def _get_mu_sigma_single(self, autoencoder, x):
ae_output = autoencoder(x)
y_mu = ae_output[0]
if self.decoder_sigma_enabled:
y_sigma = ae_output[self.decoder_sig_index]
else:
y_sigma = torch.ones_like(y_mu)
if self.decoder_cluster_enabled:
y_cluster = ae_output[
self.decoder_cluster_index].detach().cpu().numpy()
#convert to numpy
y_mu = y_mu.detach().cpu().numpy()
if self.decoder_full_cov_enabled:
#convert chol tril and log noise to numpy
y_sigma = tuple(
[y_sigma_i.detach().cpu().numpy() for y_sigma_i in y_sigma])
else:
y_sigma = y_sigma.detach().cpu().numpy()
y_sigma = flatten_np(y_sigma)
log_noise = autoencoder.log_noise.detach().cpu().numpy()
if self.decoder_cluster_enabled:
return flatten_np(y_mu), y_sigma, log_noise, y_cluster
else:
return flatten_np(y_mu), y_sigma, log_noise
def calc_output_single(self,
autoencoder,
x,
select_keys=[
"y_mu", "y_sigma", "se", "bce", "cbce",
"nll_homo", "nll_sigma"
]):
"""
Computes the output of autoencoder per sample. Given the selected keys, we specify the output not only to be the reconstructed signal, but options of
squared error (`se`), Gaussian negative loglikelihood (`nll_homo`), etc. This function is used later for every sampled autoencoder from the posterior weights.
"""
#per sample
if self.decoder_cluster_enabled:
y_mu, y_sigma, log_noise, y_cluster = self._get_mu_sigma_single(
autoencoder, x)
else:
y_mu, y_sigma, log_noise = self._get_mu_sigma_single(
autoencoder, x)
#clamp it to min max
if self.output_clamp:
y_mu = np.clip(y_mu,
a_min=self.output_clamp[0],
a_max=self.output_clamp[1])
#flatten x into numpy array
x = flatten_np(x.detach().cpu().numpy())
return self._calc_output_single(x,
y_mu=y_mu,
y_sigma=y_sigma,
log_noise=log_noise,
select_keys=select_keys)
def inverse_transform(self, x_test):
if len(x_test.shape) > 2:
new_x_test = self.scaler.inverse_transform(
flatten_np(x_test)).reshape(x_test.shape)
else:
new_x_test = self.scaler.inverse_transform(x_test)
return new_x_test
def transform(self, x_test):
if len(x_test.shape) > 2:
new_x_test = self.scaler.transform(flatten_np(x_test)).reshape(
x_test.shape)
else:
new_x_test = self.scaler.transform(x_test)
if self.clip:
new_x_test = np.clip(new_x_test, 0, 1)
return new_x_test
def fit_transform(self, x_train, y_train=None):
if len(x_train.shape) > 2:
new_x_train = self.scaler.fit_transform(
flatten_np(x_train)).reshape(x_train.shape)
else:
new_x_train = self.scaler.fit_transform(x_train)
if self.clip:
new_x_train = np.clip(new_x_train, 0, 1)
return new_x_train
def fit(self, x_train, y_train=None):
if len(x_train.shape) > 2:
self.scaler.fit(flatten_np(x_train))
else:
self.scaler.fit(x_train)