(self, cond, shape,
x_T=None, ddim_use_original_steps=False,
callback=None, timesteps=None, quantize_denoised=False,
mask=None, x0=None, img_callback=None, log_every_t=100,
temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
unconditional_guidance_scale=1., unconditional_conditioning=None, dynamic_threshold=None,
ucg_schedule=None)
| 121 | |
| 122 | @torch.no_grad() |
| 123 | def ddim_sampling(self, cond, shape, |
| 124 | x_T=None, ddim_use_original_steps=False, |
| 125 | callback=None, timesteps=None, quantize_denoised=False, |
| 126 | mask=None, x0=None, img_callback=None, log_every_t=100, |
| 127 | temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, |
| 128 | unconditional_guidance_scale=1., unconditional_conditioning=None, dynamic_threshold=None, |
| 129 | ucg_schedule=None): |
| 130 | device = self.model.betas.device |
| 131 | b = shape[0] |
| 132 | if x_T is None: |
| 133 | img = torch.randn(shape, device=device) |
| 134 | else: |
| 135 | img = x_T |
| 136 | |
| 137 | if timesteps is None: |
| 138 | timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps |
| 139 | elif timesteps is not None and not ddim_use_original_steps: |
| 140 | subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1 |
| 141 | timesteps = self.ddim_timesteps[:subset_end] |
| 142 | |
| 143 | intermediates = {'x_inter': [img], 'pred_x0': [img]} |
| 144 | time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps) |
| 145 | total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0] |
| 146 | print(f"Running DDIM Sampling with {total_steps} timesteps") |
| 147 | |
| 148 | iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps) |
| 149 | |
| 150 | for i, step in enumerate(iterator): |
| 151 | index = total_steps - i - 1 |
| 152 | ts = torch.full((b,), step, device=device, dtype=torch.long) |
| 153 | |
| 154 | if mask is not None: |
| 155 | assert x0 is not None |
| 156 | img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass? |
| 157 | img = img_orig * mask + (1. - mask) * img |
| 158 | |
| 159 | if ucg_schedule is not None: |
| 160 | assert len(ucg_schedule) == len(time_range) |
| 161 | unconditional_guidance_scale = ucg_schedule[i] |
| 162 | |
| 163 | outs = self.p_sample_ddim(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps, |
| 164 | quantize_denoised=quantize_denoised, temperature=temperature, |
| 165 | noise_dropout=noise_dropout, score_corrector=score_corrector, |
| 166 | corrector_kwargs=corrector_kwargs, |
| 167 | unconditional_guidance_scale=unconditional_guidance_scale, |
| 168 | unconditional_conditioning=unconditional_conditioning, |
| 169 | dynamic_threshold=dynamic_threshold) |
| 170 | img, pred_x0 = outs |
| 171 | if callback: callback(i) |
| 172 | if img_callback: img_callback(pred_x0, i) |
| 173 | |
| 174 | if index % log_every_t == 0 or index == total_steps - 1: |
| 175 | intermediates['x_inter'].append(img) |
| 176 | intermediates['pred_x0'].append(pred_x0) |
| 177 | |
| 178 | return img, intermediates |
| 179 | |
| 180 | @torch.no_grad() |
no test coverage detected