memef4rmer/edit_anything

1"""SAMPLING ONLY."""2import torch3 4from .dpm_solver import NoiseScheduleVP, model_wrapper, DPM_Solver5 6 7MODEL_TYPES = {8    "eps": "noise",9    "v": "v"10}11 12 13class DPMSolverSampler(object):14    def __init__(self, model, **kwargs):15        super().__init__()16        self.model = model17        to_torch = lambda x: x.clone().detach().to(torch.float32).to(model.device)18        self.register_buffer('alphas_cumprod', to_torch(model.alphas_cumprod))19 20    def register_buffer(self, name, attr):21        if type(attr) == torch.Tensor:22            if attr.device != torch.device("cuda"):23                attr = attr.to(torch.device("cuda"))24        setattr(self, name, attr)25 26    @torch.no_grad()27    def sample(self,28               S,29               batch_size,30               shape,31               conditioning=None,32               callback=None,33               normals_sequence=None,34               img_callback=None,35               quantize_x0=False,36               eta=0.,37               mask=None,38               x0=None,39               temperature=1.,40               noise_dropout=0.,41               score_corrector=None,42               corrector_kwargs=None,43               verbose=True,44               x_T=None,45               log_every_t=100,46               unconditional_guidance_scale=1.,47               unconditional_conditioning=None,48               # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...49               **kwargs50               ):51        if conditioning is not None:52            if isinstance(conditioning, dict):53                cbs = conditioning[list(conditioning.keys())[0]].shape[0]54                if cbs != batch_size:55                    print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")56            else:57                if conditioning.shape[0] != batch_size:58                    print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")59 60        # sampling61        C, H, W = shape62        size = (batch_size, C, H, W)63 64        print(f'Data shape for DPM-Solver sampling is {size}, sampling steps {S}')65 66        device = self.model.betas.device67        if x_T is None:68            img = torch.randn(size, device=device)69        else:70            img = x_T71 72        ns = NoiseScheduleVP('discrete', alphas_cumprod=self.alphas_cumprod)73 74        model_fn = model_wrapper(75            lambda x, t, c: self.model.apply_model(x, t, c),76            ns,77            model_type=MODEL_TYPES[self.model.parameterization],78            guidance_type="classifier-free",79            condition=conditioning,80            unconditional_condition=unconditional_conditioning,81            guidance_scale=unconditional_guidance_scale,82        )83 84        dpm_solver = DPM_Solver(model_fn, ns, predict_x0=True, thresholding=False)85        x = dpm_solver.sample(img, steps=S, skip_type="time_uniform", method="multistep", order=2, lower_order_final=True)86 87        return x.to(device), None