File size: 11,877 Bytes
cfb7702
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
"""
    Partially ported from https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/sampling.py
"""


from typing import Dict, Union

import torch
from omegaconf import ListConfig, OmegaConf
from tqdm import tqdm

from ...modules.diffusionmodules.sampling_utils import (
    get_ancestral_step,
    linear_multistep_coeff,
    to_d,
    to_neg_log_sigma,
    to_sigma,
)
from ...util import append_dims, default, instantiate_from_config

DEFAULT_GUIDER = {"target": "sgm.modules.diffusionmodules.guiders.IdentityGuider"}


class BaseDiffusionSampler:
    def __init__(
        self,
        discretization_config: Union[Dict, ListConfig, OmegaConf],
        num_steps: Union[int, None] = None,
        guider_config: Union[Dict, ListConfig, OmegaConf, None] = None,
        verbose: bool = False,
        device: str = "cuda",
    ):
        self.num_steps = num_steps
        self.discretization = instantiate_from_config(discretization_config)
        self.guider = instantiate_from_config(
            default(
                guider_config,
                DEFAULT_GUIDER,
            )
        )
        self.verbose = verbose
        self.device = device

    def prepare_sampling_loop(self, x, cond, uc=None, num_steps=None):
        sigmas = self.discretization(
            self.num_steps if num_steps is None else num_steps, device=self.device
        )
        uc = default(uc, cond)

        x *= torch.sqrt(1.0 + sigmas[0] ** 2.0)
        num_sigmas = len(sigmas)

        s_in = x.new_ones([x.shape[0]])

        return x, s_in, sigmas, num_sigmas, cond, uc

    def denoise(self, x, denoiser, sigma, cond, uc):
        denoised = denoiser(*self.guider.prepare_inputs(x, sigma, cond, uc))
        denoised = self.guider(denoised, sigma)
        return denoised

    def get_sigma_gen(self, num_sigmas):
        sigma_generator = range(num_sigmas - 1)
        if self.verbose:
            print("#" * 30, " Sampling setting ", "#" * 30)
            print(f"Sampler: {self.__class__.__name__}")
            print(f"Discretization: {self.discretization.__class__.__name__}")
            print(f"Guider: {self.guider.__class__.__name__}")
            sigma_generator = tqdm(
                sigma_generator,
                total=num_sigmas,
                desc=f"Sampling with {self.__class__.__name__} for {num_sigmas} steps",
            )
        return sigma_generator


class SingleStepDiffusionSampler(BaseDiffusionSampler):
    def sampler_step(self, sigma, next_sigma, denoiser, x, cond, uc, *args, **kwargs):
        raise NotImplementedError

    def euler_step(self, x, d, dt):
        return x + dt * d


class EDMSampler(SingleStepDiffusionSampler):
    def __init__(
        self, s_churn=0.0, s_tmin=0.0, s_tmax=float("inf"), s_noise=1.0, *args, **kwargs
    ):
        super().__init__(*args, **kwargs)

        self.s_churn = s_churn
        self.s_tmin = s_tmin
        self.s_tmax = s_tmax
        self.s_noise = s_noise

    def sampler_step(self, sigma, next_sigma, denoiser, x, cond, uc=None, gamma=0.0):
        sigma_hat = sigma * (gamma + 1.0)
        if gamma > 0:
            eps = torch.randn_like(x) * self.s_noise
            x = x + eps * append_dims(sigma_hat**2 - sigma**2, x.ndim) ** 0.5

        denoised = self.denoise(x, denoiser, sigma_hat, cond, uc)
        d = to_d(x, sigma_hat, denoised)
        dt = append_dims(next_sigma - sigma_hat, x.ndim)

        euler_step = self.euler_step(x, d, dt)
        x = self.possible_correction_step(
            euler_step, x, d, dt, next_sigma, denoiser, cond, uc
        )
        return x

    def __call__(self, denoiser, x, cond, uc=None, num_steps=None):
        x, s_in, sigmas, num_sigmas, cond, uc = self.prepare_sampling_loop(
            x, cond, uc, num_steps
        )

        for i in self.get_sigma_gen(num_sigmas):
            gamma = (
                min(self.s_churn / (num_sigmas - 1), 2**0.5 - 1)
                if self.s_tmin <= sigmas[i] <= self.s_tmax
                else 0.0
            )
            x = self.sampler_step(
                s_in * sigmas[i],
                s_in * sigmas[i + 1],
                denoiser,
                x,
                cond,
                uc,
                gamma,
            )

        return x


class AncestralSampler(SingleStepDiffusionSampler):
    def __init__(self, eta=1.0, s_noise=1.0, *args, **kwargs):
        super().__init__(*args, **kwargs)

        self.eta = eta
        self.s_noise = s_noise
        self.noise_sampler = lambda x: torch.randn_like(x)

    def ancestral_euler_step(self, x, denoised, sigma, sigma_down):
        d = to_d(x, sigma, denoised)
        dt = append_dims(sigma_down - sigma, x.ndim)

        return self.euler_step(x, d, dt)

    def ancestral_step(self, x, sigma, next_sigma, sigma_up):
        x = torch.where(
            append_dims(next_sigma, x.ndim) > 0.0,
            x + self.noise_sampler(x) * self.s_noise * append_dims(sigma_up, x.ndim),
            x,
        )
        return x

    def __call__(self, denoiser, x, cond, uc=None, num_steps=None):
        x, s_in, sigmas, num_sigmas, cond, uc = self.prepare_sampling_loop(
            x, cond, uc, num_steps
        )

        for i in self.get_sigma_gen(num_sigmas):
            x = self.sampler_step(
                s_in * sigmas[i],
                s_in * sigmas[i + 1],
                denoiser,
                x,
                cond,
                uc,
            )

        return x


class LinearMultistepSampler(BaseDiffusionSampler):
    def __init__(
        self,
        order=4,
        *args,
        **kwargs,
    ):
        super().__init__(*args, **kwargs)

        self.order = order

    def __call__(self, denoiser, x, cond, uc=None, num_steps=None, **kwargs):
        x, s_in, sigmas, num_sigmas, cond, uc = self.prepare_sampling_loop(
            x, cond, uc, num_steps
        )

        ds = []
        sigmas_cpu = sigmas.detach().cpu().numpy()
        for i in self.get_sigma_gen(num_sigmas):
            sigma = s_in * sigmas[i]
            denoised = denoiser(
                *self.guider.prepare_inputs(x, sigma, cond, uc), **kwargs
            )
            denoised = self.guider(denoised, sigma)
            d = to_d(x, sigma, denoised)
            ds.append(d)
            if len(ds) > self.order:
                ds.pop(0)
            cur_order = min(i + 1, self.order)
            coeffs = [
                linear_multistep_coeff(cur_order, sigmas_cpu, i, j)
                for j in range(cur_order)
            ]
            x = x + sum(coeff * d for coeff, d in zip(coeffs, reversed(ds)))

        return x


class EulerEDMSampler(EDMSampler):
    def possible_correction_step(
        self, euler_step, x, d, dt, next_sigma, denoiser, cond, uc
    ):
        return euler_step


class HeunEDMSampler(EDMSampler):
    def possible_correction_step(
        self, euler_step, x, d, dt, next_sigma, denoiser, cond, uc
    ):
        if torch.sum(next_sigma) < 1e-14:
            # Save a network evaluation if all noise levels are 0
            return euler_step
        else:
            denoised = self.denoise(euler_step, denoiser, next_sigma, cond, uc)
            d_new = to_d(euler_step, next_sigma, denoised)
            d_prime = (d + d_new) / 2.0

            # apply correction if noise level is not 0
            x = torch.where(
                append_dims(next_sigma, x.ndim) > 0.0, x + d_prime * dt, euler_step
            )
            return x


class EulerAncestralSampler(AncestralSampler):
    def sampler_step(self, sigma, next_sigma, denoiser, x, cond, uc):
        sigma_down, sigma_up = get_ancestral_step(sigma, next_sigma, eta=self.eta)
        denoised = self.denoise(x, denoiser, sigma, cond, uc)
        x = self.ancestral_euler_step(x, denoised, sigma, sigma_down)
        x = self.ancestral_step(x, sigma, next_sigma, sigma_up)

        return x


class DPMPP2SAncestralSampler(AncestralSampler):
    def get_variables(self, sigma, sigma_down):
        t, t_next = [to_neg_log_sigma(s) for s in (sigma, sigma_down)]
        h = t_next - t
        s = t + 0.5 * h
        return h, s, t, t_next

    def get_mult(self, h, s, t, t_next):
        mult1 = to_sigma(s) / to_sigma(t)
        mult2 = (-0.5 * h).expm1()
        mult3 = to_sigma(t_next) / to_sigma(t)
        mult4 = (-h).expm1()

        return mult1, mult2, mult3, mult4

    def sampler_step(self, sigma, next_sigma, denoiser, x, cond, uc=None, **kwargs):
        sigma_down, sigma_up = get_ancestral_step(sigma, next_sigma, eta=self.eta)
        denoised = self.denoise(x, denoiser, sigma, cond, uc)
        x_euler = self.ancestral_euler_step(x, denoised, sigma, sigma_down)

        if torch.sum(sigma_down) < 1e-14:
            # Save a network evaluation if all noise levels are 0
            x = x_euler
        else:
            h, s, t, t_next = self.get_variables(sigma, sigma_down)
            mult = [
                append_dims(mult, x.ndim) for mult in self.get_mult(h, s, t, t_next)
            ]

            x2 = mult[0] * x - mult[1] * denoised
            denoised2 = self.denoise(x2, denoiser, to_sigma(s), cond, uc)
            x_dpmpp2s = mult[2] * x - mult[3] * denoised2

            # apply correction if noise level is not 0
            x = torch.where(append_dims(sigma_down, x.ndim) > 0.0, x_dpmpp2s, x_euler)

        x = self.ancestral_step(x, sigma, next_sigma, sigma_up)
        return x


class DPMPP2MSampler(BaseDiffusionSampler):
    def get_variables(self, sigma, next_sigma, previous_sigma=None):
        t, t_next = [to_neg_log_sigma(s) for s in (sigma, next_sigma)]
        h = t_next - t

        if previous_sigma is not None:
            h_last = t - to_neg_log_sigma(previous_sigma)
            r = h_last / h
            return h, r, t, t_next
        else:
            return h, None, t, t_next

    def get_mult(self, h, r, t, t_next, previous_sigma):
        mult1 = to_sigma(t_next) / to_sigma(t)
        mult2 = (-h).expm1()

        if previous_sigma is not None:
            mult3 = 1 + 1 / (2 * r)
            mult4 = 1 / (2 * r)
            return mult1, mult2, mult3, mult4
        else:
            return mult1, mult2

    def sampler_step(
        self,
        old_denoised,
        previous_sigma,
        sigma,
        next_sigma,
        denoiser,
        x,
        cond,
        uc=None,
    ):
        denoised = self.denoise(x, denoiser, sigma, cond, uc)

        h, r, t, t_next = self.get_variables(sigma, next_sigma, previous_sigma)
        mult = [
            append_dims(mult, x.ndim)
            for mult in self.get_mult(h, r, t, t_next, previous_sigma)
        ]

        x_standard = mult[0] * x - mult[1] * denoised
        if old_denoised is None or torch.sum(next_sigma) < 1e-14:
            # Save a network evaluation if all noise levels are 0 or on the first step
            return x_standard, denoised
        else:
            denoised_d = mult[2] * denoised - mult[3] * old_denoised
            x_advanced = mult[0] * x - mult[1] * denoised_d

            # apply correction if noise level is not 0 and not first step
            x = torch.where(
                append_dims(next_sigma, x.ndim) > 0.0, x_advanced, x_standard
            )

        return x, denoised

    def __call__(self, denoiser, x, cond, uc=None, num_steps=None, **kwargs):
        x, s_in, sigmas, num_sigmas, cond, uc = self.prepare_sampling_loop(
            x, cond, uc, num_steps
        )

        old_denoised = None
        for i in self.get_sigma_gen(num_sigmas):
            x, old_denoised = self.sampler_step(
                old_denoised,
                None if i == 0 else s_in * sigmas[i - 1],
                s_in * sigmas[i],
                s_in * sigmas[i + 1],
                denoiser,
                x,
                cond,
                uc=uc,
            )

        return x