def forward_importance_sampled(galaxy_vae, resid_image, recon_vars, was_on,
seq_tensor, use_importance_sample = True):
assert len(was_on) == len(seq_tensor)
assert resid_image.shape[0] == len(seq_tensor)
assert recon_vars.shape[0] == len(seq_tensor)
# get class weights
class_weights = galaxy_vae.get_pixel_probs(resid_image, recon_vars)
assert_diff(class_weights.sum(1), torch.Tensor([1.0]).to(device), tol = 1e-4)
# get importance sampling weights
if use_importance_sample:
attn_offset = galaxy_vae.attn_offset
prob_off = class_weights.detach()[:, -1].view(-1, 1)
importance_weights = \
get_importance_weights(resid_image.detach(), attn_offset, prob_off)
else:
importance_weights = class_weights.detach()
# sample from importance weights
a_sample = common_utils.sample_class_weights(importance_weights.detach())
a_sample[was_on == 0.] = importance_weights.shape[-1] - 1
recon_mean, recon_var, is_on, kl_z = \
galaxy_vae.sample_conditional_a(\
resid_image, recon_vars, a_sample)
return recon_mean, recon_var, is_on, kl_z, importance_weights, \
class_weights, a_sample
def reinforce_w_double_sample_baseline(\
conditional_loss_fun, log_class_weights,
class_weights_detached, seq_tensor, z_sample,
epoch, data,
grad_estimator_kwargs = None):
# This is what we call REINFORCE+ in our paper,
# where we use a second, independent sample from the discrete distribution
# to use as a baseline
assert len(z_sample) == log_class_weights.shape[0]
# compute loss from those categories
n_classes = log_class_weights.shape[1]
one_hot_z_sample = get_one_hot_encoding_from_int(z_sample, n_classes)
conditional_loss_fun_i = conditional_loss_fun(one_hot_z_sample)
assert len(conditional_loss_fun_i) == log_class_weights.shape[0]
# get log class_weights
log_class_weights_i = log_class_weights[seq_tensor, z_sample]
# get baseline
z_sample2 = sample_class_weights(class_weights_detached)
one_hot_z_sample2 = get_one_hot_encoding_from_int(z_sample2, n_classes)
baseline = conditional_loss_fun(one_hot_z_sample2)
return get_reinforce_grad_sample(conditional_loss_fun_i,
log_class_weights_i,
baseline) + conditional_loss_fun_i
def reinforce_wr(conditional_loss_fun,
log_class_weights,
class_weights_detached,
seq_tensor,
z_sample,
epoch,
data,
n_samples=1,
baseline_constant=None):
# z_sample should be a vector of categories, but is ignored as this function samples itself
# conditional_loss_fun is a function that takes in a one hot encoding
# of z and returns the loss
assert len(z_sample) == log_class_weights.shape[0]
# Sample with replacement
ind = torch.stack([
sample_class_weights(class_weights_detached) for _ in range(n_samples)
], -1)
log_p = log_class_weights.gather(-1, ind)
n_classes = log_class_weights.shape[1]
costs = torch.stack([
conditional_loss_fun(get_one_hot_encoding_from_int(
z_sample, n_classes)) for z_sample in ind.t()
], -1)
if baseline_constant is None:
adv = (costs - costs.mean(-1, keepdim=True)) * n_samples / (n_samples -
1)
else:
adv = costs - baseline_constant
# Add the costs in case there is a direct dependency on the parameters
return (adv.detach() * log_p + costs).mean(-1)
def get_importance_sampled_loss(f_z, log_q,
importance_weights = None,
use_baseline = True):
# class weights from the variational distribution
assert (log_q.detach() < 0).all()
class_weights = torch.exp(log_q.detach())
# why is the tolerance so bad?
assert_diff(class_weights.sum(1), torch.Tensor([1.0]).to(device), tol = 1e-4)
seq_tensor = torch.LongTensor([i for i in range(class_weights.shape[0])])
# sample from conditional distribution
if importance_weights is not None:
assert importance_weights.shape[0] == log_q.shape[0]
assert importance_weights.shape[1] == log_q.shape[1]
# why is the tolerance so bad?
assert_diff(importance_weights.sum(1), torch.Tensor([1.0]).to(device), tol = 1e-4)
# sample from importance weights
z_sample = common_utils.sample_class_weights(importance_weights)
# reweight accordingly
importance_weighting = class_weights[seq_tensor, z_sample] / \
importance_weights[seq_tensor, z_sample]
assert len(importance_weighting) == len(z_sample)
else:
z_sample = common_utils.sample_class_weights(class_weights)
importance_weighting = 1.0
f_z_i_sample = f_z(z_sample)
assert len(f_z_i_sample) == len(z_sample)
log_q_i_sample = log_q[seq_tensor, z_sample]
if use_baseline:
z_sample2 = common_utils.sample_class_weights(class_weights)
baseline = f_z(z_sample2).detach()
else:
baseline = 0.0
reinforce_grad_sample = \
common_utils.get_reinforce_grad_sample(f_z_i_sample, log_q_i_sample, \
baseline) + f_z_i_sample
assert len(reinforce_grad_sample) == len(z_sample)
return (reinforce_grad_sample * importance_weighting).sum()
def get_baseline(conditional_loss_fun, log_class_weights, n_samples,
deterministic):
with torch.no_grad():
n_classes = log_class_weights.size(-1)
if deterministic:
# Compute baseline deterministically as normalized weighted topk elements
weights, ind = log_class_weights.topk(n_samples, -1)
weights = weights - weights.logsumexp(-1,
keepdim=True) # normalize
baseline = (torch.stack([
conditional_loss_fun(
get_one_hot_encoding_from_int(ind[:, i], n_classes))
for i in range(n_samples)
], -1) * weights.exp()).sum(-1)
else:
# Sample with replacement baseline and compute mean
baseline = torch.stack([
conditional_loss_fun(
get_one_hot_encoding_from_int(
sample_class_weights(log_class_weights.exp()),
n_classes)) for i in range(n_samples)
], -1).mean(-1)
return baseline
def get_raoblackwell_ps_loss(
conditional_loss_fun,
log_class_weights,
topk,
grad_estimator,
grad_estimator_kwargs={'grad_estimator_kwargs': None},
epoch=None,
data=None):
"""
Returns a pseudo_loss, such that the gradient obtained by calling
pseudo_loss.backwards() is unbiased for the true loss
Parameters
----------
conditional_loss_fun : function
A function that returns the loss conditional on an instance of the
categorical random variable. It must take in a one-hot-encoding
matrix (batchsize x n_categories) and return a vector of
losses, one for each observation in the batch.
log_class_weights : torch.Tensor
A tensor of shape batchsize x n_categories of the log class weights
topk : Integer
The number of categories to sum over
grad_estimator : function
A function that returns the pseudo loss, that is, the loss which
gives a gradient estimator when .backwards() is called.
See baselines_lib for details.
grad_estimator_kwargs : dict
keyword arguments to gradient estimator
epoch : int
The epoch of the optimizer (for Gumbel-softmax, which has an annealing rate)
data : torch.Tensor
The data at which we evaluate the loss (for NVIl and RELAX, which have
a data dependent baseline)
Returns
-------
ps_loss :
a value such that ps_loss.backward() returns an
estimate of the gradient.
In general, ps_loss might not equal the actual loss.
"""
# class weights from the variational distribution
assert np.all(log_class_weights.detach().cpu().numpy() <= 0)
class_weights = torch.exp(log_class_weights.detach())
# this is the indicator C_k
concentrated_mask, topk_domain, seq_tensor = \
get_concentrated_mask(class_weights, topk)
concentrated_mask = concentrated_mask.float().detach()
############################
# compute the summed term
summed_term = 0.0
for i in range(topk):
# get categories to be summed
summed_indx = topk_domain[:, i]
# compute gradient estimate
grad_summed = \
grad_estimator(conditional_loss_fun, log_class_weights,
class_weights, seq_tensor, \
z_sample = summed_indx,
epoch = epoch,
data = data,
**grad_estimator_kwargs)
# sum
summed_weights = class_weights[seq_tensor, summed_indx].squeeze()
summed_term = summed_term + \
(grad_summed * summed_weights).sum()
############################
# compute sampled term
sampled_weight = torch.sum(class_weights * (1 - concentrated_mask),
dim=1,
keepdim=True)
if not (topk == class_weights.shape[1]):
# if we didn't sum everything
# we sample from the remaining terms
# class weights conditioned on being in the diffuse set
conditional_class_weights = (class_weights + 1e-12) * \
(1 - concentrated_mask) / (sampled_weight + 1e-12)
# sample from conditional distribution
conditional_z_sample = sample_class_weights(conditional_class_weights)
grad_sampled = grad_estimator(conditional_loss_fun,
log_class_weights,
class_weights,
seq_tensor,
z_sample=conditional_z_sample,
epoch=epoch,
data=data,
**grad_estimator_kwargs)
else:
grad_sampled = 0.
return (grad_sampled * sampled_weight.squeeze()).sum() + summed_term
def get_partial_marginal_loss(f_z, log_q, alpha, topk,
use_baseline = True,
use_term_one_baseline = True):
# class weights from the variational distribution
assert np.all(log_q.detach().cpu().numpy() <= 0)
class_weights = torch.exp(log_q.detach())
# assert np.all(np.abs(class_weights.cpu().sum(1).numpy() - 1.0) < 1e-6), \
# np.max(np.abs(class_weights.cpu().sum(1).numpy() - 1.0))
# this is the indicator C_\alpha
concentrated_mask, topk_domain, seq_tensor = \
get_concentrated_mask(class_weights, alpha, topk)
concentrated_mask = concentrated_mask.float().detach()
# the summed term
summed_term = 0.0
for i in range(topk):
summed_indx = topk_domain[:, i]
f_z_i = f_z(summed_indx)
assert len(f_z_i) == log_q.shape[0]
log_q_i = log_q[seq_tensor, summed_indx]
if (use_term_one_baseline) and (use_baseline):
# print('using term 1 baseline')
z_sample2 = common_utils.sample_class_weights(class_weights)
baseline = f_z(z_sample2).detach()
else:
baseline = 0.0
reinforce_grad_sample = \
common_utils.get_reinforce_grad_sample(f_z_i, log_q_i, baseline)
summed_term = summed_term + \
((reinforce_grad_sample + f_z_i) * \
class_weights[seq_tensor, summed_indx].squeeze()).sum()
# sampled term
sampled_weight = torch.sum(class_weights * (1 - concentrated_mask), dim = 1, keepdim = True)
if not(topk == class_weights.shape[1]):
conditional_class_weights = \
class_weights * (1 - concentrated_mask) / (sampled_weight)
conditional_z_sample = common_utils.sample_class_weights(conditional_class_weights)
# print(conditional_z_sample)
# just for my own sanity ...
assert np.all((1 - concentrated_mask)[seq_tensor, conditional_z_sample].cpu().numpy() == 1.), \
'sampled_weight {}'.format(sampled_weight)
f_z_i_sample = f_z(conditional_z_sample)
log_q_i_sample = log_q[seq_tensor, conditional_z_sample]
if use_baseline:
if not use_term_one_baseline:
# print('using alt. covariate')
# sample from the conditional distribution instead
z_sample2 = common_utils.sample_class_weights(conditional_class_weights)
baseline2 = f_z(z_sample2).detach()
else:
z_sample2 = common_utils.sample_class_weights(class_weights)
baseline2 = f_z(z_sample2).detach()
else:
baseline2 = 0.0
sampled_term = common_utils.get_reinforce_grad_sample(f_z_i_sample,
log_q_i_sample, baseline2) + f_z_i_sample
else:
sampled_term = 0.
return (sampled_term * sampled_weight.squeeze()).sum() + summed_term
def get_partial_marginal_loss(f_z,
log_q,
alpha,
topk,
use_baseline=True,
use_term_one_baseline=True):
# class weights from the variational distribution
assert np.all(log_q.detach().cpu().numpy() <= 0)
class_weights = torch.exp(log_q.detach())
# assert np.all(np.abs(class_weights.cpu().sum(1).numpy() - 1.0) < 1e-6), \
# np.max(np.abs(class_weights.cpu().sum(1).numpy() - 1.0))
# this is the indicator C_\alpha
concentrated_mask, topk_domain, seq_tensor = \
get_concentrated_mask(class_weights, alpha, topk)
concentrated_mask = concentrated_mask.float().detach()
# the summed term
summed_term = 0.0
for i in range(topk):
summed_indx = topk_domain[:, i]
f_z_i = f_z(summed_indx)
assert len(f_z_i) == log_q.shape[0]
log_q_i = log_q[seq_tensor, summed_indx]
if (use_term_one_baseline) and (use_baseline):
# print('using term 1 baseline')
z_sample2 = common_utils.sample_class_weights(class_weights)
baseline = f_z(z_sample2).detach()
else:
baseline = 0.0
reinforce_grad_sample = \
common_utils.get_reinforce_grad_sample(f_z_i, log_q_i, baseline)
summed_term = summed_term + \
((reinforce_grad_sample + f_z_i) * \
class_weights[seq_tensor, summed_indx].squeeze()).sum()
# sampled term
sampled_weight = torch.sum(class_weights * (1 - concentrated_mask),
dim=1,
keepdim=True)
if not (topk == class_weights.shape[1]):
conditional_class_weights = \
class_weights * (1 - concentrated_mask) / (sampled_weight)
conditional_z_sample = common_utils.sample_class_weights(
conditional_class_weights)
# print(conditional_z_sample)
# just for my own sanity ...
assert np.all((1 - concentrated_mask)[seq_tensor, conditional_z_sample].cpu().numpy() == 1.), \
'sampled_weight {}'.format(sampled_weight)
f_z_i_sample = f_z(conditional_z_sample)
log_q_i_sample = log_q[seq_tensor, conditional_z_sample]
if use_baseline:
if not use_term_one_baseline:
# print('using alt. covariate')
# sample from the conditional distribution instead
z_sample2 = common_utils.sample_class_weights(
conditional_class_weights)
baseline2 = f_z(z_sample2).detach()
else:
z_sample2 = common_utils.sample_class_weights(class_weights)
baseline2 = f_z(z_sample2).detach()
else:
baseline2 = 0.0
sampled_term = common_utils.get_reinforce_grad_sample(
f_z_i_sample, log_q_i_sample, baseline2) + f_z_i_sample
else:
sampled_term = 0.
return (sampled_term * sampled_weight.squeeze()).sum() + summed_term