| 89 | |
| 90 | |
| 91 | class GaussianDiffusion(nn.Module): |
| 92 | def __init__(self, |
| 93 | timesteps=1000, |
| 94 | beta_schedule="linear", |
| 95 | zero_terminal_snr=False, |
| 96 | loss_type="l2", |
| 97 | parameterization="eps", # all assuming fixed variance schedules |
| 98 | linear_start=1e-4, |
| 99 | linear_end=2e-2, |
| 100 | cosine_s=8e-3, |
| 101 | original_elbo_weight=0., |
| 102 | v_posterior=0., # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta |
| 103 | l_simple_weight=1., |
| 104 | ): |
| 105 | super().__init__() |
| 106 | self.parameterization = parameterization |
| 107 | self.loss_type = loss_type |
| 108 | print(f"[{self.__class__.__name__}]: Running in {self.parameterization}-prediction mode") |
| 109 | |
| 110 | self.v_posterior = v_posterior |
| 111 | self.original_elbo_weight = original_elbo_weight |
| 112 | self.l_simple_weight = l_simple_weight |
| 113 | |
| 114 | if zero_terminal_snr: |
| 115 | assert beta_schedule == "linear", "enforce_zero_terminal_snr only works with linear beta schedules" |
| 116 | assert parameterization == "v", "enforce_zero_terminal_snr only works with v-parameterization" |
| 117 | |
| 118 | self.register_schedule( |
| 119 | beta_schedule=beta_schedule, |
| 120 | timesteps=timesteps, |
| 121 | linear_start=linear_start, |
| 122 | linear_end=linear_end, |
| 123 | cosine_s=cosine_s, |
| 124 | zero_terminal_snr=zero_terminal_snr |
| 125 | ) |
| 126 | |
| 127 | self.loss_type = loss_type |
| 128 | |
| 129 | def register_schedule(self, beta_schedule="linear", timesteps=1000, |
| 130 | linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3, |
| 131 | zero_terminal_snr=False): |
| 132 | betas = make_beta_schedule( |
| 133 | beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s |
| 134 | ) |
| 135 | if zero_terminal_snr: |
| 136 | betas = enforce_zero_terminal_snr(betas) |
| 137 | alphas = 1. - betas |
| 138 | alphas_cumprod = np.cumprod(alphas, axis=0) |
| 139 | alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1]) |
| 140 | alphas_cumprod_next = np.append(alphas_cumprod[1:], 0.0) |
| 141 | |
| 142 | timesteps, = betas.shape |
| 143 | self.num_timesteps = int(timesteps) |
| 144 | self.linear_start = linear_start |
| 145 | self.linear_end = linear_end |
| 146 | assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep' |
| 147 | |
| 148 | to_torch = partial(torch.tensor, dtype=torch.float32) |