(self, given_betas=None, beta_schedule="linear", timesteps=1000,
linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3)
| 143 | self.ucg_prng = np.random.RandomState() |
| 144 | |
| 145 | def register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000, |
| 146 | linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): |
| 147 | if exists(given_betas): |
| 148 | betas = given_betas |
| 149 | else: |
| 150 | betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end, |
| 151 | cosine_s=cosine_s) |
| 152 | alphas = 1. - betas |
| 153 | alphas_cumprod = np.cumprod(alphas, axis=0) |
| 154 | alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1]) |
| 155 | |
| 156 | timesteps, = betas.shape |
| 157 | self.num_timesteps = int(timesteps) |
| 158 | self.linear_start = linear_start |
| 159 | self.linear_end = linear_end |
| 160 | assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep' |
| 161 | |
| 162 | to_torch = partial(torch.tensor, dtype=torch.float32) |
| 163 | |
| 164 | self.register_buffer('betas', to_torch(betas)) |
| 165 | self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) |
| 166 | self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev)) |
| 167 | |
| 168 | # calculations for diffusion q(x_t | x_{t-1}) and others |
| 169 | self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod))) |
| 170 | self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod))) |
| 171 | self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod))) |
| 172 | self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod))) |
| 173 | self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1))) |
| 174 | |
| 175 | # calculations for posterior q(x_{t-1} | x_t, x_0) |
| 176 | posterior_variance = (1 - self.v_posterior) * betas * (1. - alphas_cumprod_prev) / ( |
| 177 | 1. - alphas_cumprod) + self.v_posterior * betas |
| 178 | # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t) |
| 179 | self.register_buffer('posterior_variance', to_torch(posterior_variance)) |
| 180 | # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain |
| 181 | self.register_buffer('posterior_log_variance_clipped', to_torch(np.log(np.maximum(posterior_variance, 1e-20)))) |
| 182 | self.register_buffer('posterior_mean_coef1', to_torch( |
| 183 | betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod))) |
| 184 | self.register_buffer('posterior_mean_coef2', to_torch( |
| 185 | (1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod))) |
| 186 | |
| 187 | if self.parameterization == "eps": |
| 188 | lvlb_weights = self.betas ** 2 / ( |
| 189 | 2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod)) |
| 190 | elif self.parameterization == "x0": |
| 191 | lvlb_weights = 0.5 * np.sqrt(torch.Tensor(alphas_cumprod)) / (2. * 1 - torch.Tensor(alphas_cumprod)) |
| 192 | elif self.parameterization == "v": |
| 193 | lvlb_weights = torch.ones_like(self.betas ** 2 / ( |
| 194 | 2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod))) |
| 195 | else: |
| 196 | raise NotImplementedError("mu not supported") |
| 197 | lvlb_weights[0] = lvlb_weights[1] |
| 198 | self.register_buffer('lvlb_weights', lvlb_weights, persistent=False) |
| 199 | assert not torch.isnan(self.lvlb_weights).all() |
| 200 | |
| 201 | @contextmanager |
| 202 | def ema_scope(self, context=None): |
no test coverage detected