(self,
unet_config,
timesteps=1000,
beta_schedule="linear",
loss_type="l2",
ckpt_path=None,
ignore_keys=[],
load_only_unet=False,
monitor="val/loss",
use_ema=True,
first_stage_key="image",
image_size=256,
channels=3,
log_every_t=100,
clip_denoised=True,
linear_start=1e-4,
linear_end=2e-2,
cosine_s=8e-3,
given_betas=None,
original_elbo_weight=0.,
v_posterior=0., # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta
l_simple_weight=1.,
conditioning_key=None,
parameterization="eps", # all assuming fixed variance schedules
scheduler_config=None,
use_positional_encodings=False,
learn_logvar=False,
logvar_init=0.,
make_it_fit=False,
ucg_training=None,
reset_ema=False,
reset_num_ema_updates=False,
)
| 53 | class DDPM(pl.LightningModule): |
| 54 | # classic DDPM with Gaussian diffusion, in image space |
| 55 | def __init__(self, |
| 56 | unet_config, |
| 57 | timesteps=1000, |
| 58 | beta_schedule="linear", |
| 59 | loss_type="l2", |
| 60 | ckpt_path=None, |
| 61 | ignore_keys=[], |
| 62 | load_only_unet=False, |
| 63 | monitor="val/loss", |
| 64 | use_ema=True, |
| 65 | first_stage_key="image", |
| 66 | image_size=256, |
| 67 | channels=3, |
| 68 | log_every_t=100, |
| 69 | clip_denoised=True, |
| 70 | linear_start=1e-4, |
| 71 | linear_end=2e-2, |
| 72 | cosine_s=8e-3, |
| 73 | given_betas=None, |
| 74 | original_elbo_weight=0., |
| 75 | v_posterior=0., # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta |
| 76 | l_simple_weight=1., |
| 77 | conditioning_key=None, |
| 78 | parameterization="eps", # all assuming fixed variance schedules |
| 79 | scheduler_config=None, |
| 80 | use_positional_encodings=False, |
| 81 | learn_logvar=False, |
| 82 | logvar_init=0., |
| 83 | make_it_fit=False, |
| 84 | ucg_training=None, |
| 85 | reset_ema=False, |
| 86 | reset_num_ema_updates=False, |
| 87 | ): |
| 88 | super().__init__() |
| 89 | assert parameterization in ["eps", "x0", "v"], 'currently only supporting "eps" and "x0" and "v"' |
| 90 | self.parameterization = parameterization |
| 91 | print(f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode") |
| 92 | self.cond_stage_model = None |
| 93 | self.clip_denoised = clip_denoised |
| 94 | self.log_every_t = log_every_t |
| 95 | self.first_stage_key = first_stage_key |
| 96 | self.image_size = image_size # try conv? |
| 97 | self.channels = channels |
| 98 | self.use_positional_encodings = use_positional_encodings |
| 99 | self.model = DiffusionWrapper(unet_config, conditioning_key) |
| 100 | count_params(self.model, verbose=True) |
| 101 | self.use_ema = use_ema |
| 102 | if self.use_ema: |
| 103 | self.model_ema = LitEma(self.model) |
| 104 | print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.") |
| 105 | |
| 106 | self.use_scheduler = scheduler_config is not None |
| 107 | if self.use_scheduler: |
| 108 | self.scheduler_config = scheduler_config |
| 109 | |
| 110 | self.v_posterior = v_posterior |
| 111 | self.original_elbo_weight = original_elbo_weight |
| 112 | self.l_simple_weight = l_simple_weight |
no test coverage detected