def softmax(z):
assert z.ndim >= 1
if z.ndim <= 2:
return T.nnet.softmax(z)
else:
from TheanoUtil import time_batch_make_flat
z_flat = time_batch_make_flat(z)
assert z_flat.ndim == 2
return T.reshape(T.nnet.softmax(z_flat), z.shape)
def softmax(z):
assert z.ndim >= 1
if z.ndim <= 2:
return T.nnet.softmax(z)
else:
from TheanoUtil import time_batch_make_flat
z_flat = time_batch_make_flat(z)
assert z_flat.ndim == 2
return T.reshape(T.nnet.softmax(z_flat), z.shape)
def add_layer(self, layer):
"""
:type layer: NetworkHiddenLayer.Layer
:rtype NetworkHiddenLayer.Layer
"""
assert layer.name
layer_errors = layer.errors()
if isinstance(layer, OutputLayer) or layer.name == "output" or layer_errors is not None:
is_output_layer = True
self.output[layer.name] = layer
else:
is_output_layer = False
self.hidden[layer.name] = layer
self.add_cost_and_constraints(layer)
if layer_errors is not None:
self.errors[layer.name] = layer_errors
if is_output_layer:
if getattr(layer, "p_y_given_x", None) is None and layer.output:
# Small little hack for layers which we use as output-layers which don't set this.
from TheanoUtil import time_batch_make_flat
layer.p_y_given_x = time_batch_make_flat(layer.output)
self.declare_train_params()
return layer
def add_layer(self, layer):
"""
:type layer: NetworkHiddenLayer.Layer
:rtype NetworkHiddenLayer.Layer
"""
assert layer.name
layer_errors = layer.errors()
if isinstance(layer, OutputLayer) or layer.name == "output" or layer_errors is not None:
is_output_layer = True
self.output[layer.name] = layer
else:
is_output_layer = False
self.hidden[layer.name] = layer
if layer_errors is not None:
self.errors[layer.name] = layer_errors
if is_output_layer:
if getattr(layer, "p_y_given_x", None) is None and layer.output:
# Small little hack for layers which we use as output-layers whicgh don't set this.
from TheanoUtil import time_batch_make_flat
layer.p_y_given_x = layer.output
layer.p_y_given_x_flat = time_batch_make_flat(layer.output)
self.declare_train_params()
return layer
def __init__(self, sources, n_out, index, y_in=None, target=None, target_index=None,
sparse=False, cost_scale=1.0,
L1=0.0, L2=0.0, L2_eye=None, varreg=0.0,
output_L2_reg=0.0, output_entropy_reg=0.0, output_entropy_exp_reg=0.0,
with_bias=True,
mask="unity", dropout=0.0, batch_norm=False, layer_drop=0.0, residual=False,
carry=False,
sparse_filtering=False, gradient_scale=1.0, trainable=True, device=None,
dtype='float32',
**kwargs):
"""
:param list[NetworkBaseLayer.Layer] sources: list of source layers
:param int n_out: output dim of W_in and dim of bias
:param float L1: l1-param-norm regularization
:param float L2: l2-param-norm regularization
:param str mask: "unity" or "dropout"
:type dropout: float
"""
super(Layer, self).__init__(**kwargs)
self.index = index
self.sources = sources; ":type: list[Layer]"
self.num_sources = len(sources)
if mask is None: mask = 'none'
self.set_attr('mask', mask)
self.set_attr('dropout', dropout)
self.set_attr('sparse', sparse)
self.set_attr('sparse_filtering', sparse_filtering)
if not trainable:
self.set_attr('trainable', trainable) # only store if not default
self.gradient_scale = 0.0 # just to be sure
else:
self.gradient_scale = gradient_scale
if gradient_scale != 1.0:
self.set_attr('gradient_scale', gradient_scale)
self.set_attr('layer_drop', layer_drop)
assert not carry, "not supported anymore"
self.set_attr('residual', residual)
self.set_attr('n_out', n_out)
self.set_attr('L1', L1)
self.set_attr('L2', L2)
if L2_eye:
self.set_attr('L2_eye', L2_eye)
self.device = device # if device else str(theano.config.device)
for s in self.sources:
s.transfer_output(self.device)
self.set_attr('varreg', varreg)
if output_L2_reg:
self.set_attr('output_L2_reg', output_L2_reg)
if output_entropy_reg:
self.set_attr('output_entropy_reg', output_entropy_reg)
if output_entropy_exp_reg:
self.set_attr('output_entropy_exp_reg', output_entropy_exp_reg)
self.set_attr('batch_norm', batch_norm)
if y_in is not None:
self.y_in = {}
for k in y_in:
if not isinstance(y_in[k], T.Variable): continue
self.y_in[k] = time_batch_make_flat(y_in[k]) # TODO: better not flatten here...
self.y_in[k].n_out = getattr(y_in[k], "n_out", None)
else:
self.y_in = None
self.constraints = T.constant(0)
if target:
self.set_attr('target', target)
if target_index:
self.set_attr('target_index', target_index)
assert target_index in self.network.j
self.index = index = self.network.j[target_index]
if cost_scale != 1:
self.set_attr("cost_scale", cost_scale)
if with_bias:
self.b = self.add_param(self.create_bias(n_out), 'b_%s'%self.name)
else:
self.set_attr('with_bias', False)
self.b = numpy.float32(0)
self.mass = T.constant(1., name = "mass_%s" % self.name, dtype='float32')
self.masks = [None] * len(self.sources)
assert mask in ['dropout', 'unity', 'none'], "invalid mask: %s" % mask
if mask == "dropout" or (mask == 'none' and dropout > 0):
assert 0.0 < dropout < 1.0
# If we apply this mass during training then we don't need any mask or mass for testing.
# The expected weight should be 1 in
# E[x] = mass * (1-dropout)
# so mass has to be 1 / (1 - dropout).
self.mass = T.constant(1.0 / (1.0 - dropout), dtype='float32')
from theano.sandbox.rng_mrg import MRG_RandomStreams as RandomStreams
srng = RandomStreams(self.rng.randint(1234) + 1)
if self.depth > 1:
self.masks = [T.cast(srng.binomial(n=1, p=1 - dropout, size=(s.attrs['n_out'],self.depth)), theano.config.floatX) for s in self.sources]
else:
self.masks = [T.cast(srng.binomial(n=1, p=1 - dropout, size=(s.attrs['n_out'],)), theano.config.floatX) for s in self.sources]
def __init__(self, loss, y, dtype=None, copy_input=None, copy_output=None, time_limit=0,
grad_clip_z=None, grad_discard_out_of_bound_z=None,
**kwargs):
"""
:param theano.Variable index: index for batches
:param str loss: e.g. 'ce'
"""
super(OutputLayer, self).__init__(**kwargs)
if dtype:
self.set_attr('dtype', dtype)
if copy_input:
self.set_attr("copy_input", copy_input.name)
if grad_clip_z is not None:
self.set_attr("grad_clip_z", grad_clip_z)
if grad_discard_out_of_bound_z is not None:
self.set_attr("grad_discard_out_of_bound_z", grad_discard_out_of_bound_z)
if not copy_input:
self.z = self.b
self.W_in = [self.add_param(self.create_forward_weights(source.attrs['n_out'], self.attrs['n_out'],
name="W_in_%s_%s" % (source.name, self.name)))
for source in self.sources]
assert len(self.sources) == len(self.masks) == len(self.W_in)
assert len(self.sources) > 0
for source, m, W in zip(self.sources, self.masks, self.W_in):
source_output = source.output
#4D input from TwoD Layers -> collapse height dimension
if source_output.ndim == 4:
source_output = source_output.sum(axis=0)
if source.attrs['sparse']:
if source.output.ndim == 3:
input = source_output[:,:,0] # old sparse format
else:
assert source_output.ndim == 2
input = source.output
self.z += W[T.cast(input, 'int32')]
elif m is None:
self.z += self.dot(source_output, W)
else:
self.z += self.dot(self.mass * m * source_output, W)
else:
self.z = copy_input.output
assert self.z.ndim == 3
if grad_clip_z is not None:
grad_clip_z = numpy.float32(grad_clip_z)
self.z = theano.gradient.grad_clip(self.z, -grad_clip_z, grad_clip_z)
if grad_discard_out_of_bound_z is not None:
grad_discard_out_of_bound_z = numpy.float32(grad_discard_out_of_bound_z)
self.z = grad_discard_out_of_bound(self.z, -grad_discard_out_of_bound_z, grad_discard_out_of_bound_z)
if not copy_output:
self.y = y
else:
self.index = copy_output.index
self.y = copy_output.y_out
if isinstance(y, T.Variable):
self.y_data_flat = time_batch_make_flat(y)
else:
assert self.attrs.get("target", "").endswith("[sparse:coo]")
assert isinstance(self.y, tuple)
assert len(self.y) == 3
s0, s1, weight = self.y
from NativeOp import max_and_argmax_sparse
n_time = self.z.shape[0]
n_batch = self.z.shape[1]
mask = self.network.j[self.attrs.get("target", "").replace("[sparse:coo]", "[sparse:coo:2:0]")]
out_arg = T.zeros((n_time, n_batch), dtype="float32")
out_max = T.zeros((n_time, n_batch), dtype="float32") - numpy.float32(1e16)
out_arg, out_max = max_and_argmax_sparse(s0, s1, weight, mask, out_arg, out_max)
assert out_arg.ndim == 2
self.y_data_flat = out_arg.astype("int32")
self.norm = numpy.float32(1)
self.target_index = self.index
if time_limit == 'inf':
#target_length = self.index.shape[0]
#mass = T.cast(T.sum(self.index),'float32')
#self.index = theano.ifelse.ifelse(T.gt(self.z.shape[0],target_length),self.sources[0].index,self.index)
#self.norm = mass / T.cast(T.sum(self.index),'float32')
num = T.cast(T.sum(self.index), 'float32')
if self.eval_flag:
self.index = self.sources[0].index
else:
import theano.ifelse
padx = T.zeros((T.abs_(self.index.shape[0] - self.z.shape[0]), self.index.shape[1], self.z.shape[2]), 'float32') + self.z[-1]
pady = T.zeros((T.abs_(self.index.shape[0] - self.z.shape[0]), self.index.shape[1]), 'int32') #+ y[-1]
padi = T.ones((T.abs_(self.index.shape[0] - self.z.shape[0]), self.index.shape[1]), 'int8')
self.z = theano.ifelse.ifelse(T.lt(self.z.shape[0], self.index.shape[0]),
T.concatenate([self.z,padx],axis=0), self.z)
#self.z = theano.ifelse.ifelse(T.gt(self.z.shape[0], self.index.shape[0]),self.z[:self.index.shape[0]], self.z)
self.y_data_flat = time_batch_make_flat(theano.ifelse.ifelse(T.gt(self.z.shape[0],self.index.shape[0]),
T.concatenate([y,pady], axis=0), y))
#self.index = theano.ifelse.ifelse(T.gt(self.z.shape[0], self.index.shape[0]), T.concatenate([T.ones((self.z.shape[0] - self.index.shape[0],self.z.shape[1]),'int8'), self.index], axis=0), self.index)
self.index = theano.ifelse.ifelse(T.gt(self.z.shape[0], self.index.shape[0]),
T.concatenate([padi,self.index],axis=0),self.index)
self.norm = num / T.cast(T.sum(self.index),'float32')
elif time_limit > 0:
end = T.min([self.z.shape[0], T.constant(time_limit, 'int32')])
nom = T.cast(T.sum(self.index),'float32')
self.index = T.set_subtensor(self.index[end:], T.zeros_like(self.index[end:]))
self.norm = nom / T.cast(T.sum(self.index),'float32')
self.z = T.set_subtensor(self.z[end:], T.zeros_like(self.z[end:]))
#xs = [s.output for s in self.sources]
#self.z = AccumulatorOpInstance(*[self.b] + xs + self.W_in)
#outputs_info = None #[ T.alloc(numpy.cast[theano.config.floatX](0), index.shape[1], self.attrs['n_out']) ]
#self.z, _ = theano.scan(step,
# sequences = [s.output for s in self.sources],
# non_sequences = self.W_in + [self.b])
self.set_attr('from', ",".join([s.name for s in self.sources]))
self.i = (self.index.flatten() > 0).nonzero()
self.j = ((1 - self.index.flatten()) > 0).nonzero()
self.loss = loss.encode("utf8")
self.attrs['loss'] = self.loss
if self.loss == 'priori':
self.priori = self.shared(value=numpy.ones((self.attrs['n_out'],), dtype=theano.config.floatX), borrow=True)
#self.make_output(self.z, collapse = False)
# Note that self.output is going to be overwritten in our derived classes.
self.output = self.make_consensus(self.z) if self.depth > 1 else self.z
def __init__(self, sources, n_out, index, y_in=None, target=None, target_index=None,
sparse=False, cost_scale=1.0, input_scale=1.0,
L1=0.0, L2=0.0, L2_eye=None, varreg=0.0,
output_L2_reg=0.0, output_entropy_reg=0.0, output_entropy_exp_reg=0.0,
with_bias=True,
mask="unity", dropout=0.0, batch_drop=False, batch_norm=False, bn_use_sample=False, layer_drop=0.0, residual=False,
carry=False,
sparse_filtering=False, gradient_scale=1.0, trainable=True, device=None,
dtype='float32',
**kwargs):
"""
:param list[NetworkBaseLayer.Layer] sources: list of source layers
:param int n_out: output dim of W_in and dim of bias
:param float L1: l1-param-norm regularization
:param float L2: l2-param-norm regularization
:param str mask: "unity" or "dropout"
:type dropout: float
"""
super(Layer, self).__init__(**kwargs)
self.index = index
self.sources = sources; ":type: list[Layer]"
self.num_sources = len(sources)
self.D = max([s.D for s in sources if isinstance(s,Layer)] + [0])
if mask is None: mask = 'none'
self.set_attr('mask', mask)
self.set_attr('dropout', dropout)
self.set_attr('sparse', sparse)
self.set_attr('bn_use_sample', bn_use_sample)
self.set_attr('sparse_filtering', sparse_filtering)
if not trainable:
self.set_attr('trainable', trainable) # only store if not default
self.gradient_scale = 0.0 # just to be sure
else:
self.gradient_scale = gradient_scale
if gradient_scale != 1.0:
self.set_attr('gradient_scale', gradient_scale)
self.set_attr('layer_drop', layer_drop)
assert not carry, "not supported anymore"
self.set_attr('residual', residual)
self.set_attr('n_out', n_out)
self.set_attr('L1', L1)
self.set_attr('L2', L2)
if L2_eye:
self.set_attr('L2_eye', L2_eye)
self.device = device # if device else str(theano.config.device)
for s in self.sources:
s.transfer_output(self.device)
self.set_attr('varreg', varreg)
if output_L2_reg:
self.set_attr('output_L2_reg', output_L2_reg)
if output_entropy_reg:
self.set_attr('output_entropy_reg', output_entropy_reg)
if output_entropy_exp_reg:
self.set_attr('output_entropy_exp_reg', output_entropy_exp_reg)
self.set_attr('batch_norm', batch_norm)
self.set_attr('input_scale', input_scale)
if y_in is not None:
self.y_in = {}
for k in y_in:
if not isinstance(y_in[k], T.Variable): continue
self.y_in[k] = time_batch_make_flat(y_in[k]) # TODO: better not flatten here...
self.y_in[k].n_out = getattr(y_in[k], "n_out", None)
else:
self.y_in = None
self.constraints = T.constant(0)
if target:
self.set_attr('target', target)
if target_index:
self.set_attr('target_index', target_index)
assert target_index in self.network.j
self.index = index = self.network.j[target_index]
if cost_scale != 1:
self.set_attr("cost_scale", cost_scale)
if with_bias:
self.b = self.add_param(self.create_bias(n_out), 'b_%s'%self.name)
else:
self.set_attr('with_bias', False)
self.b = numpy.float32(0)
self.mass = T.constant(1., name = "mass_%s" % self.name, dtype='float32')
self.masks = [None] * len(self.sources)
assert mask in ['dropout', 'unity', 'none'], "invalid mask: %s" % mask
if mask == "dropout" or (mask == 'none' and dropout > 0):
assert 0.0 < dropout < 1.0
# If we apply this mass during training then we don't need any mask or mass for testing.
# The expected weight should be 1 in
# E[x] = mass * (1-dropout)
# so mass has to be 1 / (1 - dropout).
self.mass = T.constant(1.0 / (1.0 - dropout), dtype='float32')
from theano.sandbox.rng_mrg import MRG_RandomStreams as RandomStreams
srng = RandomStreams(self.rng.randint(1234) + 1)
if self.depth > 1:
self.masks = [T.cast(srng.binomial(n=1, p=1 - dropout, size=(s.attrs['n_out'],self.depth)), theano.config.floatX) for s in self.sources]
else:
if batch_drop:
self.masks = [T.cast(srng.binomial(n=1, p=1 - dropout, size=s.output.shape), theano.config.floatX) for s in self.sources]
else:
self.masks = [T.cast(srng.binomial(n=1, p=1 - dropout, size=(s.attrs['n_out'],)), theano.config.floatX) for s in self.sources]
def __init__(self, loss, y, dtype=None, copy_input=None, copy_output=None, time_limit=0,
use_source_index=False,
compute_priors=False, compute_priors_exp_average=0, compute_distortions=False,
softmax_smoothing=1.0, grad_clip_z=None, grad_discard_out_of_bound_z=None, normalize_length=False,
exclude_labels=[],
apply_softmax=True,
substract_prior_from_output=False,
input_output_similarity=None,
input_output_similarity_scale=1,
**kwargs):
"""
:param theano.Variable index: index for batches
:param str loss: e.g. 'ce'
"""
super(OutputLayer, self).__init__(**kwargs)
self.set_attr("normalize_length", normalize_length)
if dtype:
self.set_attr('dtype', dtype)
if copy_input:
self.set_attr("copy_input", copy_input.name)
if grad_clip_z is not None:
self.set_attr("grad_clip_z", grad_clip_z)
if compute_distortions:
self.set_attr("compute_distortions", compute_distortions)
if grad_discard_out_of_bound_z is not None:
self.set_attr("grad_discard_out_of_bound_z", grad_discard_out_of_bound_z)
if not apply_softmax:
self.set_attr("apply_softmax", apply_softmax)
if substract_prior_from_output:
self.set_attr("substract_prior_from_output", substract_prior_from_output)
if input_output_similarity:
self.set_attr("input_output_similarity", input_output_similarity)
self.set_attr("input_output_similarity_scale", input_output_similarity_scale)
if use_source_index:
self.set_attr("use_source_index", use_source_index)
src_index = self.sources[0].index
self.index = src_index
if not copy_input:
self.z = self.b
self.W_in = [self.add_param(self.create_forward_weights(source.attrs['n_out'], self.attrs['n_out'],
name="W_in_%s_%s" % (source.name, self.name)))
for source in self.sources]
assert len(self.sources) == len(self.masks) == len(self.W_in)
assert len(self.sources) > 0
for source, m, W in zip(self.sources, self.masks, self.W_in):
source_output = source.output
# 4D input from TwoD Layers -> collapse height dimension
if source_output.ndim == 4:
source_output = source_output.sum(axis=0)
if source.attrs['sparse']:
if source.output.ndim == 3:
input = source_output[:, :, 0] # old sparse format
else:
assert source_output.ndim == 2
input = source.output
self.z += W[T.cast(input, 'int32')]
elif m is None:
self.z += self.dot(source_output, W)
else:
self.z += self.dot(self.mass * m * source_output, W)
else:
self.z = copy_input.output
assert self.z.ndim == 3
if grad_clip_z is not None:
grad_clip_z = numpy.float32(grad_clip_z)
self.z = theano.gradient.grad_clip(self.z, -grad_clip_z, grad_clip_z)
if grad_discard_out_of_bound_z is not None:
grad_discard_out_of_bound_z = numpy.float32(grad_discard_out_of_bound_z)
self.z = grad_discard_out_of_bound(self.z, -grad_discard_out_of_bound_z, grad_discard_out_of_bound_z)
if not copy_output:
self.y = y
else:
self.index = copy_output.index
self.y = copy_output.y_out
if y is None:
self.y_data_flat = None
elif isinstance(y, T.Variable):
self.y_data_flat = time_batch_make_flat(y)
else:
assert self.attrs.get("target", "").endswith("[sparse:coo]")
assert isinstance(self.y, tuple)
assert len(self.y) == 3
s0, s1, weight = self.y
from NativeOp import max_and_argmax_sparse
n_time = self.z.shape[0]
n_batch = self.z.shape[1]
mask = self.network.j[self.attrs.get("target", "").replace("[sparse:coo]", "[sparse:coo:2:0]")]
out_arg = T.zeros((n_time, n_batch), dtype="float32")
out_max = T.zeros((n_time, n_batch), dtype="float32") - numpy.float32(1e16)
out_arg, out_max = max_and_argmax_sparse(s0, s1, weight, mask, out_arg, out_max)
assert out_arg.ndim == 2
self.y_data_flat = out_arg.astype("int32")
self.norm = numpy.float32(1)
self.target_index = self.index
if time_limit == 'inf':
# target_length = self.index.shape[0]
# mass = T.cast(T.sum(self.index),'float32')
# self.index = theano.ifelse.ifelse(T.gt(self.z.shape[0],target_length),self.sources[0].index,self.index)
# self.norm = mass / T.cast(T.sum(self.index),'float32')
num = T.cast(T.sum(self.index), 'float32')
if self.eval_flag:
self.index = self.sources[0].index
else:
import theano.ifelse
padx = T.zeros((T.abs_(self.index.shape[0] - self.z.shape[0]), self.index.shape[1], self.z.shape[2]),
'float32') + self.z[-1]
pady = T.zeros((T.abs_(self.index.shape[0] - self.z.shape[0]), self.index.shape[1]), 'int32') # + y[-1]
padi = T.ones((T.abs_(self.index.shape[0] - self.z.shape[0]), self.index.shape[1]), 'int8')
self.z = theano.ifelse.ifelse(T.lt(self.z.shape[0], self.index.shape[0]),
T.concatenate([self.z, padx], axis=0), self.z)
# self.z = theano.ifelse.ifelse(T.gt(self.z.shape[0], self.index.shape[0]),self.z[:self.index.shape[0]], self.z)
self.y_data_flat = time_batch_make_flat(theano.ifelse.ifelse(T.gt(self.z.shape[0], self.index.shape[0]),
T.concatenate([y, pady], axis=0), y))
# self.index = theano.ifelse.ifelse(T.gt(self.z.shape[0], self.index.shape[0]), T.concatenate([T.ones((self.z.shape[0] - self.index.shape[0],self.z.shape[1]),'int8'), self.index], axis=0), self.index)
self.index = theano.ifelse.ifelse(T.gt(self.z.shape[0], self.index.shape[0]),
T.concatenate([padi, self.index], axis=0), self.index)
self.norm *= num / T.cast(T.sum(self.index), 'float32')
elif time_limit > 0:
end = T.min([self.z.shape[0], T.constant(time_limit, 'int32')])
num = T.cast(T.sum(self.index), 'float32')
self.index = T.set_subtensor(self.index[end:], T.zeros_like(self.index[end:]))
self.norm = num / T.cast(T.sum(self.index), 'float32')
self.z = T.set_subtensor(self.z[end:], T.zeros_like(self.z[end:]))
# xs = [s.output for s in self.sources]
# self.z = AccumulatorOpInstance(*[self.b] + xs + self.W_in)
# outputs_info = None #[ T.alloc(numpy.cast[theano.config.floatX](0), index.shape[1], self.attrs['n_out']) ]
# self.z, _ = theano.scan(step,
# sequences = [s.output for s in self.sources],
# non_sequences = self.W_in + [self.b])
self.set_attr('from', ",".join([s.name for s in self.sources]))
index_flat = self.index.flatten()
for label in exclude_labels:
index_flat = T.set_subtensor(index_flat[(T.eq(self.y_data_flat, label) > 0).nonzero()], numpy.int8(0))
self.i = (index_flat > 0).nonzero()
self.j = ((numpy.int32(1) - index_flat) > 0).nonzero()
self.loss = as_str(loss.encode("utf8"))
self.attrs['loss'] = self.loss
if compute_priors:
self.set_attr('compute_priors', compute_priors)
if compute_priors_exp_average:
self.set_attr('compute_priors_exp_average', compute_priors_exp_average)
if softmax_smoothing != 1.0:
self.attrs['softmax_smoothing'] = softmax_smoothing
print >> log.v4, "Logits before the softmax scaled with factor ", softmax_smoothing
self.z *= numpy.float32(softmax_smoothing)
if self.loss == 'priori':
self.priori = self.shared(value=numpy.ones((self.attrs['n_out'],), dtype=theano.config.floatX), borrow=True)
if input_output_similarity:
# First a self-similarity of input and output,
# and then add -similarity or distance between those to the constraints,
# so that the input and output correlate on a frame-by-frame basis.
# Here some other similarities/distances we could try:
# http://docs.scipy.org/doc/scipy/reference/generated/scipy.spatial.distance.pdist.html
# https://brenocon.com/blog/2012/03/cosine-similarity-pearson-correlation-and-ols-coefficients/
from TheanoUtil import self_similarity_cosine
self_similarity = self_similarity_cosine # maybe other
data_layer = self.find_data_layer()
assert data_layer
assert data_layer.output.ndim == 3
n_time = data_layer.output.shape[0]
n_batch = data_layer.output.shape[1]
findex = T.cast(self.output_index(), "float32")
findex_bc = findex.reshape((n_time * n_batch,)).dimshuffle(0, 'x')
findex_sum = T.sum(findex)
data = data_layer.output.reshape((n_time * n_batch, data_layer.output.shape[2])) * findex_bc
assert self.z.ndim == 3
z = self.z.reshape((n_time * n_batch, self.z.shape[2])) * findex_bc
data_self_sim = T.flatten(self_similarity(data))
z_self_sim = T.flatten(self_similarity(z))
assert data_self_sim.ndim == z_self_sim.ndim == 1
sim = T.dot(data_self_sim, z_self_sim) # maybe others make sense
assert sim.ndim == 0
# sim is ~ proportional to T * T, so divide by T.
sim *= numpy.float32(input_output_similarity_scale) / findex_sum
self.constraints -= sim
# self.make_output(self.z, collapse = False)
# Note that self.output is going to be overwritten in our derived classes.
self.output = self.make_consensus(self.z) if self.depth > 1 else self.z
self.y_m = None # flat log(self.p_y_given_x)