| 43 | |
| 44 | |
| 45 | class DiagonalGaussianDistribution: |
| 46 | def __init__(self, parameters, deterministic=False): |
| 47 | self.parameters = parameters |
| 48 | self.mean, self.logvar = torch.chunk(parameters, 2, dim=1) |
| 49 | self.logvar = torch.clamp(self.logvar, -30.0, 20.0) |
| 50 | self.deterministic = deterministic |
| 51 | self.std = torch.exp(0.5 * self.logvar) |
| 52 | self.var = torch.exp(self.logvar) |
| 53 | if self.deterministic: |
| 54 | self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device) |
| 55 | |
| 56 | def sample(self): |
| 57 | x = self.mean + self.std * torch.randn(self.mean.shape).to(device=self.parameters.device) |
| 58 | return x |
| 59 | |
| 60 | def kl(self, other=None): |
| 61 | if self.deterministic: |
| 62 | return torch.Tensor([0.]) |
| 63 | else: |
| 64 | if other is None: |
| 65 | return 0.5 * torch.sum(torch.pow(self.mean, 2) |
| 66 | + self.var - 1.0 - self.logvar, |
| 67 | dim=[1, 2, 3]) |
| 68 | else: |
| 69 | return 0.5 * torch.sum( |
| 70 | torch.pow(self.mean - other.mean, 2) / other.var |
| 71 | + self.var / other.var - 1.0 - self.logvar + other.logvar, |
| 72 | dim=[1, 2, 3]) |
| 73 | |
| 74 | def nll(self, sample, dims=[1,2,3]): |
| 75 | if self.deterministic: |
| 76 | return torch.Tensor([0.]) |
| 77 | logtwopi = np.log(2.0 * np.pi) |
| 78 | return 0.5 * torch.sum( |
| 79 | logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var, |
| 80 | dim=dims) |
| 81 | |
| 82 | def mode(self): |
| 83 | return self.mean |
| 84 | |
| 85 | |
| 86 | """ Resnet blocks """ |