def net_dvc(image_size=(32,32)):
convos = [5,5,5]
pools = [2,2,2]
filters = [100,200,300]
tuplify = lambda x: (x,x)
convos = list(map(tuplify, convos))
conv_layers = [Convolutional(filter_size=s,num_filters=o, num_channels=i, name="Conv"+str(n))\
for s,o,i,n in zip(convos, filters, [3] + filters, range(1000))]
pool_layers = [MaxPooling(p) for p in map(tuplify, pools)]
activations = [Rectifier() for i in convos]
layers = [i for l in zip(conv_layers, activations, pool_layers) for i in l]
cnn = ConvolutionalSequence(layers, 3, image_size=image_size, name="cnn",
weights_init=Uniform(width=.1),
biases_init=Constant(0))
cnn._push_allocation_config()
cnn_output = np.prod(cnn.get_dim('output'))
mlp_size = [cnn_output,500,2]
mlp = MLP([Rectifier(), Softmax()], mlp_size, name="mlp",
weights_init=Uniform(width=.1),
biases_init=Constant(0))
seq = FeedforwardSequence([net.apply for net in [cnn,Flattener(),mlp]])
seq.push_initialization_config()
seq.initialize()
return seq
def _allocate(self):
arghs = dict(shape=self.shape, broadcastable=self.broadcastable)
sequence = []
if self.batch_normalize:
sequence.append(Standardization(**arghs))
sequence.append(SharedScale(weights_init=Constant(1), **arghs))
sequence.append(SharedShift(biases_init=Constant(0), **arghs))
sequence.append(self.activation)
self.sequence = FeedforwardSequence(
[brick.apply for brick in sequence], name="ffs")
self.children = [self.sequence]
def build_model(images, labels):
# Construct a bottom convolutional sequence
bottom_conv_sequence = convolutional_sequence((3,3), 16, (160, 160))
bottom_conv_sequence._push_allocation_config()
# Flatten layer
flattener = Flattener()
# Construct a top MLP
conv_out_dim = numpy.prod(bottom_conv_sequence.get_dim('output'))
#top_mlp = MLP([Rectifier(name='non_linear_9'), Softmax(name='non_linear_11')], [conv_out_dim, 1024, 10], weights_init=IsotropicGaussian(), biases_init=Constant(0))
top_mlp = BatchNormalizedMLP([Rectifier(name='non_linear_9'), Softmax(name='non_linear_11')], [conv_out_dim, 1024, 10], weights_init=IsotropicGaussian(), biases_init=Constant(0))
# Construct feedforward sequence
ss_seq = FeedforwardSequence([bottom_conv_sequence.apply, flattener.apply, top_mlp.apply])
ss_seq.push_initialization_config()
ss_seq.initialize()
prediction = ss_seq.apply(images)
cost_noreg = CategoricalCrossEntropy().apply(labels.flatten(), prediction)
# add regularization
selector = Selector([top_mlp])
Ws = selector.get_parameters('W')
mlp_brick_name = 'batchnormalizedmlp'
W0 = Ws['/%s/linear_0.W' % mlp_brick_name]
W1 = Ws['/%s/linear_1.W' % mlp_brick_name]
cost = cost_noreg + .01 * (W0 ** 2).mean() + .01 * (W1 ** 2).mean()
return cost
def build_model(images, labels):
vgg = VGG(layer='conv4_4')
vgg.push_initialization_config()
vgg.initialize()
tdb = top_direction_block()
tdb.push_initialization_config()
tdb.initialize()
# Construct feedforward sequence
ss_seq = FeedforwardSequence([vgg.apply, tdb.apply])
ss_seq.push_initialization_config()
ss_seq.initialize()
prediction = ss_seq.apply(images)
cost = StructuredCost().apply(labels, theano.tensor.clip(prediction, 1e-5, 1 - 1e-5))
cg = ComputationGraph(cost)
cg_dropout = apply_dropout(cg, [VariableFilter(roles=[OUTPUT])(cg.variables)[0]], .5)
cost_dropout = cg_dropout.outputs[0]
# define learned parameters
selector = Selector([ss_seq])
W = selector.get_parameters()
parameters = []
parameters += [v for k, v in W.items()]
return cost_dropout, parameters
def build_model(images, labels):
# Construct a bottom convolutional sequence
bottom_conv_sequence = convolutional_sequence((3, 3), 64, (150, 150))
bottom_conv_sequence._push_allocation_config()
# Flatten layer
flattener = Flattener()
# Construct a top MLP
conv_out_dim = numpy.prod(bottom_conv_sequence.get_dim('output'))
top_mlp = MLP([
LeakyRectifier(name='non_linear_9'),
LeakyRectifier(name='non_linear_10'),
Softmax(name='non_linear_11')
], [conv_out_dim, 2048, 612, 10],
weights_init=IsotropicGaussian(),
biases_init=Constant(1))
# Construct feedforward sequence
ss_seq = FeedforwardSequence(
[bottom_conv_sequence.apply, flattener.apply, top_mlp.apply])
ss_seq.push_initialization_config()
ss_seq.initialize()
prediction = ss_seq.apply(images)
cost = CategoricalCrossEntropy().apply(labels.flatten(), prediction)
return cost
def construct_model(task, patch_shape, initargs, n_channels, n_spatial_dims, hidden_dim,
batch_normalize,
hyperparameters, patch_cnn_spec=None, patch_mlp_spec=None,
prefork_area_mlp_spec=[], postmerge_area_mlp_spec=[], response_mlp_spec=[],
**kwargs):
patch_transforms = []
if patch_cnn_spec:
patch_transforms.append(masonry.construct_cnn(
name="patch_cnn",
layer_specs=patch_cnn_spec,
input_shape=patch_shape,
n_channels=n_channels,
batch_normalize=batch_normalize).apply)
shape = patch_transforms[-1].brick.get_dim("output")
else:
shape = (n_channels,) + tuple(patch_shape)
patch_transforms.append(masonry.FeedforwardFlattener(input_shape=shape).apply)
if patch_mlp_spec:
patch_transforms.append(masonry.construct_mlp(
name="patch_mlp",
hidden_dims=patch_mlp_spec,
input_dim=patch_transforms[-1].brick.output_dim,
batch_normalize=batch_normalize,
initargs=initargs).apply)
patch_transform = FeedforwardSequence(patch_transforms, name="ffs")
prefork_area_transform = masonry.construct_mlp(
name="prefork_area_mlp",
input_dim=hidden_dim,
hidden_dims=prefork_area_mlp_spec,
batch_normalize=batch_normalize,
initargs=initargs)
postmerge_area_transform = masonry.construct_mlp(
name="postmerge_area_mlp",
input_dim=2*n_spatial_dims,
hidden_dims=postmerge_area_mlp_spec,
batch_normalize=batch_normalize,
initargs=initargs)
# LSTM requires the input to have dim=4*hidden_dim
response_mlp_spec.append(4*hidden_dim)
response_transform = masonry.construct_mlp(
name="response_mlp",
hidden_dims=response_mlp_spec[1:],
input_dim=response_mlp_spec[0],
batch_normalize=batch_normalize,
initargs=initargs)
emitter = task.get_emitter(**hyperparameters)
return Ram(patch_transform=patch_transform.apply,
prefork_area_transform=prefork_area_transform.apply,
postmerge_area_transform=postmerge_area_transform.apply,
response_transform=response_transform.apply,
emitter=emitter,
**hyperparameters)
def __init__(self, visual_dim, textual_dim, output_dim, hidden_size,
init_ranges, **kwargs):
(visual_range, textual_range, linear_range_1, linear_range_2,
linear_range_3) = init_ranges
visual_layer = FeedforwardSequence([
BatchNormalization(input_dim=visual_dim).apply,
LinearMaxout(
input_dim=visual_dim,
output_dim=hidden_size,
weights_init=initialization.Uniform(width=visual_range),
use_bias=False,
biases_init=initialization.Constant(0),
num_pieces=2).apply
],
name='visual_layer')
textual_layer = FeedforwardSequence([
BatchNormalization(input_dim=textual_dim).apply,
LinearMaxout(
input_dim=textual_dim,
output_dim=hidden_size,
weights_init=initialization.Uniform(width=textual_range),
biases_init=initialization.Constant(0),
use_bias=False,
num_pieces=2).apply
],
name='textual_layer')
logistic_mlp = MLPGenreClassifier(
hidden_size, output_dim, hidden_size,
[linear_range_1, linear_range_2, linear_range_3])
# logistic_mlp = Sequence([
# BatchNormalization(input_dim=hidden_size, name='bn1').apply,
# Linear(hidden_size, output_dim, name='linear_output', use_bias=False,
# weights_init=initialization.Uniform(width=linear_range_1)).apply,
# Logistic().apply
#], name='logistic_mlp')
children = [visual_layer, textual_layer, logistic_mlp]
kwargs.setdefault('use_bias', False)
kwargs.setdefault('children', children)
super(LinearSumClassifier, self).__init__(**kwargs)
def _allocate(self):
arghs = dict(shape=self.shape,
broadcastable=self.broadcastable)
sequence = []
if self.batch_normalize:
sequence.append(Standardization(**arghs))
sequence.append(SharedScale(
weights_init=Constant(1),
**arghs))
sequence.append(SharedShift(
biases_init=Constant(0),
**arghs))
sequence.append(self.activation)
self.sequence = FeedforwardSequence([
brick.apply for brick in sequence
], name="ffs")
self.children = [self.sequence]
class NormalizedActivation(Initializable, Feedforward):
@lazy(allocation="shape broadcastable".split())
def __init__(self, shape, broadcastable, activation=None, batch_normalize=False, **kwargs):
super(NormalizedActivation, self).__init__(**kwargs)
self.shape = shape
self.broadcastable = broadcastable
self.activation = activation or Rectifier()
self.batch_normalize = batch_normalize
@property
def broadcastable(self):
return self._broadcastable or [False]*len(self.shape)
@broadcastable.setter
def broadcastable(self, broadcastable):
self._broadcastable = broadcastable
def _allocate(self):
arghs = dict(shape=self.shape,
broadcastable=self.broadcastable)
sequence = []
if self.batch_normalize:
sequence.append(Standardization(**arghs))
sequence.append(SharedScale(
weights_init=Constant(1),
**arghs))
sequence.append(SharedShift(
biases_init=Constant(0),
**arghs))
sequence.append(self.activation)
self.sequence = FeedforwardSequence([
brick.apply for brick in sequence
], name="ffs")
self.children = [self.sequence]
@application(inputs=["input_"], outputs=["output"])
def apply(self, input_):
return self.sequence.apply(input_)
def get_dim(self, name):
try:
return dict(input_=self.shape,
output=self.shape)
except:
return super(NormalizedActivation, self).get_dim(name)
class NormalizedActivation(Initializable, Feedforward):
@lazy(allocation="shape broadcastable".split())
def __init__(self,
shape,
broadcastable,
activation=None,
batch_normalize=False,
**kwargs):
super(NormalizedActivation, self).__init__(**kwargs)
self.shape = shape
self.broadcastable = broadcastable
self.activation = activation or Rectifier()
self.batch_normalize = batch_normalize
@property
def broadcastable(self):
return self._broadcastable or [False] * len(self.shape)
@broadcastable.setter
def broadcastable(self, broadcastable):
self._broadcastable = broadcastable
def _allocate(self):
arghs = dict(shape=self.shape, broadcastable=self.broadcastable)
sequence = []
if self.batch_normalize:
sequence.append(Standardization(**arghs))
sequence.append(SharedScale(weights_init=Constant(1), **arghs))
sequence.append(SharedShift(biases_init=Constant(0), **arghs))
sequence.append(self.activation)
self.sequence = FeedforwardSequence(
[brick.apply for brick in sequence], name="ffs")
self.children = [self.sequence]
@application(inputs=["input_"], outputs=["output"])
def apply(self, input_):
return self.sequence.apply(input_)
def get_dim(self, name):
try:
return dict(input_=self.shape, output=self.shape)
except:
return super(NormalizedActivation, self).get_dim(name)
def build_fork_lookup(vocab_size, args):
x = tensor.lmatrix('features')
virtual_dim = 6
time_length = 5
mini_batch_size = 2
skip_connections = True
layers = 3
# Build the model
output_names = []
output_dims = []
for d in range(layers):
if d > 0:
suffix = '_' + str(d)
else:
suffix = ''
if d == 0 or skip_connections:
output_names.append("inputs" + suffix)
output_dims.append(virtual_dim)
print output_names
print output_dims
lookup = LookupTable(length=vocab_size, dim=virtual_dim)
lookup.weights_init = initialization.IsotropicGaussian(0.1)
lookup.biases_init = initialization.Constant(0)
fork = Fork(output_names=output_names,
input_dim=time_length,
output_dims=output_dims,
prototype=FeedforwardSequence([lookup.apply]))
# Return list of 3D Tensor, one for each layer
# (Batch X Time X embedding_dim)
pre_rnn = fork.apply(x)
fork.initialize()
f = theano.function([x], pre_rnn)
return f
def net_dvc(image_size=(32, 32)):
convos = [5, 5, 5]
pools = [2, 2, 2]
filters = [100, 200, 300]
tuplify = lambda x: (x, x)
convos = list(map(tuplify, convos))
conv_layers = [Convolutional(filter_size=s,num_filters=o, num_channels=i, name="Conv"+str(n))\
for s,o,i,n in zip(convos, filters, [3] + filters, range(1000))]
pool_layers = [MaxPooling(p) for p in map(tuplify, pools)]
activations = [Rectifier() for i in convos]
layers = [i for l in zip(conv_layers, activations, pool_layers) for i in l]
cnn = ConvolutionalSequence(layers,
3,
image_size=image_size,
name="cnn",
weights_init=Uniform(width=.1),
biases_init=Constant(0))
cnn._push_allocation_config()
cnn_output = np.prod(cnn.get_dim('output'))
mlp_size = [cnn_output, 500, 2]
mlp = MLP([Rectifier(), Softmax()],
mlp_size,
name="mlp",
weights_init=Uniform(width=.1),
biases_init=Constant(0))
seq = FeedforwardSequence([net.apply for net in [cnn, Flattener(), mlp]])
seq.push_initialization_config()
seq.initialize()
return seq
def build_model_hard(vocab_size, args, dtype=floatX):
logger.info('Building model ...')
# Parameters for the model
context = args.context
state_dim = args.state_dim
layers = args.layers
skip_connections = args.skip_connections
# Symbolic variables
# In both cases: Time X Batch
x = tensor.lmatrix('features')
y = tensor.lmatrix('targets')
# Build the model
output_names = []
output_dims = []
for d in range(layers):
if d > 0:
suffix = '_' + str(d)
else:
suffix = ''
if d == 0 or skip_connections:
output_names.append("inputs" + suffix)
output_dims.append(state_dim)
lookup = LookupTable(length=vocab_size, dim=state_dim)
lookup.weights_init = initialization.IsotropicGaussian(0.1)
lookup.biases_init = initialization.Constant(0)
fork = Fork(output_names=output_names,
input_dim=args.mini_batch_size,
output_dims=output_dims,
prototype=FeedforwardSequence([lookup.apply]))
transitions = [SimpleRecurrent(dim=state_dim, activation=Tanh())]
for i in range(layers - 1):
mlp = MLP(activations=[Logistic()],
dims=[2 * state_dim, 1],
weights_init=initialization.IsotropicGaussian(0.1),
biases_init=initialization.Constant(0),
name="mlp_" + str(i))
transitions.append(
HardGatedRecurrent(dim=state_dim, mlp=mlp, activation=Tanh()))
rnn = RecurrentStack(transitions, skip_connections=skip_connections)
# dim = layers * state_dim
output_layer = Linear(input_dim=layers * state_dim,
output_dim=vocab_size,
name="output_layer")
# Return list of 3D Tensor, one for each layer
# (Time X Batch X embedding_dim)
pre_rnn = fork.apply(x)
# Give a name to the input of each layer
if skip_connections:
for t in range(len(pre_rnn)):
pre_rnn[t].name = "pre_rnn_" + str(t)
else:
pre_rnn.name = "pre_rnn"
# Prepare inputs for the RNN
kwargs = OrderedDict()
init_states = {}
for d in range(layers):
if d > 0:
suffix = '_' + str(d)
else:
suffix = ''
if skip_connections:
kwargs['inputs' + suffix] = pre_rnn[d]
elif d == 0:
kwargs['inputs' + suffix] = pre_rnn
init_states[d] = theano.shared(numpy.zeros(
(args.mini_batch_size, state_dim)).astype(floatX),
name='state0_%d' % d)
kwargs['states' + suffix] = init_states[d]
# Apply the RNN to the inputs
h = rnn.apply(low_memory=True, **kwargs)
# Now we have correctly:
# h = [state_1, state_2, state_3 ...]
# Save all the last states
last_states = {}
for d in range(layers):
last_states[d] = h[d][-1, :, :]
# Concatenate all the states
if layers > 1:
h = tensor.concatenate(h, axis=2)
h.name = "hidden_state"
# The updates of the hidden states
updates = []
for d in range(layers):
updates.append((init_states[d], last_states[d]))
presoft = output_layer.apply(h[context:, :, :])
# Define the cost
# Compute the probability distribution
time, batch, feat = presoft.shape
presoft.name = 'presoft'
cross_entropy = Softmax().categorical_cross_entropy(
y[context:, :].flatten(), presoft.reshape((batch * time, feat)))
cross_entropy = cross_entropy / tensor.log(2)
cross_entropy.name = "cross_entropy"
# TODO: add regularisation for the cost
# the log(1) is here in order to differentiate the two variables
# for monitoring
cost = cross_entropy + tensor.log(1)
cost.name = "regularized_cost"
# Initialize the model
logger.info('Initializing...')
fork.initialize()
rnn.weights_init = initialization.Orthogonal()
rnn.biases_init = initialization.Constant(0)
rnn.initialize()
output_layer.weights_init = initialization.IsotropicGaussian(0.1)
output_layer.biases_init = initialization.Constant(0)
output_layer.initialize()
return cost, cross_entropy, updates
def get_prernn(args):
# time x batch
x_mask = tensor.fmatrix('mask')
# Compute the state dim
if args.rnn_type == 'lstm':
state_dim = 4 * args.state_dim
else:
state_dim = args.state_dim
# Prepare the arguments for the fork
output_names = []
output_dims = []
for d in range(args.layers):
if d > 0:
suffix = RECURRENTSTACK_SEPARATOR + str(d)
else:
suffix = ''
if d == 0 or args.skip_connections:
output_names.append("inputs" + suffix)
output_dims.append(state_dim)
# Prepare the brick to be forked (LookupTable or Linear)
# Check if the dataset provides indices (in the case of a
# fixed vocabulary, x is 2D tensor) or if it gives raw values
# (x is 3D tensor)
if has_indices(args.dataset):
features = args.mini_batch_size
x = tensor.lmatrix('features')
vocab_size = get_output_size(args.dataset)
lookup = LookupTable(length=vocab_size, dim=state_dim)
lookup.weights_init = initialization.IsotropicGaussian(0.1)
lookup.biases_init = initialization.Constant(0)
forked = FeedforwardSequence([lookup.apply])
if not has_mask(args.dataset):
x_mask = tensor.ones_like(x, dtype=floatX)
else:
x = tensor.tensor3('features', dtype=floatX)
if args.used_inputs is not None:
x = tensor.set_subtensor(x[args.used_inputs:, :, :],
tensor.zeros_like(x[args.used_inputs:,
:, :],
dtype=floatX))
features = get_output_size(args.dataset)
forked = Linear(input_dim=features, output_dim=state_dim)
forked.weights_init = initialization.IsotropicGaussian(0.1)
forked.biases_init = initialization.Constant(0)
if not has_mask(args.dataset):
x_mask = tensor.ones_like(x[:, :, 0], dtype=floatX)
# Define the fork
fork = Fork(output_names=output_names, input_dim=features,
output_dims=output_dims,
prototype=forked)
fork.initialize()
# Apply the fork
prernn = fork.apply(x)
# Give a name to the input of each layer
if args.skip_connections:
for t in range(len(prernn)):
prernn[t].name = "pre_rnn_" + str(t)
else:
prernn.name = "pre_rnn"
return prernn, x_mask
def get_prernn(args):
# time x batch
x_mask = tensor.fmatrix('mask')
# Compute the state dim
if args.rnn_type == 'lstm':
state_dim = 4 * args.state_dim
else:
state_dim = args.state_dim
# Prepare the arguments for the fork
output_names = []
output_dims = []
for d in range(args.layers):
if d > 0:
suffix = RECURRENTSTACK_SEPARATOR + str(d)
else:
suffix = ''
if d == 0 or args.skip_connections:
output_names.append("inputs" + suffix)
output_dims.append(state_dim)
# Prepare the brick to be forked (LookupTable or Linear)
# Check if the dataset provides indices (in the case of a
# fixed vocabulary, x is 2D tensor) or if it gives raw values
# (x is 3D tensor)
if has_indices(args.dataset):
features = args.mini_batch_size
x = tensor.lmatrix('features')
vocab_size = get_output_size(args.dataset)
lookup = LookupTable(length=vocab_size, dim=state_dim)
lookup.weights_init = initialization.IsotropicGaussian(0.1)
lookup.biases_init = initialization.Constant(0)
forked = FeedforwardSequence([lookup.apply])
if not has_mask(args.dataset):
x_mask = tensor.ones_like(x, dtype=floatX)
else:
x = tensor.tensor3('features', dtype=floatX)
if args.used_inputs is not None:
x = tensor.set_subtensor(
x[args.used_inputs:, :, :],
tensor.zeros_like(x[args.used_inputs:, :, :], dtype=floatX))
features = get_output_size(args.dataset)
forked = Linear(input_dim=features, output_dim=state_dim)
forked.weights_init = initialization.IsotropicGaussian(0.1)
forked.biases_init = initialization.Constant(0)
if not has_mask(args.dataset):
x_mask = tensor.ones_like(x[:, :, 0], dtype=floatX)
# Define the fork
fork = Fork(output_names=output_names,
input_dim=features,
output_dims=output_dims,
prototype=forked)
fork.initialize()
# Apply the fork
prernn = fork.apply(x)
# Give a name to the input of each layer
if args.skip_connections:
for t in range(len(prernn)):
prernn[t].name = "pre_rnn_" + str(t)
else:
prernn.name = "pre_rnn"
return prernn, x_mask
def build_fork_lookup(vocab_size, time_length, args):
x = tensor.lmatrix('features')
virtual_dim = 6
state_dim = 6
skip_connections = False
layers = 1
# Build the model
output_names = []
output_dims = []
for d in range(layers):
if d > 0:
suffix = '_' + str(d)
else:
suffix = ''
if d == 0 or skip_connections:
output_names.append("inputs" + suffix)
output_dims.append(virtual_dim)
lookup = LookupTable(length=vocab_size, dim=virtual_dim)
lookup.weights_init = initialization.IsotropicGaussian(0.1)
lookup.biases_init = initialization.Constant(0)
fork = Fork(output_names=output_names, input_dim=time_length,
output_dims=output_dims,
prototype=FeedforwardSequence(
[lookup.apply]))
# Note that this order of the periods makes faster modules flow in slower
# ones with is the opposite of the original paper
transitions = [ClockworkBase(dim=state_dim, activation=Tanh(),
period=2 ** i) for i in range(layers)]
rnn = RecurrentStack(transitions, skip_connections=skip_connections)
# Return list of 3D Tensor, one for each layer
# (Batch X Time X embedding_dim)
pre_rnn = fork.apply(x)
# Give time as the first index for each element in the list:
# (Time X Batch X embedding_dim)
if layers > 1 and skip_connections:
for t in range(len(pre_rnn)):
pre_rnn[t] = pre_rnn[t].dimshuffle(1, 0, 2)
else:
pre_rnn = pre_rnn.dimshuffle(1, 0, 2)
f_pre_rnn = theano.function([x], pre_rnn)
# Prepare inputs for the RNN
kwargs = OrderedDict()
for d in range(layers):
if d > 0:
suffix = '_' + str(d)
else:
suffix = ''
if d == 0 or skip_connections:
if skip_connections:
kwargs['inputs' + suffix] = pre_rnn[d]
else:
kwargs['inputs' + suffix] = pre_rnn
print kwargs
# Apply the RNN to the inputs
h = rnn.apply(low_memory=True, **kwargs)
fork.initialize()
rnn.weights_init = initialization.Orthogonal()
rnn.biases_init = initialization.Constant(0)
rnn.initialize()
f_h = theano.function([x], h)
return f_pre_rnn, f_h
def build_model(images, labels):
vgg = VGG(layer='conv3_4')
vgg.push_initialization_config()
vgg.initialize()
sb = SubstractBatch()
# Construct a bottom convolutional sequence
layers = [
Convolutional(filter_size=(3, 3),
num_filters=100,
use_bias=True,
tied_biases=True,
name='final_conv0'),
BatchNormalization(name='batchnorm_1'),
Rectifier(name='final_conv0_act'),
Convolutional(filter_size=(3, 3),
num_filters=100,
use_bias=True,
tied_biases=True,
name='final_conv1'),
BatchNormalization(name='batchnorm_2'),
Rectifier(name='final_conv1_act'),
MaxPooling(pooling_size=(2, 2), name='maxpool_final')
]
bottom_conv_sequence = ConvolutionalSequence(
layers,
num_channels=256,
image_size=(40, 40),
biases_init=Constant(0.),
weights_init=IsotropicGaussian(0.01))
bottom_conv_sequence._push_allocation_config()
# Flatten layer
flattener = Flattener()
# Construct a top MLP
conv_out_dim = numpy.prod(bottom_conv_sequence.get_dim('output'))
print 'dim output conv:', bottom_conv_sequence.get_dim('output')
# conv_out_dim = 20 * 40 * 40
top_mlp = BatchNormalizedMLP(
[Rectifier(name='non_linear_9'),
Softmax(name='non_linear_11')], [conv_out_dim, 1024, 10],
weights_init=IsotropicGaussian(),
biases_init=Constant(0))
# Construct feedforward sequence
ss_seq = FeedforwardSequence([
vgg.apply, bottom_conv_sequence.apply, flattener.apply, top_mlp.apply
])
ss_seq.push_initialization_config()
ss_seq.initialize()
prediction = ss_seq.apply(images)
cost_noreg = CategoricalCrossEntropy().apply(labels.flatten(), prediction)
# add regularization
selector = Selector([top_mlp])
Ws = selector.get_parameters('W')
mlp_brick_name = 'batchnormalizedmlp'
W0 = Ws['/%s/linear_0.W' % mlp_brick_name]
W1 = Ws['/%s/linear_1.W' % mlp_brick_name]
cost = cost_noreg + .0001 * (W0**2).sum() + .001 * (W1**2).sum()
# define learned parameters
selector = Selector([ss_seq])
Ws = selector.get_parameters('W')
bs = selector.get_parameters('b')
BNSCs = selector.get_parameters('batch_norm_scale')
BNSHs = selector.get_parameters('batch_norm_shift')
parameters_top = []
parameters_top += [v for k, v in Ws.items()]
parameters_top += [v for k, v in bs.items()]
parameters_top += [v for k, v in BNSCs.items()]
parameters_top += [v for k, v in BNSHs.items()]
selector = Selector([vgg])
convs = selector.get_parameters()
parameters_all = []
parameters_all += parameters_top
parameters_all += [v for k, v in convs.items()]
return cost, [parameters_top, parameters_all]
def build_model_lstm(vocab_size, args, dtype=floatX):
logger.info('Building model ...')
# Parameters for the model
context = args.context
state_dim = args.state_dim
layers = args.layers
skip_connections = args.skip_connections
virtual_dim = 4 * state_dim
# Symbolic variables
# In both cases: Time X Batch
x = tensor.lmatrix('features')
y = tensor.lmatrix('targets')
# Build the model
output_names = []
output_dims = []
for d in range(layers):
if d > 0:
suffix = '_' + str(d)
else:
suffix = ''
if d == 0 or skip_connections:
output_names.append("inputs" + suffix)
output_dims.append(virtual_dim)
lookup = LookupTable(length=vocab_size, dim=virtual_dim)
lookup.weights_init = initialization.IsotropicGaussian(0.1)
lookup.biases_init = initialization.Constant(0)
# Make sure time_length is what we need
fork = Fork(output_names=output_names,
input_dim=args.mini_batch_size,
output_dims=output_dims,
prototype=FeedforwardSequence([lookup.apply]))
transitions = [
LSTM(dim=state_dim, activation=Tanh()) for _ in range(layers)
]
rnn = RecurrentStack(transitions, skip_connections=skip_connections)
# If skip_connections: dim = layers * state_dim
# else: dim = state_dim
output_layer = Linear(input_dim=skip_connections * layers * state_dim +
(1 - skip_connections) * state_dim,
output_dim=vocab_size,
name="output_layer")
# Return list of 3D Tensor, one for each layer
# (Time X Batch X embedding_dim)
pre_rnn = fork.apply(x)
# Give a name to the input of each layer
if skip_connections:
for t in range(len(pre_rnn)):
pre_rnn[t].name = "pre_rnn_" + str(t)
else:
pre_rnn.name = "pre_rnn"
# Prepare inputs for the RNN
kwargs = OrderedDict()
init_states = {}
init_cells = {}
for d in range(layers):
if d > 0:
suffix = '_' + str(d)
else:
suffix = ''
if skip_connections:
kwargs['inputs' + suffix] = pre_rnn[d]
elif d == 0:
kwargs['inputs'] = pre_rnn
init_states[d] = theano.shared(numpy.zeros(
(args.mini_batch_size, state_dim)).astype(floatX),
name='state0_%d' % d)
init_cells[d] = theano.shared(numpy.zeros(
(args.mini_batch_size, state_dim)).astype(floatX),
name='cell0_%d' % d)
kwargs['states' + suffix] = init_states[d]
kwargs['cells' + suffix] = init_cells[d]
# Apply the RNN to the inputs
h = rnn.apply(low_memory=True, **kwargs)
# h = [state, cell, in, forget, out, state_1,
# cell_1, in_1, forget_1, out_1 ...]
last_states = {}
last_cells = {}
for d in range(layers):
last_states[d] = h[5 * d][-1, :, :]
last_cells[d] = h[5 * d + 1][-1, :, :]
# The updates of the hidden states
updates = []
for d in range(layers):
updates.append((init_states[d], last_states[d]))
updates.append((init_cells[d], last_states[d]))
# h = [state, cell, in, forget, out, state_1,
# cell_1, in_1, forget_1, out_1 ...]
# Extract the values
in_gates = h[2::5]
forget_gates = h[3::5]
out_gates = h[4::5]
gate_values = {
"in_gates": in_gates,
"forget_gates": forget_gates,
"out_gates": out_gates
}
h = h[::5]
# Now we have correctly:
# h = [state, state_1, state_2 ...] if layers > 1
# h = [state] if layers == 1
# If we have skip connections, concatenate all the states
# Else only consider the state of the highest layer
if layers > 1:
if skip_connections:
h = tensor.concatenate(h, axis=2)
else:
h = h[-1]
else:
h = h[0]
h.name = "hidden_state"
presoft = output_layer.apply(h[context:, :, :])
# Define the cost
# Compute the probability distribution
time, batch, feat = presoft.shape
presoft.name = 'presoft'
cross_entropy = Softmax().categorical_cross_entropy(
y[context:, :].flatten(), presoft.reshape((batch * time, feat)))
cross_entropy = cross_entropy / tensor.log(2)
cross_entropy.name = "cross_entropy"
# TODO: add regularisation for the cost
# the log(1) is here in order to differentiate the two variables
# for monitoring
cost = cross_entropy + tensor.log(1)
cost.name = "regularized_cost"
# Initialize the model
logger.info('Initializing...')
fork.initialize()
# Dont initialize as Orthogonal if we are about to load new parameters
if args.load_path is not None:
rnn.weights_init = initialization.Constant(0)
else:
rnn.weights_init = initialization.Orthogonal()
rnn.biases_init = initialization.Constant(0)
rnn.initialize()
output_layer.weights_init = initialization.IsotropicGaussian(0.1)
output_layer.biases_init = initialization.Constant(0)
output_layer.initialize()
return cost, cross_entropy, updates, gate_values
