def sample(self, inf_data_in):
if self.sample_stored_output is not None:
output_data = self.sample_stored_output
else:
if inf_data_in is not None:
inf_layer_data = self.inf_pre(inf_data_in)
else:
inf_layer_data = None
if self.sublayer is not None:
next_inf_data = self.inf_post(inf_layer_data)
if self.inf_batchnorm is not None:
next_inf_data = self.inf_batchnorm(next_inf_data)
gen_data_in = self.sublayer.sample(next_inf_data)
else:
gen_data_in = None
if not (gen_data_in is None):
gen_layer_data = self.gen_pre(gen_data_in)
else:
gen_layer_data = None
if self.latent_q is not None or self.latent_p is not None:
p_data = gen_layer_data
q_data = torch_ext.cat((inf_layer_data, gen_layer_data), 1)
latent_q_dist = self.latent_q(q_data)
if p_data is None:
latent_p_dist = tdist.Normal(
torch.zeros_like(latent_q_dist.loc),
torch.ones_like(latent_q_dist.loc))
else:
latent_p_dist = self.latent_p(p_data)
# latent_q_dist = torch_ext.combine(latent_q_dist, latent_p_dist)
latent_sample = latent_p_dist.rsample()
latent_dec = self.latent_dec(latent_sample)
else:
latent_dec = None
if p_data is not None and self.bypass is not None:
bypass = self.bypass(p_data)
else:
bypass = None
output_data = self.gen_post(bypass, latent_dec)
self.sample_stored_output = output_data
if self.gen_batchnorm is not None:
output_data = self.gen_batchnorm(output_data)
return output_data
def forward(self, x, *additional_input):
x = torch_ext.cat((x, *additional_input), 1)
return self.layers(x) + self.skip(x)
def forward(self, inf_data_in, lengths=None):
if inf_data_in is not None:
if len(inf_data_in.size()) > 2:
inf_layer_data = self.inf_pre(inf_data_in.mean(-1))
else:
inf_layer_data = self.inf_pre(inf_data_in)
else:
inf_layer_data = None
if self.sublayer is not None:
next_inf_data = self.inf_post(inf_layer_data)
if self.inf_batchnorm is not None:
next_inf_data = self.inf_batchnorm(next_inf_data)
gen_data_in, ls = self.sublayer(next_inf_data)
else:
ls = []
gen_data_in = None
if not (gen_data_in is None):
gen_layer_data = self.gen_pre(gen_data_in)
else:
gen_layer_data = None
p_data = gen_layer_data
q_data = torch_ext.cat((inf_layer_data, gen_layer_data), 1)
if self.latent_q is not None or self.latent_p is not None:
latent_q_dist = self.latent_q(q_data)
if p_data is None:
latent_p_dist = tdist.Normal(
torch.zeros_like(latent_q_dist.loc),
torch.ones_like(latent_q_dist.loc))
else:
latent_p_dist = self.latent_p(p_data)
latent_q_sample = latent_q_dist.rsample()
latent_dec = self.latent_dec(latent_q_sample)
this_ls = [
LatentState(latent_q_dist, latent_p_dist, latent_q_sample)
]
else:
latent_dec = None
this_ls = []
if p_data is not None and self.bypass is not None:
bypass = self.bypass(p_data)
else:
bypass = None
output_data = self.gen_post(bypass, latent_dec)
new_ls = ls + this_ls
if self.gen_batchnorm is not None:
output_data = self.gen_batchnorm(output_data)
if inf_data_in is not None:
if len(inf_data_in.size()) > 2:
output_data = output_data[...,
None].repeat(1, 1,
inf_data_in.size()[2])
return output_data, new_ls
def forward(self, x, *cond):
return self.layers(torch_ext.cat((x, *cond), 1)) + self.skip(x)