def divergence_loss(self, posterior, prior):
if isinstance(posterior, DistributionList) and isinstance(
prior, DistributionList):
result = 0.0
scalars = {}
for idx, (s, o) in enumerate(zip(posterior.items, prior.items)):
match_result = match(s, o)
scalars[f"layer {idx}"] = float(match_result)
result = result + match_result
self.writer.add_scalars("layer kld", scalars, self.step_id)
return result
return match(posterior, prior)
def forward(self, inputs): posterior = self.encoder(inputs) code = posterior.rsample() prior = Normal(torch.zeros_like(code), torch.ones_like(code)) self.loss = match(posterior, prior) reconstruction = self.decoder(code) return reconstruction
def match(self, other):
result = 0.0
# print(self.items[0].loc, other.items[0].loc)
# print(self.items[0].scale, other.items[0].scale)
for s, o in zip(self.items, other.items):
match_result = match(s, o)
result = result + match_result
return result
def stack_loss(self, task_args, scale_args):
task_loss = 0.0
count = 0
for idx, ta in enumerate(zip(*task_args)):
level_loss = self.task_loss(*ta)
self.current_losses[f"level {idx}"] = float(level_loss)
task_loss += level_loss
count += 1
task_loss = task_loss / count
self.current_losses["task"] = float(task_loss)
prior_loss = match(MatchableList(scale_args.prior),
MatchableList(detach(scale_args.posterior)))
self.current_losses["prior"] = float(prior_loss)
policy_loss = self.policy_loss(task_args, scale_args)
self.current_losses["policy"] = float(policy_loss)
result = self.task_weight * task_loss
result = result + self.prior_weight * prior_loss
result = result + self.policy_weight * policy_loss
return result
def prior_loss(self, data, reconstruction, prior, codes):
rec_loss = match(data, reconstruction).mean()
kl_loss = match(prior, codes).mean()
self.current_losses["kullback leibler"] = float(kl_loss)
self.current_losses["reconstruction"] = float(rec_loss)
return rec_loss + kl_loss
def prior_loss(self, generated_codes, prior_codes):
result = match(generated_codes, prior_codes).mean()
self.current_losses["prior"] = float(result)
return result
def loss(self, reconstruction, target): return match(reconstruction, target)
def divergence_loss(self, posterior, prior): return match(posterior, prior)
