def make_latent_distributions(self, obs, out_obs):
"""
根据 batch, mask 输出 proposal 网络和 prior 网络的输出
No no_proposal is True, return None instead of proposal distribution.
"""
assert obs.numpy().shape[-1] == 1
# Proposal 网络输入是 原始图像 和 mask
proposal_params = self.proposal_network(tf.concat([obs, out_obs], axis=-1)) # (None,1,1,32)
proposal = normal_parse_params(proposal_params, 1e-3)
# Prior 网络输入是 mask之后的图像 和 mask
prior_params = self.prior_network(obs) # 在通道上进行连接
prior = normal_parse_params(prior_params, 1e-3)
return proposal, prior
def generate_samples_params(self, obs, k=100):
""" k 代表采样的个数. 从 prior network 输出分布中采样, 随后输入到 generative network 中采样
"""
prior_params = self.prior_network(obs)
prior = normal_parse_params(prior_params, 1e-3)
#
samples = []
for i in range(k):
latent = self.reparameterize(prior) # (batch,1,1,16) 重参数化并采样
sample_params = self.generative_network(latent) # (batch,28,28,1)
samples.append(sample_params)
return samples