def validation_epoch_end(self, outputs):
metrics = [
'val_loss', 'val/recon_loss', 'val/kl_div_z', 'val/kl_div_b',
'val_rec_err'
]
tensorboard_logs = {}
for m in metrics:
loss_f = lambda x: x['log'][m]
losses = list(map(loss_f, outputs))
rec_err = sum(losses) / len(losses)
self.logger.experiment.add_scalar(m, rec_err, self.global_step)
self.log(m, rec_err)
tensorboard_logs[m] = rec_err
if (self.gt_eigvectors is not None) and (self.gt_eigs is not None):
ms = metric_subspace(self.vae, self.gt_eigvectors, self.gt_eigs)
ma = metric_alignment(self.vae, self.gt_eigvectors)
mp = metric_procrustes(self.vae, self.gt_eigvectors)
mr = metric_pairwise(self.vae, self.gt_eigvectors, self.gt_eigs)
tlog = {'subspace_distance': ms, 'alignment': ma, 'procrustes': mp}
self.logger.experiment.add_scalar('procrustes', mp,
self.global_step)
self.logger.experiment.add_scalar('pairwise_r', mr,
self.global_step)
self.logger.experiment.add_scalar('subspace_distance', ms,
self.global_step)
self.logger.experiment.add_scalar('alignment', ma,
self.global_step)
tensorboard_logs = {**tensorboard_logs, **tlog}
return {'val_loss': rec_err, 'log': tensorboard_logs}
def validation_epoch_end(self, outputs):
loss_f = lambda x: x['log']['val_rec_err']
losses = list(map(loss_f, outputs))
rec_err = sum(losses) / len(losses)
self.logger.experiment.add_scalar('val_rec_err', rec_err,
self.global_step)
loss_f = lambda x: x['log']['val_loss']
losses = list(map(loss_f, outputs))
loss = sum(losses) / len(losses)
self.logger.experiment.add_scalar('val_loss', loss, self.global_step)
self.log('val_loss', loss)
# Commenting out, since it is too slow...
# ortho, eig_err = metric_orthogonality(self.vae)
# self.logger.experiment.add_scalar('orthogonality',
# ortho, self.global_step)
# tensorboard_logs = dict(
# [('val_loss', loss),
# ('orthogonality', ortho),
# ('eigenvalue-error', eig_err)]
# )
tensorboard_logs = {'val_loss': loss, 'val_rec_error': rec_err}
if (self.gt_eigvectors is not None) and (self.gt_eigs is not None):
ms = metric_subspace(self.vae, self.gt_eigvectors, self.gt_eigs)
ma = metric_alignment(self.vae, self.gt_eigvectors)
mp = metric_procrustes(self.vae, self.gt_eigvectors)
mr = metric_pairwise(self.vae, self.gt_eigvectors, self.gt_eigs)
tlog = {'subspace_distance': ms, 'alignment': ma, 'procrustes': mp}
self.logger.experiment.add_scalar('procrustes', mp,
self.global_step)
self.logger.experiment.add_scalar('pairwise_r', mr,
self.global_step)
self.logger.experiment.add_scalar('subspace_distance', ms,
self.global_step)
self.logger.experiment.add_scalar('alignment', ma,
self.global_step)
tensorboard_logs = {**tensorboard_logs, **tlog}
return {'val_loss': loss, 'log': tensorboard_logs}
def validation_epoch_end(self, outputs):
loss_f = lambda x: x['log']['val_rec_err']
losses = list(map(loss_f, outputs))
rec_err = sum(losses) / len(losses)
self.logger.experiment.add_scalar('val_rec_err', rec_err,
self.global_step)
if self.hparams.encoder_depth == 1:
mt = metric_transpose_theorem(self.model)
self.logger.experiment.add_scalar('transpose', mt,
self.global_step)
ortho, eig_err = metric_orthogonality(self.model)
self.logger.experiment.add_scalar('orthogonality', ortho,
self.global_step)
tensorboard_logs = dict([
('val_loss', rec_err),
# ('transpose', mt),
('orthogonality', ortho),
('eigenvalue-error', eig_err)
])
if (self.gt_eigvectors is not None) and (self.gt_eigs is not None):
ms = metric_subspace(self.model, self.gt_eigvectors, self.gt_eigs)
ma = metric_alignment(self.model, self.gt_eigvectors)
mp = metric_procrustes(self.model, self.gt_eigvectors)
mr = metric_pairwise(self.model, self.gt_eigvectors, self.gt_eigs)
tlog = {'subspace_distance': ms, 'alignment': ma, 'procrustes': mp}
self.logger.experiment.add_scalar('procrustes', mp,
self.global_step)
self.logger.experiment.add_scalar('pairwise_r', mr,
self.global_step)
self.logger.experiment.add_scalar('subspace_distance', ms,
self.global_step)
self.logger.experiment.add_scalar('alignment', ma,
self.global_step)
tensorboard_logs = {**tensorboard_logs, **tlog}
return {'val_loss': rec_err, 'log': tensorboard_logs}