class ProjectionLayer(Layer):
"""
This layer can be used to project discrete labels into a continous space
as done in e.g. language models. It takes labels as an input (IndexSpace)
and maps them to their continous embeddings and concatenates them.
Parameters
----------
dim : int
The dimension of the embeddings. Note that this means that the
output dimension is (dim * number of input labels)
layer_name : string
Layer name
irange : numeric
The range of the uniform distribution used to initialize the
embeddings. Can't be used with istdev.
istdev : numeric
The standard deviation of the normal distribution used to
initialize the embeddings. Can't be used with irange.
"""
def __init__(self, dim, layer_name, irange=None, istdev=None):
"""
Initializes a projection layer.
"""
super(ProjectionLayer, self).__init__()
self.dim = dim
self.layer_name = layer_name
if irange is None and istdev is None:
raise ValueError("ProjectionLayer needs either irange or"
"istdev in order to intitalize the projections.")
elif irange is not None and istdev is not None:
raise ValueError("ProjectionLayer was passed both irange and "
"istdev but needs only one")
else:
self._irange = irange
self._istdev = istdev
@wraps(Layer.get_monitoring_channels)
def get_monitoring_channels(self):
W, = self.transformer.get_params()
assert W.ndim == 2
sq_W = T.sqr(W)
row_norms = T.sqrt(sq_W.sum(axis=1))
col_norms = T.sqrt(sq_W.sum(axis=0))
return OrderedDict([('row_norms_min', row_norms.min()),
('row_norms_mean', row_norms.mean()),
('row_norms_max', row_norms.max()),
('col_norms_min', col_norms.min()),
('col_norms_mean', col_norms.mean()),
('col_norms_max', col_norms.max()), ])
@wraps(Layer.set_input_space)
def set_input_space(self, space):
if isinstance(space, IndexSpace):
self.input_dim = space.dim
self.input_space = space
else:
raise ValueError("ProjectionLayer needs an IndexSpace as input")
self.output_space = VectorSpace(self.dim * self.input_dim)
rng = self.mlp.rng
if self._irange is not None:
W = rng.uniform(-self._irange,
self._irange,
(space.max_labels, self.dim))
else:
W = rng.randn(space.max_labels, self.dim) * self._istdev
W = sharedX(W)
W.name = self.layer_name + '_W'
self.transformer = MatrixMul(W)
W, = self.transformer.get_params()
assert W.name is not None
@wraps(Layer.fprop)
def fprop(self, state_below):
z = self.transformer.project(state_below)
return z
@wraps(Layer.get_params)
def get_params(self):
W, = self.transformer.get_params()
assert W.name is not None
params = [W]
return params
class vLBLSoft(Model):
def __init__(self, dict_size, dim, context_length, k, irange = 0.1, seed = 22):
super(vLBLSoft, self).__init__()
rng = np.random.RandomState(seed)
self.rng = rng
self.context_length = context_length
self.dim = dim
self.dict_size = dict_size
C_values = np.asarray(rng.normal(0, math.sqrt(irange),
size=(dim,context_length)),
dtype=theano.config.floatX)
self.C = theano.shared(value=C_values, name='C', borrow=True)
W_context = rng.uniform(-irange, irange, (dict_size, dim))
W_context = sharedX(W_context,name='W_context')
W_target = rng.uniform(-irange, irange, (dict_size, dim))
W_target = sharedX(W_target,name='W_target')
self.projector_context = MatrixMul(W_context)
self.projector_target = MatrixMul(W_target)
self.W_context = W_context
self.W_target = W_target
self.W_target = W_context
b_values = np.asarray(rng.normal(0, math.sqrt(irange), size=(dict_size,)),
dtype=theano.config.floatX)
self.b = theano.shared(value=b_values, name='b', borrow=True)
self.input_space = IndexSpace(dim = context_length, max_labels = dict_size)
self.output_space = IndexSpace(dim = 1, max_labels = dict_size)
self.allY = T.as_tensor_variable(np.arange(dict_size,dtype=np.int64).reshape(dict_size,1))
def get_params(self):
#get W from projector
rval1 = self.projector_context.get_params()
rval2 = self.projector_target.get_params()
#add C, b
rval1.extend([self.C, self.b])
rval1.extend(rval2)
return rval1
def fprop(self, state_below):
"""
state_below is r_w?
"""
state_below = state_below.reshape((state_below.shape[0], self.dim, self.context_length))
rval = self.C.dimshuffle('x', 0, 1) * state_below
rval = rval.sum(axis=2)
return rval
def get_default_cost(self):
return Default()
def get_monitoring_data_specs(self):
"""
Returns data specs requiring both inputs and targets."""
space = CompositeSpace((self.get_input_space(),
self.get_output_space()))
source = (self.get_input_source(), self.get_target_source())
return (space, source)
def get_monitoring_channels(self, data):
rval = OrderedDict()
rval['nll'] = self.cost_from_X(data)
rval['perplexity'] = 10 ** (rval['nll']/np.log(10).astype('float32'))
return rval
def score(self, all_q_w, q_h):
sallwh = T.dot(q_h,all_q_w.T) + self.b.dimshuffle('x',0)
return sallwh
def cost_from_X(self, data):
X, Y = data
X = self.projector_context.project(X)
q_h = self.fprop(X)
rval = self.cost(Y,q_h)
return rval
def cost(self,Y,q_h):
all_q_w = self.W_target
s = self.score(all_q_w,q_h)
p_w_given_h = T.nnet.softmax(s)
return T.cast(-T.mean(T.diag(T.log(p_w_given_h)[T.arange(Y.shape[0]), Y])),theano.config.floatX)
def apply_dropout(self, state, include_prob, scale, theano_rng, input_space, mask_value=0, per_example=True):
"""
per_example : bool, optional
Sample a different mask value for every example in a batch.
Defaults to `True`. If `False`, sample one mask per mini-batch.
"""
if include_prob in [None, 1.0, 1]:
return state
assert scale is not None
if isinstance(state, tuple):
return tuple(self.apply_dropout(substate, include_prob,
scale, theano_rng, mask_value)
for substate in state)
if per_example:
mask = theano_rng.binomial(p=include_prob, size=state.shape,
dtype=state.dtype)
else:
batch = input_space.get_origin_batch(1)
mask = theano_rng.binomial(p=include_prob, size=batch.shape,
dtype=state.dtype)
rebroadcast = T.Rebroadcast(*zip(xrange(batch.ndim),
[s == 1 for s in batch.shape]))
mask = rebroadcast(mask)
if mask_value == 0:
rval = state * mask * scale
else:
rval = T.switch(mask, state * scale, mask_value)
return T.cast(rval, state.dtype)
def dropout_fprop(self, state_below, default_input_include_prob=0.5,
input_include_probs=None, default_input_scale=2.,
input_scales=None, per_example=True):
if input_include_probs is None:
input_include_probs = {}
if input_scales is None:
input_scales = {}
theano_rng = MRG_RandomStreams(max(self.rng.randint(2 ** 15), 1))
include_prob = default_input_include_prob
scale = default_input_scale
state_below = self.apply_dropout(
state=state_below,
include_prob=include_prob,
theano_rng=theano_rng,
scale=scale,
#check
mask_value=0,
input_space=self.get_input_space(),
per_example=per_example
)
state_below = self.fprop(state_below)
return state_below
class ProjectionLayer(Layer):
"""
This layer can be used to project discrete labels into a continous space
as done in e.g. language models. It takes labels as an input (IndexSpace)
and maps them to their continous embeddings and concatenates them.
Parameters
----------
dim : int
The dimension of the embeddings. Note that this means that the
output dimension is (dim * number of input labels)
layer_name : string
Layer name
irange : numeric
The range of the uniform distribution used to initialize the
embeddings. Can't be used with istdev.
istdev : numeric
The standard deviation of the normal distribution used to
initialize the embeddings. Can't be used with irange.
"""
def __init__(self, dim, layer_name, irange=None, istdev=None):
"""
Initializes a projection layer.
"""
super(ProjectionLayer, self).__init__()
self.dim = dim
self.layer_name = layer_name
if irange is None and istdev is None:
raise ValueError("ProjectionLayer needs either irange or"
"istdev in order to intitalize the projections.")
elif irange is not None and istdev is not None:
raise ValueError("ProjectionLayer was passed both irange and "
"istdev but needs only one")
else:
self._irange = irange
self._istdev = istdev
@wraps(Layer.get_layer_monitoring_channels)
def get_layer_monitoring_channels(self, *args, **kwargs):
W, = self.transformer.get_params()
assert W.ndim == 2
sq_W = T.sqr(W)
row_norms = T.sqrt(sq_W.sum(axis=1))
col_norms = T.sqrt(sq_W.sum(axis=0))
return OrderedDict([
('row_norms_min', row_norms.min()),
('row_norms_mean', row_norms.mean()),
('row_norms_max', row_norms.max()),
('col_norms_min', col_norms.min()),
('col_norms_mean', col_norms.mean()),
('col_norms_max', col_norms.max()),
])
def _check_input_space_and_get_max_labels(self, space):
if isinstance(space, IndexSpace):
return space.max_labels
if isinstance(space, CompositeSpace):
ml = []
for c in space.components:
ml.append(self._check_input_space_and_get_max_labels(c))
# check that all of them are equal
if len(set(ml)) != 1:
raise ValueError("Composite space is empty or containing "
"incompatible index spaces")
return ml[0]
raise ValueError("ProjectionLayer needs an IndexSpace or a "
"CompositeSpace of them as input")
def _build_output_space(self, space):
if isinstance(space, IndexSpace):
return VectorSpace(self.dim * space.dim)
if isinstance(space, CompositeSpace):
return CompositeSpace(
[self._build_output_space(c) for c in space.components])
assert False
@wraps(Layer.set_input_space)
def set_input_space(self, space):
max_labels = self._check_input_space_and_get_max_labels(space)
self.input_space = space
self.output_space = self._build_output_space(space)
rng = self.mlp.rng
if self._irange is not None:
W = rng.uniform(-self._irange, self._irange,
(max_labels, self.dim))
else:
W = rng.randn(max_labels, self.dim) * self._istdev
W = sharedX(W)
W.name = self.layer_name + '_W'
self.transformer = MatrixMul(W)
W, = self.transformer.get_params()
assert W.name is not None
def _fprop_recursive(self, state_below):
if isinstance(state_below, tuple):
return tuple(self._fprop_recursive(s) for s in state_below)
return self.transformer.project(state_below)
@wraps(Layer.fprop)
def fprop(self, state_below):
return self._fprop_recursive(state_below)
@wraps(Layer.get_params)
def get_params(self):
W, = self.transformer.get_params()
assert W.name is not None
params = [W]
return params
@wraps(Layer.get_weight_decay)
def get_weight_decay(self, coeff):
if isinstance(coeff, str):
coeff = float(coeff)
assert isinstance(coeff, float) or hasattr(coeff, 'dtype')
W, = self.transformer.get_params()
return coeff * T.sqr(W).sum()
@wraps(Layer.get_l1_weight_decay)
def get_l1_weight_decay(self, coeff):
if isinstance(coeff, str):
coeff = float(coeff)
assert isinstance(coeff, float) or hasattr(coeff, 'dtype')
W, = self.transformer.get_params()
return coeff * abs(W).sum()
class vLBL(Model):
def __init__(self, dict_size, dim, context_length, k, irange = 0.1, seed = 22, max_row_norm=None,max_col_norm=None):
rng = np.random.RandomState(seed)
self.rng = rng
self.k = k
self.context_length = context_length
self.dim = dim
self.dict_size = dict_size
C = rng.randn(dim, context_length)
self.C = sharedX(C)
W = rng.uniform(-irange, irange, (dict_size, dim))
W = sharedX(W)
self.W=W
# TODO maybe have another projector for tagets
self.projector = MatrixMul(W)
self.b = sharedX(np.zeros((dict_size,)), name = 'vLBL_b')
self.input_space = IndexSpace(dim = context_length, max_labels = dict_size)
#self.output_space = IndexSpace(dim = 1, max_labels = dict_size)
self.output_space = VectorSpace(dim = dict_size)
self.max_row_norm=max_row_norm
self.max_col_norm=max_col_norm
def get_params(self):
rval = self.projector.get_params()
rval.extend([self.C, self.b])
return rval
def get_default_cost(self):
return Default()
def fprop(self, state_below):
"q^(h) from EQ. 2"
state_below = state_below.reshape((state_below.shape[0], self.dim, self.context_length))
rval = self.C.dimshuffle('x', 0, 1) * state_below
rval = rval.sum(axis=2)
return rval
def _modify_updates(self,updates):
#for param in self.get_params():
if self.max_row_norm is not None:
W = self.W
if W in updates:
updated_W = updates[W]
row_norms = T.sqrt(T.sum(T.sqr(updated_W), axis=1))
desired_norms = T.clip(row_norms, 0, self.max_row_norm)
scales = desired_norms / (1e-7 + row_norms)
updates[W] = updated_W * scales.dimshuffle(0, 'x')
if self.max_col_norm is not None:
assert self.max_row_norm is None
W = self.W
if W in updates:
updated_W = updates[W]
col_norms = T.sqrt(T.sum(T.sqr(updated_W), axis=0))
desired_norms = T.clip(col_norms, 0, self.max_col_norm)
updates[W] = updated_W * (desired_norms / (1e-7 + col_norms))
def score(self,q_h):
q_w = self.projector._W
rval = T.dot(q_h, q_w.T) + self.b.dimshuffle('x', 0)
return rval
def cost(self,Y,q_h):
z = self.score(q_h)
z = z - z.max(axis=1).dimshuffle(0, 'x')
log_prob = z - T.log(T.exp(z).sum(axis=1).dimshuffle(0, 'x'))
log_prob_of = (Y * log_prob).sum(axis=1)
assert log_prob_of.ndim == 1
rval = as_floatX(log_prob_of.mean())
return - rval
def cost_from_X(self, data):
X, Y = data
X = self.projector.project(X)
q_h = self.fprop(X)
return self.cost(Y,q_h)
def get_monitoring_data_specs(self):
space = CompositeSpace((self.get_input_space(),
self.get_output_space()))
source = (self.get_input_source(), self.get_target_source())
return (space, source)
def get_monitoring_channels(self, data):
X, Y = data
rval = OrderedDict()
W_context = self.W
W_target = self.W
b = self.b
C = self.C
sq_W_context = T.sqr(W_context)
# sq_W_target = T.sqr(W_target)
sq_b = T.sqr(b)
sq_c = T.sqr(C)
row_norms_W_context = T.sqrt(sq_W_context.sum(axis=1))
col_norms_W_context = T.sqrt(sq_W_context.sum(axis=0))
# row_norms_W_target = T.sqrt(sq_W_target.sum(axis=1))
# col_norms_W_target = T.sqrt(sq_W_target.sum(axis=0))
col_norms_b = T.sqrt(sq_b.sum(axis=0))
col_norms_c = T.sqrt(sq_c.sum(axis=0))
rval = OrderedDict([
('W_context_row_norms_min' , row_norms_W_context.min()),
('W_context_row_norms_mean' , row_norms_W_context.mean()),
('W_context_row_norms_max' , row_norms_W_context.max()),
('W_context_col_norms_min' , col_norms_W_context.min()),
('W_context_col_norms_mean' , col_norms_W_context.mean()),
('W_context_col_norms_max' , col_norms_W_context.max()),
# ('W_target_row_norms_min' , row_norms_W_target.min()),
# ('W_target_row_norms_mean' , row_norms_W_target.mean()),
# ('W_target_row_norms_max' , row_norms_W_target.max()),
# ('W_target_col_norms_min' , col_norms_W_target.min()),
# ('W_target_col_norms_mean' , col_norms_W_target.mean()),
# ('W_target_col_norms_max' , col_norms_W_target.max()),
('b_col_norms_min' , col_norms_b.min()),
('b_col_norms_mean' , col_norms_b.mean()),
('b_col_norms_max' , col_norms_b.max()),
('c_col_norms_min' , col_norms_c.min()),
('c_col_norms_mean' , col_norms_c.mean()),
('c_col_norms_max' , col_norms_c.max()),
])
nll = self.cost_from_X(data)
rval['perplexity'] = as_floatX(10 ** (nll/np.log(10)))
return rval
def apply_dropout(self, state, include_prob, scale, theano_rng, input_space, mask_value=0, per_example=True):
"""
per_example : bool, optional
Sample a different mask value for every example in a batch.
Defaults to `True`. If `False`, sample one mask per mini-batch.
"""
if include_prob in [None, 1.0, 1]:
return state
assert scale is not None
if isinstance(state, tuple):
return tuple(self.apply_dropout(substate, include_prob,
scale, theano_rng, mask_value)
for substate in state)
if per_example:
mask = theano_rng.binomial(p=include_prob, size=state.shape,
dtype=state.dtype)
else:
batch = input_space.get_origin_batch(1)
mask = theano_rng.binomial(p=include_prob, size=batch.shape,
dtype=state.dtype)
rebroadcast = T.Rebroadcast(*zip(xrange(batch.ndim),
[s == 1 for s in batch.shape]))
mask = rebroadcast(mask)
if mask_value == 0:
rval = state * mask * scale
else:
rval = T.switch(mask, state * scale, mask_value)
return T.cast(rval, state.dtype)
def dropout_fprop(self, state_below, default_input_include_prob=0.5,
input_include_probs=None, default_input_scale=2.,
input_scales=None, per_example=True):
if input_include_probs is None:
input_include_probs = {}
if input_scales is None:
input_scales = {}
theano_rng = MRG_RandomStreams(max(self.rng.randint(2 ** 15), 1))
include_prob = default_input_include_prob
scale = default_input_scale
state_below = self.apply_dropout(
state=state_below,
include_prob=include_prob,
theano_rng=theano_rng,
scale=scale,
#check
mask_value=0,
input_space=self.get_input_space(),
per_example=per_example
)
state_below = self.fprop(state_below)
return state_below
class vLBL(Model):
def __init__(self, dict_size, dim, context_length, k, irange = 0.1, seed = 22):
rng = np.random.RandomState(seed)
self.rng = rng
self.k = k
self.context_length = context_length
self.dim = dim
self.dict_size = dict_size
C = rng.randn(dim, context_length)
self.C = sharedX(C)
W = rng.uniform(-irange, irange, (dict_size, dim))
W = sharedX(W)
# TODO maybe have another projector for tagets
self.projector = MatrixMul(W)
self.b = sharedX(np.zeros((dict_size,)), name = 'vLBL_b')
self.set_spaces()
self.rng = np.random.RandomState(2014)
def set_spaces(self):
self.input_space = IndexSpace(dim=self.context_length, max_labels=self.dict_size)
self.output_space = VectorSpace(dim=self.dict_size)
def get_params(self):
rval = self.projector.get_params()
rval.extend([self.C, self.b])
return rval
def context(self, state_below):
"q^(h) from EQ. 2"
state_below = state_below.reshape((state_below.shape[0], self.dim, self.context_length))
rval = self.C.dimshuffle('x', 0, 1) * state_below
rval = rval.sum(axis=2)
return rval
def score(self, X, Y=None):
X = self.projector.project(X)
q_h = self.context(X)
# this is used during training
if Y is not None:
q_w = self.projector.project(Y).reshape((Y.shape[0], self.dim))
rval = (q_w * q_h).sum(axis=1) + self.b[Y].flatten()
# during nll
else:
q_w = self.projector._W
rval = T.dot(q_h, q_w.T) + self.b.dimshuffle('x', 0)
return rval
def cost_from_X(self, data):
X, Y = data
z = self.score(X)
z = z - z.max(axis=1).dimshuffle(0, 'x')
log_prob = z - T.log(T.exp(z).sum(axis=1).dimshuffle(0, 'x'))
log_prob_of = (Y * log_prob).sum(axis=1)
assert log_prob_of.ndim == 1
rval = as_floatX(log_prob_of.mean())
return - rval
def get_monitoring_data_specs(self):
space = CompositeSpace((self.get_input_space(),
self.get_output_space()))
source = (self.get_input_source(), self.get_target_source())
return (space, source)
def get_monitoring_channels(self, data):
X, Y = data
rval = OrderedDict()
nll = self.cost_from_X(data)
rval['perplexity'] = as_floatX(10 ** (nll/np.log(10)))
return rval
class ProjectionLayer(Layer):
"""
This layer can be used to project discrete labels into a continous space
as done in e.g. language models. It takes labels as an input (IndexSpace)
and maps them to their continous embeddings and concatenates them.
Parameters
----------
dim : int
The dimension of the embeddings. Note that this means that the
output dimension is (dim * number of input labels)
layer_name : string
Layer name
irange : numeric
The range of the uniform distribution used to initialize the
embeddings. Can't be used with istdev.
istdev : numeric
The standard deviation of the normal distribution used to
initialize the embeddings. Can't be used with irange.
"""
def __init__(self, dim, layer_name, irange=None, istdev=None):
"""
Initializes a projection layer.
"""
super(ProjectionLayer, self).__init__()
self.dim = dim
self.layer_name = layer_name
if irange is None and istdev is None:
raise ValueError("ProjectionLayer needs either irange or"
"istdev in order to intitalize the projections.")
elif irange is not None and istdev is not None:
raise ValueError("ProjectionLayer was passed both irange and "
"istdev but needs only one")
else:
self._irange = irange
self._istdev = istdev
@wraps(Layer.get_monitoring_channels)
def get_monitoring_channels(self):
W, = self.transformer.get_params()
assert W.ndim == 2
sq_W = T.sqr(W)
row_norms = T.sqrt(sq_W.sum(axis=1))
col_norms = T.sqrt(sq_W.sum(axis=0))
return OrderedDict([
('row_norms_min', row_norms.min()),
('row_norms_mean', row_norms.mean()),
('row_norms_max', row_norms.max()),
('col_norms_min', col_norms.min()),
('col_norms_mean', col_norms.mean()),
('col_norms_max', col_norms.max()),
])
@wraps(Layer.set_input_space)
def set_input_space(self, space):
if isinstance(space, IndexSpace):
self.input_dim = space.dim
self.input_space = space
else:
raise ValueError("ProjectionLayer needs an IndexSpace as input")
self.output_space = VectorSpace(self.dim * self.input_dim)
rng = self.mlp.rng
if self._irange is not None:
W = rng.uniform(-self._irange, self._irange,
(space.max_labels, self.dim))
else:
W = rng.randn(space.max_labels, self.dim) * self._istdev
W = sharedX(W)
W.name = self.layer_name + '_W'
self.transformer = MatrixMul(W)
W, = self.transformer.get_params()
assert W.name is not None
@wraps(Layer.fprop)
def fprop(self, state_below):
z = self.transformer.project(state_below)
return z
@wraps(Layer.get_params)
def get_params(self):
W, = self.transformer.get_params()
assert W.name is not None
params = [W]
return params
class vLBLSoft(Model):
def __init__(self, dict_size, dim, context_length, k, irange=0.1, seed=22):
super(vLBLSoft, self).__init__()
rng = np.random.RandomState(seed)
self.rng = rng
self.context_length = context_length
self.dim = dim
self.dict_size = dict_size
C_values = np.asarray(rng.normal(0, math.sqrt(irange), size=(dim, context_length)), dtype=theano.config.floatX)
self.C = theano.shared(value=C_values, name="C", borrow=True)
W_context = rng.uniform(-irange, irange, (dict_size, dim))
W_context = sharedX(W_context, name="W_context")
W_target = rng.uniform(-irange, irange, (dict_size, dim))
W_target = sharedX(W_target, name="W_target")
self.projector_context = MatrixMul(W_context)
self.projector_target = MatrixMul(W_target)
self.W_context = W_context
self.W_target = W_target
self.W_target = W_context
b_values = np.asarray(rng.normal(0, math.sqrt(irange), size=(dict_size,)), dtype=theano.config.floatX)
self.b = theano.shared(value=b_values, name="b", borrow=True)
self.input_space = IndexSpace(dim=context_length, max_labels=dict_size)
self.output_space = IndexSpace(dim=1, max_labels=dict_size)
self.allY = T.as_tensor_variable(np.arange(dict_size, dtype=np.int64).reshape(dict_size, 1))
def get_params(self):
# get W from projector
rval1 = self.projector_context.get_params()
rval2 = self.projector_target.get_params()
# add C, b
rval1.extend([self.C, self.b])
rval1.extend(rval2)
return rval1
def fprop(self, state_below):
"""
state_below is r_w?
"""
state_below = state_below.reshape((state_below.shape[0], self.dim, self.context_length))
rval = self.C.dimshuffle("x", 0, 1) * state_below
rval = rval.sum(axis=2)
return rval
def get_default_cost(self):
return Default()
def get_monitoring_data_specs(self):
"""
Returns data specs requiring both inputs and targets.
Returns
-------
data_specs: TODO
The data specifications for both inputs and targets.
"""
space = CompositeSpace((self.get_input_space(), self.get_output_space()))
source = (self.get_input_source(), self.get_target_source())
return (space, source)
def get_monitoring_channels(self, data):
rval = OrderedDict()
# W_context = self.W_context
# W_target = self.W_target
# b = self.b
# C = self.C
# sq_W_context = T.sqr(W_context)
# sq_W_target = T.sqr(W_target)
# sq_b = T.sqr(b)
# sq_c = T.sqr(C)
# row_norms_W_context = T.sqrt(sq_W_context.sum(axis=1))
# col_norms_W_context = T.sqrt(sq_W_context.sum(axis=0))
# row_norms_W_target = T.sqrt(sq_W_target.sum(axis=1))
# col_norms_W_target = T.sqrt(sq_W_target.sum(axis=0))
# col_norms_b = T.sqrt(sq_b.sum(axis=0))
# col_norms_c = T.sqrt(sq_c.sum(axis=0))
# rval = OrderedDict([
# ('W_context_row_norms_min' , row_norms_W_context.min()),
# ('W_context_row_norms_mean' , row_norms_W_context.mean()),
# ('W_context_row_norms_max' , row_norms_W_context.max()),
# ('W_context_col_norms_min' , col_norms_W_context.min()),
# ('W_context_col_norms_mean' , col_norms_W_context.mean()),
# ('W_context_col_norms_max' , col_norms_W_context.max()),
# ('W_target_row_norms_min' , row_norms_W_target.min()),
# ('W_target_row_norms_mean' , row_norms_W_target.mean()),
# ('W_target_row_norms_max' , row_norms_W_target.max()),
# ('W_target_col_norms_min' , col_norms_W_target.min()),
# ('W_target_col_norms_mean' , col_norms_W_target.mean()),
# ('W_target_col_norms_max' , col_norms_W_target.max()),
# ('b_col_norms_min' , col_norms_b.min()),
# ('b_col_norms_mean' , col_norms_b.mean()),
# ('b_col_norms_max' , col_norms_b.max()),
# ('c_col_norms_min' , col_norms_c.min()),
# ('c_col_norms_mean' , col_norms_c.mean()),
# ('c_col_norms_max' , col_norms_c.max()),
# ])
rval["nll"] = self.cost_from_X(data)
rval["perplexity"] = 10 ** (rval["nll"] / np.log(10).astype("float32"))
return rval
def score(self, all_q_w, q_h):
sallwh = T.dot(q_h, all_q_w.T) + self.b.dimshuffle("x", 0)
return sallwh
def cost_from_X(self, data):
X, Y = data
X = self.projector_context.project(X)
q_h = self.fprop(X)
rval = self.cost(Y, q_h)
return rval
def cost(self, Y, q_h):
all_q_w = self.projector_target.project(self.allY)
s = self.score(all_q_w, q_h)
p_w_given_h = T.nnet.softmax(s)
return T.cast(-T.mean(T.diag(T.log(p_w_given_h)[T.arange(Y.shape[0]), Y])), theano.config.floatX)
def apply_dropout(self, state, include_prob, scale, theano_rng, input_space, mask_value=0, per_example=True):
"""
Parameters
----------
state: WRITEME
include_prob : WRITEME
scale : WRITEME
theano_rng : WRITEME
input_space : WRITEME
mask_value : WRITEME
per_example : bool, optional
Sample a different mask value for every example in a batch.
Defaults to `True`. If `False`, sample one mask per mini-batch.
"""
if include_prob in [None, 1.0, 1]:
return state
assert scale is not None
if isinstance(state, tuple):
return tuple(
self.apply_dropout(substate, include_prob, scale, theano_rng, mask_value) for substate in state
)
# TODO: all of this assumes that if it's not a tuple, it's
# a dense tensor. It hasn't been tested with sparse types.
# A method to format the mask (or any other values) as
# the given symbolic type should be added to the Spaces
# interface.
if per_example:
mask = theano_rng.binomial(p=include_prob, size=state.shape, dtype=state.dtype)
else:
batch = input_space.get_origin_batch(1)
mask = theano_rng.binomial(p=include_prob, size=batch.shape, dtype=state.dtype)
rebroadcast = T.Rebroadcast(*zip(xrange(batch.ndim), [s == 1 for s in batch.shape]))
mask = rebroadcast(mask)
if mask_value == 0:
rval = state * mask * scale
else:
rval = T.switch(mask, state * scale, mask_value)
return T.cast(rval, state.dtype)
def dropout_fprop(
self,
state_below,
default_input_include_prob=0.5,
input_include_probs=None,
default_input_scale=2.0,
input_scales=None,
per_example=True,
):
"""
Returns the output of the MLP, when applying dropout to the input and
intermediate layers.
Parameters
----------
state_below : WRITEME
The input to the MLP
default_input_include_prob : WRITEME
input_include_probs : WRITEME
default_input_scale : WRITEME
input_scales : WRITEME
per_example : bool, optional
Sample a different mask value for every example in a batch.
Defaults to `True`. If `False`, sample one mask per mini-batch.
Notes
-----
Each input to each layer is randomly included or
excluded for each example. The probability of inclusion is independent
for each input and each example. Each layer uses
`default_input_include_prob` unless that layer's name appears as a key
in input_include_probs, in which case the input inclusion probability
is given by the corresponding value.
Each feature is also multiplied by a scale factor. The scale factor for
each layer's input scale is determined by the same scheme as the input
probabilities.
"""
warnings.warn(
"dropout doesn't use fixed_var_descr so it won't work "
"with algorithms that make more than one theano "
"function call per batch, such as BGD. Implementing "
"fixed_var descr could increase the memory usage "
"though."
)
if input_include_probs is None:
input_include_probs = {}
if input_scales is None:
input_scales = {}
# self._validate_layer_names(list(input_include_probs.keys()))
# self._validate_layer_names(list(input_scales.keys()))
theano_rng = MRG_RandomStreams(max(self.rng.randint(2 ** 15), 1))
# for layer in self.layers:
# layer_name = layer.layer_name
# if layer_name in input_include_probs:
# include_prob = input_include_probs[layer_name]
# else:
include_prob = default_input_include_prob
# if layer_name in input_scales:
# scale = input_scales[layer_name]
# else:
scale = default_input_scale
state_below = self.apply_dropout(
state=state_below,
include_prob=include_prob,
theano_rng=theano_rng,
scale=scale,
# check
mask_value=0,
input_space=self.get_input_space(),
per_example=per_example,
)
state_below = self.fprop(state_below)
return state_below
class ProjectionLayer(Layer):
"""
This layer can be used to project discrete labels into a continous space
as done in e.g. language models. It takes labels as an input (IndexSpace)
and maps them to their continous embeddings and concatenates them.
Parameters
----------
dim : int
The dimension of the embeddings. Note that this means that the
output dimension is (dim * number of input labels)
layer_name : string
Layer name
irange : numeric
The range of the uniform distribution used to initialize the
embeddings. Can't be used with istdev.
istdev : numeric
The standard deviation of the normal distribution used to
initialize the embeddings. Can't be used with irange.
"""
def __init__(self, dim, layer_name, irange=None, istdev=None):
"""
Initializes a projection layer.
"""
super(ProjectionLayer, self).__init__()
self.dim = dim
self.layer_name = layer_name
if irange is None and istdev is None:
raise ValueError("ProjectionLayer needs either irange or"
"istdev in order to intitalize the projections.")
elif irange is not None and istdev is not None:
raise ValueError("ProjectionLayer was passed both irange and "
"istdev but needs only one")
else:
self._irange = irange
self._istdev = istdev
@wraps(Layer.get_layer_monitoring_channels)
def get_layer_monitoring_channels(self, *args, **kwargs):
W, = self.transformer.get_params()
assert W.ndim == 2
sq_W = T.sqr(W)
row_norms = T.sqrt(sq_W.sum(axis=1))
col_norms = T.sqrt(sq_W.sum(axis=0))
return OrderedDict([('row_norms_min', row_norms.min()),
('row_norms_mean', row_norms.mean()),
('row_norms_max', row_norms.max()),
('col_norms_min', col_norms.min()),
('col_norms_mean', col_norms.mean()),
('col_norms_max', col_norms.max()), ])
def _check_input_space_and_get_max_labels(self, space):
if isinstance(space, IndexSpace):
return space.max_labels
if isinstance(space, CompositeSpace):
ml = []
for c in space.components:
ml.append(self._check_input_space_and_get_max_labels(c))
# check that all of them are equal
if len(set(ml)) != 1:
raise ValueError("Composite space is empty or containing "
"incompatible index spaces")
return ml[0]
raise ValueError("ProjectionLayer needs an IndexSpace or a "
"CompositeSpace of them as input")
def _build_output_space(self, space):
if isinstance(space, IndexSpace):
return VectorSpace(self.dim * space.dim)
if isinstance(space, CompositeSpace):
return CompositeSpace([self._build_output_space(c)
for c in space.components])
assert False
@wraps(Layer.set_input_space)
def set_input_space(self, space):
max_labels = self._check_input_space_and_get_max_labels(space)
self.input_space = space
self.output_space = self._build_output_space(space)
rng = self.mlp.rng
if self._irange is not None:
W = rng.uniform(-self._irange,
self._irange,
(max_labels, self.dim))
else:
W = rng.randn(max_labels, self.dim) * self._istdev
W = sharedX(W)
W.name = self.layer_name + '_W'
self.transformer = MatrixMul(W)
W, = self.transformer.get_params()
assert W.name is not None
def _fprop_recursive(self, state_below):
if isinstance(state_below, tuple):
return tuple(self._fprop_recursive(s) for s in state_below)
return self.transformer.project(state_below)
@wraps(Layer.fprop)
def fprop(self, state_below):
return self._fprop_recursive(state_below)
@wraps(Layer.get_params)
def get_params(self):
W, = self.transformer.get_params()
assert W.name is not None
params = [W]
return params
@wraps(Layer.get_weight_decay)
def get_weight_decay(self, coeff):
if isinstance(coeff, str):
coeff = float(coeff)
assert isinstance(coeff, float) or hasattr(coeff, 'dtype')
W, = self.transformer.get_params()
return coeff * T.sqr(W).sum()
@wraps(Layer.get_l1_weight_decay)
def get_l1_weight_decay(self, coeff):
if isinstance(coeff, str):
coeff = float(coeff)
assert isinstance(coeff, float) or hasattr(coeff, 'dtype')
W, = self.transformer.get_params()
return coeff * abs(W).sum()