nonlinearity=default_nonlinearity,
use_dropout=False,
**kwargs):
logger.warning("Unrecognized options to the model: %s", kwargs)
l = InputLayer(inputs_shape)
l = conv(l, 32, 5, nonlinearity=None)
l = nonlin(l, nonlinearity=nonlinearity)
l = maxpool(l, 2)
l = conv(l, 64, 5, nonlinearity=None)
l = nonlin(l, nonlinearity=nonlinearity)
l = maxpool(l, 2)
if use_dropout:
l = dropout(l, p=0.5)
l = dense(l, 512, nonlinearity=None)
l = nonlin(l, nonlinearity=nonlinearity)
if use_dropout:
l = dropout(l, p=0.5)
# output layers
logits = dense(l, 10, nonlinearity=None)
return logits
def build_network(self, ra_input_var, mc_input_var):
print('Building raw network with parameters:')
pp = pprint.PrettyPrinter(indent=4)
pp.pprint(self.net_opts)
ra_network_1 = layers.InputLayer((None, 1, None), ra_input_var)
ra_network_1 = self.set_conv_layer(ra_network_1, 'ra_conv_1', dropout=False, pad='same')
ra_network_1 = self.set_pool_layer(ra_network_1, 'ra_pool_1')
ra_network_1 = self.set_conv_layer(ra_network_1, 'ra_conv_2', pad='same')
ra_network_1 = self.set_pool_layer(ra_network_1, 'ra_pool_2')
ra_network_1 = self.set_conv_layer(ra_network_1, 'ra_conv_3', pad='same')
ra_network_1 = self.set_pool_layer(ra_network_1, 'ra_pool_3')
ra_network_1 = self.set_conv_layer(ra_network_1, 'ra_conv_4', pad='same')
ra_network_1 = self.set_pool_layer(ra_network_1, 'ra_pool_4')
concat_list = [ra_network_1]
mc_input = layers.InputLayer((None, 2, None), mc_input_var)
concat_list.append(mc_input)
network = layers.ConcatLayer(concat_list, axis=1, cropping=[None, None, 'center'])
network = self.set_conv_layer(network, 'conv_1')
network = self.set_pool_layer(network, 'pool_1')
network = self.set_conv_layer(network, 'conv_2')
network = self.set_pool_layer(network, 'pool_2')
network = self.set_conv_layer(network, 'conv_3')
network = layers.GlobalPoolLayer(network, getattr(T, self.net_opts['global_pool_func']))
# print(layers.get_output_shape(network))
# network = layers.DenseLayer(layers.dropout(network, p=self.net_opts['dropout_p']),
# self.net_opts['dens_1'],
# nonlinearity=lasagne.nonlinearities.rectify)
network = layers.DenseLayer(layers.dropout(network, p=self.net_opts['dropout_p']),
self.net_opts['dens_2'],
nonlinearity=lasagne.nonlinearities.rectify)
network = layers.DenseLayer(layers.dropout(network, p=self.net_opts['dropout_p']),
self.net_opts['num_class'],
nonlinearity=lasagne.nonlinearities.softmax)
# print(layers.get_output_shape(network))
self.network = network
return self.network
def bid_layer(input_layer, rnn_dim, batch_size, rnn_shape, cell,
add_dense=True, dropout_p=0.2, depth=1, **cell_args):
"""
batch_size: int or symbolic_var (e.g. input_var.shape[0])
context: int
"""
if cell == 'lstm':
cell = LSTMLayer
elif cell == 'gru':
cell = GRULayer
else:
raise ValueError('cell must be one of "lstm", "gru"')
rnn = input_layer
for n in range(depth):
fwd = cell(rnn, rnn_dim, only_return_final=False, **cell_args)
# No need to reverse output of bwd_lstm since backwards is defined:
# backwards : bool
# process the sequence backwards and then reverse the output again
# such that the output from the layer is always from x1x1 to xnxn.
bwd = cell(rnn, rnn_dim, only_return_final=False, backwards=True,
**cell_args)
if add_dense:
# reshape for dense
fwd = ReshapeLayer(fwd, (-1, rnn_dim))
bwd = ReshapeLayer(bwd, (-1, rnn_dim))
fwd = DenseLayer(fwd, num_units=rnn_dim, nonlinearity=tanh)
bwd = DenseLayer(bwd, num_units=rnn_dim, nonlinearity=tanh)
# dropout
fwd = dropout(fwd, p=dropout_p)
bwd = dropout(bwd, p=dropout_p)
# reshape back to input format
fwd = ReshapeLayer(fwd, rnn_shape)
bwd = ReshapeLayer(bwd, rnn_shape)
# merge over lstm output dim (axis=2)
rnn = ElemwiseSumLayer(incomings=[fwd, bwd])
return rnn
def build_model_audio(modelfile, meanstd_file, input_dim, excerpt_size):
"""
Builds the CNN architecture defined by Jan et. al. @ISMIR2015 and later loads the saved model
and the mean std file.
"""
# Build CNN architecture
net = {}
net['input'] = InputLayer((None, 1, excerpt_size, input_dim))
kwargs = dict(nonlinearity=lasagne.nonlinearities.leaky_rectify,
W=lasagne.init.Orthogonal())
net['Conv1_1'] = ConvLayer(net['input'], 64, 3, **kwargs)
net['Conv1_2'] = ConvLayer(net['Conv1_1'], 32, 3, **kwargs)
net['pool1'] = MaxPool2DLayer(net['Conv1_2'], 3)
net['Conv2_1'] = ConvLayer(net['pool1'], 128, 3, **kwargs)
net['Conv2_2'] = ConvLayer(net['Conv2_1'], 64, 3, **kwargs)
net['pool2'] = MaxPool2DLayer(net['Conv2_2'], 3)
net['fc3'] = DenseLayer(dropout(net['pool2'], 0.5), 256, **kwargs)
net['fc4'] = DenseLayer(dropout(net['fc3'], 0.5), 64, **kwargs)
net['score'] = DenseLayer(dropout(net['fc4'], 0.5),
1,
nonlinearity=lasagne.nonlinearities.sigmoid,
W=lasagne.init.Orthogonal())
# load saved weights
with np.load(modelfile) as f:
lasagne.layers.set_all_param_values(
net['score'], [f['param%d' % i] for i in range(len(f.files))])
# - load mean/std
with np.load(meanstd_file) as f:
mean = f['mean']
std = f['std']
mean = mean.astype(floatX)
istd = np.reciprocal(std).astype(floatX)
return net, mean, istd
def mlp(inputs_shape,
layer_dims,
nonlinearity=default_nonlinearity,
use_dropout=False,
**kwargs):
logger.warning("Unrecognized options to the model: %s", kwargs)
l = lasagne.layers.InputLayer(inputs_shape)
W_init = lasagne.init.GlorotUniform()
for i, layer_size in enumerate(layer_dims[:-1]):
assert layer_size >= 0
l = dense(l, layer_size, W=W_init, nonlinearity=None)
l = nonlin(l, nonlinearity=nonlinearity)
if use_dropout:
l = dropout(l)
logits = dense(l, layer_dims[-1], W=W_init, nonlinearity=None)
return logits
def __build_24_net__(self):
network = layers.InputLayer((None, 3, 24, 24),
input_var=self.__input_var__)
network = layers.dropout(network, p=0.1)
network = layers.Conv2DLayer(network,
num_filters=64,
filter_size=(5, 5),
stride=1,
nonlinearity=relu)
network = layers.batch_norm(network)
network = layers.MaxPool2DLayer(network, pool_size=(3, 3), stride=2)
network = layers.DropoutLayer(network, p=0.5)
network = layers.batch_norm(network)
network = layers.DenseLayer(network, num_units=64, nonlinearity=relu)
network = layers.DropoutLayer(network, p=0.5)
network = layers.DenseLayer(network, num_units=2, nonlinearity=softmax)
return network
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_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 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
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 __init__(self, train_list_raw, test_list_raw, png_folder, batch_size, dropout, l2, mode, batch_norm, **kwargs):
print "==> not used params in DMN class:", kwargs.keys()
self.train_list_raw = train_list_raw
self.test_list_raw = test_list_raw
self.png_folder = png_folder
self.batch_size = batch_size
self.dropout = dropout
self.l2 = l2
self.mode = mode
self.batch_norm = batch_norm
self.input_var = T.tensor4('input_var')
self.answer_var = T.ivector('answer_var')
print "==> building network"
example = np.random.uniform(size=(self.batch_size, 1, 256, 858), low=0.0, high=1.0).astype(np.float32) #########
answer = np.random.randint(low=0, high=176, size=(self.batch_size,)) #########
network = layers.InputLayer(shape=(None, 1, 256, 858), input_var=self.input_var)
print layers.get_output(network).eval({self.input_var:example}).shape
# CONV-RELU-POOL 1
network = layers.Conv2DLayer(incoming=network, num_filters=16, filter_size=(7, 7),
stride=1, nonlinearity=rectify)
print layers.get_output(network).eval({self.input_var:example}).shape
network = layers.MaxPool2DLayer(incoming=network, pool_size=(3, 3), stride=2, ignore_border=False)
print layers.get_output(network).eval({self.input_var:example}).shape
if (self.batch_norm):
network = layers.BatchNormLayer(incoming=network)
# CONV-RELU-POOL 2
network = layers.Conv2DLayer(incoming=network, num_filters=32, filter_size=(5, 5),
stride=1, nonlinearity=rectify)
print layers.get_output(network).eval({self.input_var:example}).shape
network = layers.MaxPool2DLayer(incoming=network, pool_size=(3, 3), stride=2, ignore_border=False)
print layers.get_output(network).eval({self.input_var:example}).shape
if (self.batch_norm):
network = layers.BatchNormLayer(incoming=network)
# CONV-RELU-POOL 3
network = layers.Conv2DLayer(incoming=network, num_filters=64, filter_size=(3, 3),
stride=1, nonlinearity=rectify)
print layers.get_output(network).eval({self.input_var:example}).shape
network = layers.MaxPool2DLayer(incoming=network, pool_size=(3, 3), stride=2, ignore_border=False)
print layers.get_output(network).eval({self.input_var:example}).shape
if (self.batch_norm):
network = layers.BatchNormLayer(incoming=network)
# CONV-RELU-POOL 4
network = layers.Conv2DLayer(incoming=network, num_filters=128, filter_size=(3, 3),
stride=1, nonlinearity=rectify)
print layers.get_output(network).eval({self.input_var:example}).shape
network = layers.MaxPool2DLayer(incoming=network, pool_size=(3, 3), stride=2, ignore_border=False)
print layers.get_output(network).eval({self.input_var:example}).shape
if (self.batch_norm):
network = layers.BatchNormLayer(incoming=network)
# CONV-RELU-POOL 5
network = layers.Conv2DLayer(incoming=network, num_filters=128, filter_size=(3, 3),
stride=1, nonlinearity=rectify)
print layers.get_output(network).eval({self.input_var:example}).shape
network = layers.MaxPool2DLayer(incoming=network, pool_size=(3, 3), stride=2, ignore_border=False)
print layers.get_output(network).eval({self.input_var:example}).shape
if (self.batch_norm):
network = layers.BatchNormLayer(incoming=network)
# CONV-RELU-POOL 6
network = layers.Conv2DLayer(incoming=network, num_filters=256, filter_size=(3, 3),
stride=1, nonlinearity=rectify)
print layers.get_output(network).eval({self.input_var:example}).shape
network = layers.MaxPool2DLayer(incoming=network, pool_size=(3, 3), stride=(3, 2), ignore_border=False)
print layers.get_output(network).eval({self.input_var:example}).shape
if (self.batch_norm):
network = layers.BatchNormLayer(incoming=network)
# DENSE 1
network = layers.DenseLayer(incoming=network, num_units=1024, nonlinearity=rectify)
if (self.batch_norm):
network = layers.BatchNormLayer(incoming=network)
if (self.dropout > 0):
network = layers.dropout(network, self.dropout)
print layers.get_output(network).eval({self.input_var:example}).shape
"""
# DENSE 2
network = layers.DenseLayer(incoming=network, num_units=1024, nonlinearity=rectify)
if (self.batch_norm):
network = layers.BatchNormLayer(incoming=network)
if (self.dropout > 0):
network = layers.dropout(network, self.dropout)
print layers.get_output(network).eval({self.input_var:example}).shape
"""
# Last layer: classification
network = layers.DenseLayer(incoming=network, num_units=176, nonlinearity=softmax)
print layers.get_output(network).eval({self.input_var:example}).shape
self.params = layers.get_all_params(network, trainable=True)
self.prediction = layers.get_output(network)
self.loss_ce = lasagne.objectives.categorical_crossentropy(self.prediction, self.answer_var).mean()
if (self.l2 > 0):
self.loss_l2 = self.l2 * lasagne.regularization.regularize_network_params(network,
lasagne.regularization.l2)
else:
self.loss_l2 = 0
self.loss = self.loss_ce + self.loss_l2
#updates = lasagne.updates.adadelta(self.loss, self.params)
updates = lasagne.updates.momentum(self.loss, self.params, learning_rate=0.003)
if self.mode == 'train':
print "==> compiling train_fn"
self.train_fn = theano.function(inputs=[self.input_var, self.answer_var],
outputs=[self.prediction, self.loss],
updates=updates)
print "==> compiling test_fn"
self.test_fn = theano.function(inputs=[self.input_var, self.answer_var],
outputs=[self.prediction, self.loss])
def __init__(
self,
train_list_raw,
test_list_raw,
png_folder,
batch_size,
dropout,
l2,
mode,
batch_norm,
rnn_num_units,
**kwargs
):
print "==> not used params in DMN class:", kwargs.keys()
self.train_list_raw = train_list_raw
self.test_list_raw = test_list_raw
self.png_folder = png_folder
self.batch_size = batch_size
self.dropout = dropout
self.l2 = l2
self.mode = mode
self.batch_norm = batch_norm
self.num_units = rnn_num_units
self.input_var = T.tensor4("input_var")
self.answer_var = T.ivector("answer_var")
print "==> building network"
example = np.random.uniform(size=(self.batch_size, 1, 128, 768), low=0.0, high=1.0).astype(
np.float32
) #########
answer = np.random.randint(low=0, high=176, size=(self.batch_size,)) #########
network = layers.InputLayer(shape=(None, 1, 128, 768), input_var=self.input_var)
print layers.get_output(network).eval({self.input_var: example}).shape
# CONV-RELU-POOL 1
network = layers.Conv2DLayer(
incoming=network, num_filters=16, filter_size=(7, 7), stride=1, nonlinearity=rectify
)
print layers.get_output(network).eval({self.input_var: example}).shape
network = layers.MaxPool2DLayer(incoming=network, pool_size=(3, 3), stride=2, pad=2)
print layers.get_output(network).eval({self.input_var: example}).shape
if self.batch_norm:
network = layers.BatchNormLayer(incoming=network)
# CONV-RELU-POOL 2
network = layers.Conv2DLayer(
incoming=network, num_filters=32, filter_size=(5, 5), stride=1, nonlinearity=rectify
)
print layers.get_output(network).eval({self.input_var: example}).shape
network = layers.MaxPool2DLayer(incoming=network, pool_size=(3, 3), stride=2, pad=2)
print layers.get_output(network).eval({self.input_var: example}).shape
if self.batch_norm:
network = layers.BatchNormLayer(incoming=network)
# CONV-RELU-POOL 3
network = layers.Conv2DLayer(
incoming=network, num_filters=32, filter_size=(3, 3), stride=1, nonlinearity=rectify
)
print layers.get_output(network).eval({self.input_var: example}).shape
network = layers.MaxPool2DLayer(incoming=network, pool_size=(3, 3), stride=2, pad=2)
print layers.get_output(network).eval({self.input_var: example}).shape
if self.batch_norm:
network = layers.BatchNormLayer(incoming=network)
# CONV-RELU-POOL 4
network = layers.Conv2DLayer(
incoming=network, num_filters=32, filter_size=(3, 3), stride=1, nonlinearity=rectify
)
print layers.get_output(network).eval({self.input_var: example}).shape
network = layers.MaxPool2DLayer(incoming=network, pool_size=(3, 3), stride=2, pad=2)
print layers.get_output(network).eval({self.input_var: example}).shape
if self.batch_norm:
network = layers.BatchNormLayer(incoming=network)
self.params = layers.get_all_params(network, trainable=True)
output = layers.get_output(network)
num_channels = 32
filter_W = 48
filter_H = 8
# NOTE: these constants are shapes of last pool layer, it can be symbolic
# explicit values are better for optimizations
channels = []
for channel_index in range(num_channels):
channels.append(output[:, channel_index, :, :].transpose((0, 2, 1)))
rnn_network_outputs = []
W_in_to_updategate = None
W_hid_to_updategate = None
b_updategate = None
W_in_to_resetgate = None
W_hid_to_resetgate = None
b_resetgate = None
W_in_to_hidden_update = None
W_hid_to_hidden_update = None
b_hidden_update = None
for channel_index in range(num_channels):
rnn_input_var = channels[channel_index]
# InputLayer
network = layers.InputLayer(shape=(None, filter_W, filter_H), input_var=rnn_input_var)
if channel_index == 0:
# GRULayer
network = layers.GRULayer(incoming=network, num_units=self.num_units, only_return_final=True)
W_in_to_updategate = network.W_in_to_updategate
W_hid_to_updategate = network.W_hid_to_updategate
b_updategate = network.b_updategate
W_in_to_resetgate = network.W_in_to_resetgate
W_hid_to_resetgate = network.W_hid_to_resetgate
b_resetgate = network.b_resetgate
W_in_to_hidden_update = network.W_in_to_hidden_update
W_hid_to_hidden_update = network.W_hid_to_hidden_update
b_hidden_update = network.b_hidden_update
# add params
self.params += layers.get_all_params(network, trainable=True)
else:
# GRULayer, but shared
network = layers.GRULayer(
incoming=network,
num_units=self.num_units,
only_return_final=True,
resetgate=layers.Gate(W_in=W_in_to_resetgate, W_hid=W_hid_to_resetgate, b=b_resetgate),
updategate=layers.Gate(W_in=W_in_to_updategate, W_hid=W_hid_to_updategate, b=b_updategate),
hidden_update=layers.Gate(
W_in=W_in_to_hidden_update, W_hid=W_hid_to_hidden_update, b=b_hidden_update
),
)
rnn_network_outputs.append(layers.get_output(network))
all_output_var = T.concatenate(rnn_network_outputs, axis=1)
print all_output_var.eval({self.input_var: example}).shape
# InputLayer
network = layers.InputLayer(shape=(None, self.num_units * num_channels), input_var=all_output_var)
# Dropout Layer
if self.dropout > 0:
network = layers.dropout(network, self.dropout)
# BatchNormalization Layer
if self.batch_norm:
network = layers.BatchNormLayer(incoming=network)
# Last layer: classification
network = layers.DenseLayer(incoming=network, num_units=176, nonlinearity=softmax)
print layers.get_output(network).eval({self.input_var: example}).shape
self.params += layers.get_all_params(network, trainable=True)
self.prediction = layers.get_output(network)
# print "==> param shapes", [x.eval().shape for x in self.params]
self.loss_ce = lasagne.objectives.categorical_crossentropy(self.prediction, self.answer_var).mean()
if self.l2 > 0:
self.loss_l2 = self.l2 * lasagne.regularization.apply_penalty(self.params, lasagne.regularization.l2)
else:
self.loss_l2 = 0
self.loss = self.loss_ce + self.loss_l2
# updates = lasagne.updates.adadelta(self.loss, self.params)
updates = lasagne.updates.momentum(self.loss, self.params, learning_rate=0.003)
if self.mode == "train":
print "==> compiling train_fn"
self.train_fn = theano.function(
inputs=[self.input_var, self.answer_var], outputs=[self.prediction, self.loss], updates=updates
)
print "==> compiling test_fn"
self.test_fn = theano.function(inputs=[self.input_var, self.answer_var], outputs=[self.prediction, self.loss])
def buildCNNFingerprint(input_atom, input_bonds, input_atom_index, input_bond_index, input_mask, \
max_atom_len, max_bond_len, input_atom_dim, input_bond_dim, input_index_dim, fingerprint_dim,
batch_size, output_dim, final_layer_type, fingerprint_network_architecture=[],\
neural_net=[]):
dropout_prob = 0.0
network_vals = {}
#take in input layers for the atom and the bonds
l_in_atom = InputLayer(shape=(None,max_atom_len,input_atom_dim), \
input_var=input_atom)
l_in_bond = InputLayer(shape=(None,max_bond_len,input_bond_dim), \
input_var=input_bonds)
#take in layers for the indexing into the atoms and bonds
l_index_atom = InputLayer(shape=(None,max_atom_len,input_index_dim), \
input_var=input_atom_index)
l_index_bond = InputLayer(shape=(None,max_atom_len,input_index_dim), \
input_var=input_bond_index)
#take in input for the mask
l_mask = InputLayer(shape=(None,max_atom_len), input_var=input_mask)
#get the number of hidden units for the first layer
first_hidden_units_num = fingerprint_network_architecture[0]
#get the embedding of the sequences we are encoding
network_vals['hiddens_for_atoms'] = FingerprintHiddensLayer([l_index_atom,l_index_bond,\
l_in_atom,l_in_bond,l_in_atom,l_mask], input_atom_dim, input_atom_dim,
input_bond_dim,first_hidden_units_num,max_atom_len)
#sparsify
network_vals['sparse_for_atoms'] = SparsifyFingerprintLayer([network_vals['hiddens_for_atoms']],\
first_hidden_units_num, fingerprint_dim)
network_vals['fingerprints'] = FingerprintGen([network_vals['sparse_for_atoms'],l_mask])
for i,curr_num_hiddens in enumerate(fingerprint_network_architecture[1:]):
prev_hidden_units = fingerprint_network_architecture[i]
network_vals['hiddens_for_atoms'] = FingerprintHiddensLayer([l_index_atom,l_index_bond,\
l_in_atom,l_in_bond,network_vals['hiddens_for_atoms'],l_mask], prev_hidden_units,
input_atom_dim, input_bond_dim, curr_num_hiddens, max_atom_len)
network_vals['sparse_for_atoms'] = SparsifyFingerprintLayer([network_vals['hiddens_for_atoms']],\
curr_num_hiddens, fingerprint_dim)
newFingerprints = FingerprintGen([network_vals['sparse_for_atoms'],l_mask])
network_vals['fingerprints'] = FingerprintMerge([network_vals['fingerprints'],newFingerprints])
#then run through the neural net on top of the fingerprint
if neural_net != []:
#do dropout
network_vals['drop0'] = lasagne.layers.dropout(network_vals['fingerprints'],p=dropout_prob)
#run through the layers I have
for layerNum,hiddenUnits in enumerate(neural_net):
oldLayerNum = layerNum
currLayerNum = layerNum + 1
network_vals['dense'+str(currLayerNum)] = DenseLayer(network_vals['drop'+str(oldLayerNum)], \
hiddenUnits,nonlinearity=lasagne.nonlinearities.rectify)
network_vals['drop'+str(currLayerNum)] = dropout(network_vals['dense'+str(currLayerNum)],p=dropout_prob)
network_vals['final_out'] = network_vals['drop'+str(currLayerNum)]
else:
network_vals['final_out'] = network_vals['fingerprints']
#finally, project it into the dimensionality we want
network_vals['output'] = DenseLayer(network_vals['final_out'],num_units=output_dim, nonlinearity=final_layer_type)
return network_vals
def build_critic(input_var=None, cond_var=None, n_conds=0, arch=0,
with_BatchNorm=True, loss_type='wgan'):
from lasagne.layers import (
InputLayer, Conv2DLayer, DenseLayer, MaxPool2DLayer, concat,
dropout, flatten)
from lasagne.nonlinearities import rectify, LeakyRectify
from lasagne.init import GlorotUniform # Normal
lrelu = LeakyRectify(0.2)
layer = InputLayer(
shape=(None, 1, 128, 128), input_var=input_var, name='d_in_data')
# init = Normal(0.02, 0.0)
init = GlorotUniform()
if cond_var:
# class: from data or from generator input
layer_cond = InputLayer(
shape=(None, n_conds), input_var=cond_var, name='d_in_condition')
layer_cond = BatchNorm(DenseLayer(
layer_cond, 1024, W=init, b=None, nonlinearity=lrelu),
with_BatchNorm)
if arch == 'dcgan':
# DCGAN inspired
layer = BatchNorm(Conv2DLayer(
layer, 32, 4, stride=2, pad=1, W=init, b=None, nonlinearity=lrelu),
with_BatchNorm)
layer = BatchNorm(Conv2DLayer(
layer, 64, 4, stride=2, pad=1, W=init, b=None, nonlinearity=lrelu),
with_BatchNorm)
layer = BatchNorm(Conv2DLayer(
layer, 128, 4, stride=2, pad=1, W=init, b=None, nonlinearity=lrelu),
with_BatchNorm)
layer = BatchNorm(Conv2DLayer(
layer, 256, 4, stride=2, pad=1, W=init, b=None, nonlinearity=lrelu),
with_BatchNorm)
layer = BatchNorm(Conv2DLayer(
layer, 512, 4, stride=2, pad=1, W=init, b=None, nonlinearity=lrelu),
with_BatchNorm)
elif arch == 'cont-enc':
# convolution layers
layer = BatchNorm(Conv2DLayer(
layer, 64, 4, stride=2, pad=1, W=init, nonlinearity=lrelu),
with_BatchNorm)
layer = BatchNorm(Conv2DLayer(
layer, 64, 4, stride=2, pad=1, W=init, nonlinearity=lrelu),
with_BatchNorm)
layer = BatchNorm(Conv2DLayer(
layer, 128, 4, stride=2, pad=1, W=init, nonlinearity=lrelu),
with_BatchNorm)
layer = BatchNorm(Conv2DLayer(
layer, 256, 4, stride=2, pad=1, W=init, nonlinearity=lrelu),
with_BatchNorm)
layer = BatchNorm(Conv2DLayer(
layer, 512, 4, stride=2, pad=1, W=init, nonlinearity=lrelu),
with_BatchNorm)
elif arch == 'mnist':
# Jan Schluechter's MNIST discriminator
# convolution layers
layer = BatchNorm(Conv2DLayer(
layer, 128, 5, stride=2, pad='same', W=init, b=None,
nonlinearity=lrelu), with_BatchNorm)
layer = BatchNorm(Conv2DLayer(
layer, 128, 5, stride=2, pad='same', W=init, b=None,
nonlinearity=lrelu), with_BatchNorm)
layer = BatchNorm(Conv2DLayer(
layer, 128, 5, stride=2, pad='same', W=init, b=None,
nonlinearity=lrelu), with_BatchNorm)
# layer = BatchNorm(Conv2DLayer(
# layer, 128, 5, stride=2, pad='same', W=init, b=None,
# nonlinearity=lrelu), with_BatchNorm)
# fully-connected layer
# layer = BatchNorm(DenseLayer(
# layer, 1024, W=init, b=None, nonlinearity=lrelu), with_BatchNorm)
elif arch == 'lsgan':
layer = batch_norm(Conv2DLayer(
layer, 256, 5, stride=2, pad='same', nonlinearity=lrelu))
layer = batch_norm(Conv2DLayer(
layer, 256, 5, stride=2, pad='same', nonlinearity=lrelu))
layer = batch_norm(Conv2DLayer(
layer, 256, 5, stride=2, pad='same', nonlinearity=lrelu))
elif arch == 'crepe':
# CREPE
# form words from sequence of characters
layer = BatchNorm(Conv2DLayer(
layer, 1024, (128, 7), W=init, b=None, nonlinearity=lrelu),
with_BatchNorm)
layer = MaxPool2DLayer(layer, (1, 3))
# temporal convolution, 7-gram
layer = BatchNorm(Conv2DLayer(
layer, 512, (1, 7), W=init, b=None, nonlinearity=lrelu),
with_BatchNorm)
layer = MaxPool2DLayer(layer, (1, 3))
# temporal convolution, 3-gram
layer = BatchNorm(Conv2DLayer(
layer, 256, (1, 3), W=init, b=None, nonlinearity=lrelu),
with_BatchNorm)
layer = BatchNorm(Conv2DLayer(
layer, 256, (1, 3), W=init, b=None, nonlinearity=lrelu),
with_BatchNorm)
layer = BatchNorm(Conv2DLayer(
layer, 256, (1, 3), W=init, b=None, nonlinearity=lrelu),
with_BatchNorm)
layer = BatchNorm(Conv2DLayer(
layer, 256, (1, 3), W=init, b=None, nonlinearity=lrelu),
with_BatchNorm)
layer = flatten(layer)
# fully-connected layers
layer = dropout(DenseLayer(
layer, 1024, W=init, b=None, nonlinearity=rectify))
layer = dropout(DenseLayer(
layer, 1024, W=init, b=None, nonlinearity=rectify))
else:
raise Exception("Model architecture {} is not supported".format(arch))
# output layer (linear and without bias)
if cond_var is not None:
layer = DenseLayer(layer, 1024, nonlinearity=lrelu, b=None)
layer = concat([layer, layer_cond])
layer = DenseLayer(layer, 1, b=None, nonlinearity=None)
print("Critic output:", layer.output_shape)
return layer
input_var = T.tensor4('inputs')
target_var = T.ivector('targets')
domain_target_var = T.ivector('domain_targets')
num_classes = 2
net = InputLayer(shape=(None, 3, 32, 32), input_var=input_var)
halfway_net = DenseLayer(net, num_units=1024)
net = DenseLayer(halfway_net,
num_units=num_classes,
nonlinearity=lasagne.nonlinearities.softmax)
# gradient reversal branch
gr_branch = DenseLayer(halfway_net, num_units=512)
gr_branch = DenseLayer(dropout(gr_branch),
num_units=2,
nonlinearity=lasagne.nonlinearities.softmax)
###################################
# Define and compile Theano
# I.e., you've got one output layer for the source task classification, and another output layer for the domain classification, and both share the same input layer (and a part of the network).
# You'd then define two ordinary loss functions:
pred_sourcetask, pred_domainclass = lasagne.layers.get_output(
[net, gr_branch])
loss_sourcetask = lasagne.objectives.categorical_crossentropy(
pred_sourcetask, target_var).mean()
loss_domainclass = lasagne.objectives.categorical_crossentropy(
pred_domainclass, domain_target_var).mean()
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"]