def TrainFn(fnsim,
embeddings,
leftop,
rightop,
loss=loss.hinge,
loss_margin=1.0,
op='',
method='SGD',
decay=0.999,
epsilon=1e-6,
max_learning_rate=None,
weight_L1_param_regularizer=None,
weight_L2_param_regularizer=None,
weight_contractive_regularizer_left=None,
weight_contractive_regularizer_right=None):
"""
This function returns a theano function to perform a training iteration, contrasting couples of positive and negative triplets. members are given
as sparse matrices. for one positive triplet there is one negative triplet.
:param fnsim: similarity function (on theano variables).
:param embeddings: an embeddings instance.
:param leftop: class for the 'left' operator.
:param rightop: class for the 'right' operator.
"""
embedding, relationl, relationr = parse_embeddings(embeddings)
# Inputs
inpr, inpl, inpo = S.csr_matrix('inpr'), S.csr_matrix(
'inpl'), S.csr_matrix('inpo')
inpln, inprn, inpon = S.csr_matrix('inpln'), S.csr_matrix(
'inprn'), S.csr_matrix('inpon')
# Learning rates for parameters and embeddings
rate_params = T.scalar('rate_params')
rate_embeddings = T.scalar('rate_embeddings')
# E: D x N, inp: N x B -> <E, inp>: D x B -> <E, inp>.T: B x D
# Positive triplet functions
lhs = S.dot(embedding.E, inpl).T
rhs = S.dot(embedding.E, inpr).T
rell = S.dot(relationl.E, inpo).T
relr = S.dot(relationr.E, inpo).T
# Negative triplet functions
lhsn = S.dot(embedding.E, inpln).T
rhsn = S.dot(embedding.E, inprn).T
relln = S.dot(relationl.E, inpon).T
relrn = S.dot(relationr.E, inpon).T
# Similarity Function, applied to g_lhs and g_rhs
lop, rop = leftop(lhs, rell), rightop(rhs, relr)
lopn, ropn = leftop(lhsn, relln), rightop(rhsn, relrn)
simi = fnsim(lop, rop)
simin = fnsim(lopn, ropn)
supported_loss_args = inspect.getargspec(loss)[0]
loss_args = {} if 'margin' not in supported_loss_args else {
'margin': loss_margin
}
cost, out = loss(simi, simin, **loss_args)
# <EXPERIMENTAL_CODE>
# Should I also plug examples from corrupted triples ?
if weight_contractive_regularizer_left is not None:
cost = cost + (weight_contractive_regularizer_left *
R.contractive_regularizer(lop, lhs))
if weight_contractive_regularizer_right is not None:
cost = cost + (weight_contractive_regularizer_right *
R.contractive_regularizer(rop, rhs))
for rel_param in set([relationl.E, relationr.E]):
if weight_L1_param_regularizer is not None:
cost = cost + (weight_L1_param_regularizer *
R.L1_regularizer(rel_param))
if weight_L2_param_regularizer is not None:
cost = cost + (weight_L2_param_regularizer *
R.L2_regularizer(rel_param))
# </EXPERIMENTAL_CODE>
params = leftop.params + rightop.params + (fnsim.params if hasattr(
fnsim, 'params') else [])
params = list(set(params))
embeds = [embedding.E] + ([relationr.E, relationl.E] if
(type(embeddings) == list) else [])
embeds = list(set(embeds))
# The function updates the implicit function arguments according to the updates.
updates = collections.OrderedDict()
if (method == 'SGD'):
pass # do nothing
elif (method == 'MOMENTUM'):
param_previous_update_map = collections.OrderedDict()
for param in params + embeds:
# Allocate the previous updates
previous_update_value = numpy.zeros(param.get_value().shape,
dtype=theano.config.floatX)
param_previous_update = theano.shared(value=previous_update_value,
name='su_' + param.name)
param_previous_update_map[param] = param_previous_update
elif (method == 'ADAGRAD'):
param_squared_gradients_map = collections.OrderedDict()
for param in params + embeds:
# Allocate the sums of squared gradients
squared_gradients_value = numpy.zeros(param.get_value().shape,
dtype=theano.config.floatX)
param_squared_gradients = theano.shared(
value=squared_gradients_value, name='sg_' + param.name)
param_squared_gradients_map[param] = param_squared_gradients
elif (method == 'ADADELTA'):
param_squared_gradients_map = collections.OrderedDict()
param_squared_updates_map = collections.OrderedDict()
for param in params + embeds:
# Allocate the sums of squared gradients
squared_gradients_value = numpy.zeros(param.get_value().shape,
dtype=theano.config.floatX)
param_squared_gradients = theano.shared(
value=squared_gradients_value, name='sg_' + param.name)
param_squared_gradients_map[param] = param_squared_gradients
# Allocate the sums of squared updates
squared_updates_value = numpy.zeros(param.get_value().shape,
dtype=theano.config.floatX)
param_squared_updates = theano.shared(value=squared_updates_value,
name='su_' + param.name)
param_squared_updates_map[param] = param_squared_updates
elif (method == 'RMSPROP'):
param_squared_gradients_map = collections.OrderedDict()
for param in params + embeds:
# Allocate the sums of squared gradients
squared_gradients_value = numpy.zeros(param.get_value().shape,
dtype=theano.config.floatX)
param_squared_gradients = theano.shared(
value=squared_gradients_value, name='sg_' + param.name)
param_squared_gradients_map[param] = param_squared_gradients
else:
raise ValueError('Unknown method: %s' % (method))
# Parameter Gradients
gradientsparams = T.grad(cost, params)
# Embeddings gradients
gradientsembeds = T.grad(cost, embeds)
# Learning Rates
rates_params = [rate_params for i in range(len(params))]
# In TransE etc. the rate for predicates' embeddings (that do not get normalized) is rate_params, not rate_embeddings
rates_embeddings = [rate_embeddings,
rate_params, rate_params] if len(embeds) > 1 else [
rate_embeddings
] # [rate_embeddings for i in range(len(embeds))]
for param, gradient, rate in zip(params + embeds,
gradientsparams + gradientsembeds,
rates_params + rates_embeddings):
if (method == 'SGD'): # SGD
U.sgd(param, rate, gradient, updates)
elif (method == 'MOMENTUM'): # SGD+MOMENTUM
param_previous_update = param_previous_update_map[param]
U.momentum(param, rate, decay, gradient, updates,
param_previous_update)
elif (method == 'ADAGRAD'): # ADAGRAD
param_squared_gradients = param_squared_gradients_map[param]
U.adagrad(param, rate, epsilon, gradient, updates,
param_squared_gradients)
elif (method == 'ADADELTA'): # ADADELTA
param_squared_gradients = param_squared_gradients_map[param]
param_squared_updates = param_squared_updates_map[param]
U.adadelta(param, rate, decay, epsilon, gradient, updates,
param_squared_gradients, param_squared_updates)
elif (method == 'RMSPROP'): # RMSPROP
param_squared_gradients = param_squared_gradients_map[param]
U.rmsprop(param, rate, decay, max_learning_rate, epsilon, gradient,
updates, param_squared_gradients)
else:
raise ValueError('Unknown method: %s' % (method))
"""
Theano function inputs.
:input rate_embeddings: learning/decay rate for the embeddings.
:input rate_params: learning/decay rate for the parameters.
:input inpl: sparse csr matrix representing the indexes of the positive triplet 'left' member, shape=(#examples,N [Embeddings]).
:input inpr: sparse csr matrix representing the indexes of the positive triplet 'right' member, shape=(#examples,N [Embeddings]).
:input inpo: sparse csr matrix representing the indexes of the positive triplet relation member, shape=(#examples,N [Embeddings]).
:input inpln: sparse csr matrix representing the indexes of the negative triplet 'left' member, shape=(#examples,N [Embeddings]).
:input inprn: sparse csr matrix representing the indexes of the negative triplet 'right' member, shape=(#examples,N [Embeddings]).
:input inpon: sparse csr matrix representing the indexes of the negative triplet relation member, shape=(#examples,N [Embeddings]).
Theano function output.
:output mean(cost): average cost.
:output mean(out): ratio of examples for which the margin is violated,
i.e. for which an update occurs.
"""
return theano.function(
[rate_embeddings, rate_params, inpl, inpr, inpo, inpln, inprn, inpon],
[T.mean(cost), T.mean(out)],
updates=updates,
on_unused_input='ignore')
def TrainFn1Member(fnsim,
embeddings,
leftop,
rightop,
rel=True,
loss=loss.hinge,
loss_margin=1.0,
op=None,
method='SGD',
decay=0.999,
epsilon=1e-6,
max_learning_rate=None,
weight_L1_param_regularizer=None,
weight_L2_param_regularizer=None,
weight_contractive_regularizer_left=None,
weight_contractive_regularizer_right=None):
embedding, relationl, relationr = parse_embeddings(embeddings)
# Inputs
inpr, inpl, inpo = S.csr_matrix('inpr'), S.csr_matrix(
'inpl'), S.csr_matrix('inpo')
inpln, inprn = S.csr_matrix('inpln'), S.csr_matrix('inprn')
# Learning rates for parameters and embeddings
rate_params = T.scalar('rate_params')
rate_embeddings = T.scalar('rate_embeddings')
# Graph
lhs = S.dot(embedding.E, inpl).T
rhs = S.dot(embedding.E, inpr).T
rell = S.dot(relationl.E, inpo).T
relr = S.dot(relationr.E, inpo).T
lhsn = S.dot(embedding.E, inpln).T
rhsn = S.dot(embedding.E, inprn).T
simi = fnsim(leftop(lhs, rell), rightop(rhs, relr))
# Negative 'left' member
similn = fnsim(leftop(lhsn, rell), rightop(rhs, relr))
# Negative 'right' member
simirn = fnsim(leftop(lhs, rell), rightop(rhsn, relr))
costl, outl = loss(simi, similn, margin=loss_margin)
costr, outr = loss(simi, simirn, margin=loss_margin)
cost, out = costl + costr, T.concatenate([outl, outr])
# List of inputs of the function
list_in = [rate_embeddings, rate_params, inpl, inpr, inpo, inpln, inprn]
if rel:
# If rel is True, we also consider a negative relation member
inpon = S.csr_matrix()
relln = S.dot(relationl.E, inpon).T
relrn = S.dot(relationr.E, inpon).T
simion = fnsim(leftop(lhs, relln), rightop(rhs, relrn))
costo, outo = loss(simi, simion, margin=loss_margin)
cost += costo
out = T.concatenate([out, outo])
list_in += [inpon]
# <EXPERIMENTAL_CODE>
# Should I also plug examples from corrupted triples ?
if weight_contractive_regularizer_left is not None:
cost = cost + (weight_contractive_regularizer_left *
R.contractive_regularizer(lop, lhs))
if weight_contractive_regularizer_right is not None:
cost = cost + (weight_contractive_regularizer_right *
R.contractive_regularizer(rop, rhs))
for rel_param in set([relationl.E, relationr.E]):
if weight_L1_param_regularizer is not None:
cost = cost + (weight_L1_param_regularizer *
R.L1_regularizer(rel_param))
if weight_L2_param_regularizer is not None:
cost = cost + (weight_L2_param_regularizer *
R.L2_regularizer(rel_param))
# </EXPERIMENTAL_CODE>
params = leftop.params + rightop.params + (fnsim.params if hasattr(
fnsim, 'params') else [])
embeds = [embedding.E] + ([relationr.E, relationl.E] if
(type(embeddings) == list) else [])
# The function updates the implicit function arguments according to the updates.
updates = collections.OrderedDict()
if (method == 'SGD'):
pass # do nothing
elif (method == 'MOMENTUM'):
param_previous_update_map = collections.OrderedDict()
for param in params + embeds:
# Allocate the previous updates
previous_update_value = numpy.zeros(param.get_value().shape,
dtype=theano.config.floatX)
param_previous_update = theano.shared(value=previous_update_value,
name='su_' + param.name)
param_previous_update_map[param] = param_previous_update
elif (method == 'ADAGRAD'):
param_squared_gradients_map = collections.OrderedDict()
for param in params + embeds:
# Allocate the sums of squared gradients
squared_gradients_value = numpy.zeros(param.get_value().shape,
dtype=theano.config.floatX)
param_squared_gradients = theano.shared(
value=squared_gradients_value, name='sg_' + param.name)
param_squared_gradients_map[param] = param_squared_gradients
elif (method == 'ADADELTA'):
param_squared_gradients_map = collections.OrderedDict()
param_squared_updates_map = collections.OrderedDict()
for param in params + embeds:
# Allocate the sums of squared gradients
squared_gradients_value = numpy.zeros(param.get_value().shape,
dtype=theano.config.floatX)
param_squared_gradients = theano.shared(
value=squared_gradients_value, name='sg_' + param.name)
param_squared_gradients_map[param] = param_squared_gradients
# Allocate the sums of squared updates
squared_updates_value = numpy.zeros(param.get_value().shape,
dtype=theano.config.floatX)
param_squared_updates = theano.shared(value=squared_updates_value,
name='su_' + param.name)
param_squared_updates_map[param] = param_squared_updates
elif (method == 'RMSPROP'):
param_squared_gradients_map = collections.OrderedDict()
for param in params + embeds:
# Allocate the sums of squared gradients
squared_gradients_value = numpy.zeros(param.get_value().shape,
dtype=theano.config.floatX)
param_squared_gradients = theano.shared(
value=squared_gradients_value, name='sg_' + param.name)
param_squared_gradients_map[param] = param_squared_gradients
else:
raise ValueError('Unknown method: %s' % (method))
# Parameter Gradients
gradientsparams = T.grad(cost, params)
# Embeddings gradients
gradientsembeds = T.grad(cost, embeds)
# Learning Rates
rates_params = [rate_params for i in range(len(params))]
# In TransE etc. the rate for predicates' embeddings (that do not get normalized) is rate_params, not rate_embeddings
rates_embeddings = [rate_embeddings,
rate_params, rate_params] if len(embeds) > 1 else [
rate_embeddings
] # [rate_embeddings for i in range(len(embeds))]
for param, gradient, rate in zip(params + embeds,
gradientsparams + gradientsembeds,
rates_params + rates_embeddings):
if (method == 'SGD'): # SGD
U.sgd(param, rate, gradient, updates)
elif (method == 'MOMENTUM'): # SGD+MOMENTUM
param_previous_update = param_previous_update_map[param]
U.momentum(param, rate, decay, gradient, updates,
param_previous_update)
elif (method == 'ADAGRAD'): # ADAGRAD
param_squared_gradients = param_squared_gradients_map[param]
U.adagrad(param, rate, epsilon, gradient, updates,
param_squared_gradients)
elif (method == 'ADADELTA'): # ADADELTA
param_squared_gradients = param_squared_gradients_map[param]
param_squared_updates = param_squared_updates_map[param]
U.adadelta(param, rate, decay, epsilon, gradient, updates,
param_squared_gradients, param_squared_updates)
elif (method == 'RMSPROP'): # RMSPROP
param_squared_gradients = param_squared_gradients_map[param]
U.rmsprop(param, rate, decay, max_learning_rate, epsilon, gradient,
updates, param_squared_gradients)
else:
raise ValueError('Unknown method: %s' % (method))
return theano.function(list_in, [T.mean(cost), T.mean(out)],
updates=updates,
on_unused_input='ignore')