def learn_posterior(self,
obs_xs,
n_samples,
n_comps,
maxepochs=5000,
lreg=0.01,
minibatch=100,
step=ss.Adam(),
store_sims=False,
logger=sys.stdout,
rng=np.random):
"""
Trains a Bayesian MDN to learn the posterior using the proposal.
"""
# TODO: deal with tuning maxepochs
# create an svi mdn
if self.mdn_prop is None:
self.mdn_post = mdns.MDN_SVI(n_inputs=len(obs_xs),
n_outputs=self.prior.n_dims,
n_hiddens=self.n_hiddens,
act_fun=self.act_fun,
n_components=n_comps,
rng=rng)
else:
self.mdn_post = mdns.replicate_gaussian_mdn(self.mdn_prop,
n_comps,
rng=rng)
logger.write('Learning posterior\n')
# simulate data
ps, xs = self._sim_data(n_samples, store_sims, logger, rng)
# train mdn
self._train_mdn(ps, xs, self.mdn_post, maxepochs, lreg,
min(minibatch, n_samples), step, logger)
try:
# calculate the approximate posterior
self.posterior = self._calc_posterior(self.mdn_post, obs_xs)
except pdfs.gaussian.ImproperCovarianceError:
logger.write(
'WARNING: learning posterior failed due to negative variance.\n'
)
self.posterior = self.proposal
return self.posterior
def learn_proposal(self,
obs_xs,
n_samples,
n_rounds,
maxepochs=1000,
lreg=0.01,
minibatch=50,
step=ss.Adam(),
store_sims=False,
logger=sys.stdout,
rng=np.random):
"""
Iteratively trains an bayesian MDN to learn a gaussian proposal.
"""
# TODO: deal with tuning maxepochs
# create mdn, if haven't already
if self.mdn_prop is None:
self.mdn_prop = mdns.MDN_SVI(n_inputs=len(obs_xs),
n_outputs=self.prior.n_dims,
n_hiddens=self.n_hiddens,
act_fun=self.act_fun,
n_components=1,
rng=rng)
for i in xrange(n_rounds):
logger.write('Learning proposal, round {0}\n'.format(i + 1))
# simulate new batch of data
ps, xs = self._sim_data(n_samples, store_sims, logger, rng)
# train mdn
self._train_mdn(ps, xs, self.mdn_prop, maxepochs, lreg,
min(minibatch, n_samples), step, logger)
try:
# calculate the proposal
self.proposal = self._calc_posterior(
self.mdn_prop, obs_xs).project_to_gaussian()
self.all_proposals.append(self.proposal)
except pdfs.gaussian.ImproperCovarianceError:
logger.write(
'WARNING: learning proposal failed in iteration {0} due to negative variance.\n'
.format(i + 1))
break
return self.proposal
def __init__(self,
model,
trn_data,
trn_loss,
trn_target=None,
val_data=None,
val_loss=None,
val_target=None,
step=ss.Adam(),
max_norm=None):
"""
Constructs and configures the trainer.
:param model: the model to be trained
:param trn_data: training inputs and (possibly) training targets
:param trn_loss: theano variable representing the training loss to minimize
:param trn_target: theano variable representing the training target
:param val_data: validation inputs and (possibly) validation targets
:param val_loss: theano variable representing the validation loss
:param val_target: theano variable representing the validation target
:param step: step size strategy object
:param max_norm: constrain the gradients to have this maximum norm, ignore if None
"""
assert isinstance(
step, ss.StepStrategy), 'step must be a step strategy object'
SGD_Template.__init__(self, model, trn_data, trn_target, val_data,
val_target)
# compile theano function for a single training update
idx = tt.ivector('idx')
grads = tt.grad(trn_loss, model.parms)
grads = [tt.switch(tt.isnan(g), 0., g) for g in grads]
grads = grads if max_norm is None else util.ml.total_norm_constraint(
grads, max_norm)
self.make_update = theano.function(
inputs=[idx],
outputs=trn_loss,
givens=zip(self.trn_inputs, [x[idx] for x in self.trn_data]),
updates=step.updates(model.parms, grads))
if self.do_validation:
# compile theano function for validation
self.validate = theano.function(
inputs=[],
outputs=val_loss,
givens=zip(self.val_inputs, self.val_data) +
self.batch_norm_givens)
def train(model, a):
assert is_data_loaded(), 'Dataset hasn\'t been loaded'
regularizer = lf.WeightDecay(model.parms, weight_decay_rate)
trainer = trainers.SGD(model=model,
trn_data=[data.trn.x],
trn_loss=model.trn_loss + regularizer,
val_data=[data.val.x],
val_loss=model.trn_loss,
step=ss.Adam(a=a))
trainer.train(minibatch=minibatch,
patience=patience,
monitor_every=monitor_every)
def __init__(self,
model,
trn_data,
trn_loss,
trn_target=None,
val_data=None,
val_loss=None,
val_target=None,
step=ss.Adam()):
"""
Constructs and configures the trainer.
:param model: the model to be trained
:param trn_data: train inputs and (possibly) train targets
:param trn_loss: theano variable representing the train loss to minimize
:param trn_target: theano variable representing the train target
:param val_data: validation inputs and (possibly) validation targets
:param val_loss: theano variable representing the validation loss
:param val_target: theano variable representing the validation target
:param step: step size strategy object
:return: None
"""
# parse input
# TODO: it would be good to type check the other inputs too
assert isinstance(
step, ss.StepStrategy), 'Step must be a step strategy object.'
# prepare train data
n_trn_data_list = set([x.shape[0] for x in trn_data])
assert len(
n_trn_data_list) == 1, 'Number of train data is not consistent.'
self.n_trn_data = list(n_trn_data_list)[0]
trn_data = [
theano.shared(x.astype(dtype), borrow=True) for x in trn_data
]
#! privatise this
# compile theano function for a single training update
grads = tt.grad(trn_loss, model.parms)
idx = tt.ivector('idx')
trn_inputs = [model.input
] if trn_target is None else [model.input, trn_target]
self.make_update = theano.function(
inputs=[idx],
outputs=trn_loss,
givens=list(zip(trn_inputs, [x[idx] for x in trn_data])),
updates=step.updates(model.parms, grads))
# private version
# self.make_update = theano.function(
# inputs=[idx],
# outputs=trn_loss,
# givens=list(zip(trn_inputs, [x[idx] for x in trn_data])),
# updates=step.updates(model.parms, grads)
# )
# if model uses batch norm, compile a theano function for setting up stats
if getattr(model, 'batch_norm', False):
batch_norm_givens = [(bn.m, bn.bm)
for bn in model.bns] + [(bn.v, bn.bv)
for bn in model.bns]
self.set_batch_norm_stats = theano.function(
inputs=[],
givens=list(zip(trn_inputs, trn_data)),
updates=[(bn.bm, bn.m)
for bn in model.bns] + [(bn.bv, bn.v)
for bn in model.bns])
else:
self.set_batch_norm_stats = None
batch_norm_givens = []
# if validation data is given, then set up validation too
self.do_validation = val_data is not None
if self.do_validation:
# prepare validation data
n_val_data_list = set([x.shape[0] for x in val_data])
assert len(n_val_data_list
) == 1, 'Number of validation data is not consistent.'
self.n_val_data = list(n_val_data_list)[0]
val_data = [
theano.shared(x.astype(dtype), borrow=True) for x in val_data
]
# compile theano function for validation
val_inputs = [model.input] if val_target is None else [
model.input, val_target
]
self.validate = theano.function(
inputs=[],
outputs=val_loss,
givens=list(zip(val_inputs, val_data)) + batch_norm_givens)
# create checkpointer to store best model
self.checkpointer = ModelCheckpointer(model)
self.best_val_loss = float('inf')
# initialize some variables
self.trn_loss = float('inf')
self.idx_stream = ds.IndexSubSampler(self.n_trn_data)
def __init__(self,
model,
trn_data,
trn_losses,
trn_weights=None,
trn_reg=None,
trn_target=None,
val_data=None,
val_losses=None,
val_weights=None,
val_reg=None,
val_target=None,
step=ss.Adam(),
max_norm=None):
"""
:param model: the model to be trained
:param trn_data: training inputs and (possibly) training targets
:param trn_losses: theano variable representing the training losses at training points
:param trn_weights: weights for training points
:param trn_reg: theano variable representing the training regularizer
:param trn_target: theano variable representing the training target
:param val_data: validation inputs and (possibly) validation targets
:param val_losses: theano variable representing the validation losses at validation points
:param val_weights: weights for validation points
:param val_reg: theano variable representing the validation regularizer
:param val_target: theano variable representing the validation target
:param step: step size strategy object
:param max_norm: constrain the gradients to have this maximum norm, ignore if None
"""
assert isinstance(
step, ss.StepStrategy), 'step must be a step strategy object'
SGD_Template.__init__(self, model, trn_data, trn_target, val_data,
val_target)
# prepare training weights
trn_weights = np.ones(
self.n_trn_data,
dtype=dtype) if trn_weights is None else trn_weights
trn_weights = theano.shared(trn_weights.astype(dtype), borrow=True)
# prepare training regularizer
trn_reg = 0.0 if trn_reg is None else trn_reg
# compile theano function for a single training update
idx = tt.ivector('idx')
trn_loss = tt.mean(trn_weights[idx] * trn_losses) + trn_reg
grads = tt.grad(trn_loss, model.parms)
grads = [tt.switch(tt.isnan(g), 0., g) for g in grads]
grads = grads if max_norm is None else util.ml.total_norm_constraint(
grads, max_norm)
self.make_update = theano.function(
inputs=[idx],
outputs=trn_loss,
givens=zip(self.trn_inputs, [x[idx] for x in self.trn_data]),
updates=step.updates(model.parms, grads))
if self.do_validation:
# prepare validation weights
val_weights = np.ones(
self.n_val_data,
dtype=dtype) if val_weights is None else val_weights
val_weights = theano.shared(val_weights.astype(dtype), borrow=True)
# prepare validation regularizer
val_reg = 0.0 if val_reg is None else val_reg
# compile theano function for validation
self.validate = theano.function(
inputs=[],
outputs=tt.mean(val_weights * val_losses) + val_reg,
givens=zip(self.val_inputs, self.val_data) +
self.batch_norm_givens)
def learn_posterior(self,
obs_xs,
n_samples,
n_rounds,
maxepochs=1000,
minibatch=100,
step=ss.Adam(),
normalize_weights=True,
store_sims=False,
logger=sys.stdout,
rng=np.random):
"""
Sequentially trains an SVI MDN to learn the posterior. Previous posteriors guide simulations.
Simulated data are importance weighted when retraining the model.
"""
# create an svi mdn
if self.mdn is None:
self.mdn = mdns.MDN_SVI(n_inputs=len(obs_xs),
n_outputs=self.prior.n_dims,
n_hiddens=self.n_hiddens,
act_fun=self.act_fun,
n_components=self.n_comps,
rng=rng)
self.regularizer = lf.SviRegularizer(self.mdn.mps, self.mdn.sps,
self.lreg)
for i in xrange(n_rounds):
logger.write('Learning posterior, round {0}\n'.format(i + 1))
# simulate data
logger.write('simulating data... ')
ps, xs = simulators.sim_data(self.posterior.gen,
self.sim_model,
n_samples,
rng=rng)
logger.write('done\n')
# importance weights
if normalize_weights:
log_ws = self.prior.eval(ps) - self.posterior.eval(ps)
ws = n_samples * np.exp(log_ws - logsumexp(log_ws))
else:
ws = np.exp(self.prior.eval(ps) - self.posterior.eval(ps))
if store_sims:
self.all_ps.append(ps)
self.all_xs.append(xs)
self.all_ws.append(ws)
# train model
logger.write('training model...\n')
trainer = trainers.WeightedSGD(model=self.mdn,
trn_data=[xs, ps],
trn_losses=-self.mdn.L,
trn_weights=ws,
trn_reg=self.regularizer /
n_samples,
trn_target=self.mdn.y,
step=step,
max_norm=0.1)
trainer.train(minibatch=minibatch,
maxepochs=maxepochs,
monitor_every=1,
logger=logger)
logger.write('training model done\n')
# update regularizer
m0s = [mp.get_value() for mp in self.mdn.mps]
s0s = [sp.get_value() for sp in self.mdn.sps]
self.regularizer = lf.SviRegularizer_DiagCov(
self.mdn.mps, self.mdn.sps, m0s, s0s)
self.posterior = self.mdn.get_mog(obs_xs)
self.all_posteriors.append(self.posterior)
return self.posterior
def learn_conditional_density(model,
xs,
ys,
ws=None,
regularizer=None,
val_frac=0.05,
step=ss.Adam(a=1.e-4),
minibatch=100,
patience=20,
monitor_every=1,
logger=sys.stdout,
rng=np.random):
"""
Train model to learn the conditional density p(y|x).
"""
xs = np.asarray(xs, np.float32)
ys = np.asarray(ys, np.float32)
n_data = xs.shape[0]
assert ys.shape[0] == n_data, 'wrong sizes'
# shuffle data, so that training and validation sets come from the same distribution
idx = rng.permutation(n_data)
xs = xs[idx]
ys = ys[idx]
# split data into training and validation sets
n_trn = int(n_data - val_frac * n_data)
xs_trn, xs_val = xs[:n_trn], xs[n_trn:]
ys_trn, ys_val = ys[:n_trn], ys[n_trn:]
if ws is None:
# train model without weights
trainer = trainers.SGD(model=model,
trn_data=[xs_trn, ys_trn],
trn_loss=model.trn_loss if regularizer is None
else model.trn_loss + regularizer,
trn_target=model.y,
val_data=[xs_val, ys_val],
val_loss=model.trn_loss,
val_target=model.y,
step=step)
trainer.train(minibatch=minibatch,
patience=patience,
monitor_every=monitor_every,
logger=logger)
else:
# prepare weights
ws = np.asarray(ws, np.float32)
assert ws.size == n_data, 'wrong sizes'
ws = ws[idx]
ws_trn, ws_val = ws[:n_trn], ws[n_trn:]
# train model with weights
trainer = trainers.WeightedSGD(model=model,
trn_data=[xs_trn, ys_trn],
trn_losses=-model.L,
trn_weights=ws_trn,
trn_reg=regularizer,
trn_target=model.y,
val_data=[xs_val, ys_val],
val_losses=-model.L,
val_weights=ws_val,
val_target=model.y,
step=step)
trainer.train(minibatch=minibatch,
patience=patience,
monitor_every=monitor_every,
logger=logger)
return model