File size: 12,920 Bytes
b910c09
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import einops
from tqdm import tqdm
from functools import partial
import torch.nn.functional as F
from abc import ABC, abstractmethod

from patch_flow.flow_pf import pad_v_like_x_patches


# ===================================================================================================


def forward_with_cfg_and_uncertainty(x, t, model, cfg_scale=1.0, uc_cond=None, cond_key="y", **model_kwargs):
    """Function to include sampling with Classifier-Free Guidance (CFG)"""
    if cfg_scale == 1.0:  # without CFG
        model_output = model(x, t, **model_kwargs, return_uncertainty=True)
        model_vt, model_uq = model_output
        out = {"vt": model_vt, "uq": model_uq, "uq_uc": None, "uq_c": model_uq, "vt_uc": None, "vt_c": model_vt}

    else:  # with CFG
        assert cond_key in model_kwargs, f"Condition key '{cond_key}' for CFG not found in model_kwargs"
        assert uc_cond is not None, "Unconditional condition not provided for CFG"
        kwargs = model_kwargs.copy()
        c = kwargs[cond_key]
        x_in = torch.cat([x] * 2)
        t_in = torch.cat([t] * 2)
        if uc_cond.shape[0] == 1:
            uc_cond = einops.repeat(uc_cond, "1 ... -> bs ...", bs=x.shape[0])
        c_in = torch.cat([uc_cond, c])
        kwargs[cond_key] = c_in
        model_output = model(x_in, t_in, **kwargs, return_uncertainty=True)
        model_vt, model_uq = model_output
        model_vt_uc, model_vt_c = model_vt.chunk(2)
        model_uq_uc, model_uq_c = model_uq.chunk(2)
        guided_vt = model_vt_uc + cfg_scale * (model_vt_c - model_vt_uc)
        guided_uq = model_uq_uc + cfg_scale * (model_uq_c - model_uq_uc)

        out = {
            "vt": guided_vt,
            "uq": guided_uq,
            "uq_uc": model_uq_uc,
            "uq_c": model_uq_c,
            "vt_uc": model_vt_uc,
            "vt_c": model_vt_c,
        }

    return out


def forward_with_cfg(x, t, model, cfg_scale=1.0, uc_cond=None, cond_key="y", **model_kwargs):
    """Function to include sampling with Classifier-Free Guidance (CFG)"""
    if cfg_scale == 1.0:  # without CFG
        model_output = model(x, t, **model_kwargs)

    else:  # with CFG
        assert cond_key in model_kwargs, f"Condition key '{cond_key}' for CFG not found in model_kwargs"
        assert uc_cond is not None, "Unconditional condition not provided for CFG"
        kwargs = model_kwargs.copy()
        c = kwargs[cond_key]
        x_in = torch.cat([x] * 2)
        t_in = torch.cat([t] * 2)
        if uc_cond.shape[0] == 1:
            uc_cond = einops.repeat(uc_cond, "1 ... -> bs ...", bs=x.shape[0])
        c_in = torch.cat([uc_cond, c])
        kwargs[cond_key] = c_in
        model_uc, model_c = model(x_in, t_in, **kwargs).chunk(2)
        model_output = model_uc + cfg_scale * (model_c - model_uc)

    return model_output


def patch_reduce_pool(x: torch.Tensor, n: int, mode: str = "mean"):
    """Patch reducing with pooling (downsampling)"""
    if mode == "mean":
        return F.avg_pool2d(x, kernel_size=n, stride=n)
    elif mode == "max":
        return F.max_pool2d(x, kernel_size=n, stride=n)
    elif mode == "min":
        return -F.max_pool2d(-x, kernel_size=n, stride=n)
    else:
        raise ValueError("mode must be 'mean', 'max', or 'min'")


def patch_reduce(x: torch.Tensor, n: int, mode: str = "mean"):
    """Patch reduce and upsample to original size"""
    y = patch_reduce_pool(x, n=n, mode=mode)
    y = y.repeat_interleave(n, dim=-1).repeat_interleave(n, dim=-2)
    assert y.shape == x.shape
    return y


# ======================================================================================


class SamplerBase(ABC):
    @abstractmethod
    def __repr__(self) -> str: ...

    @abstractmethod
    def __call__(self, model, x, timesteps: list[float], progress: bool = True, **kwargs): ...


# ======================================================================================
# Base samplers


def euler(model, x, timesteps: list[float], progress=True, **kwargs):
    bs, dev = x.shape[0], x.device

    xt = x
    for t_curr, t_next in tqdm(zip(timesteps[:-1], timesteps[1:]), disable=not progress, total=len(timesteps) - 1):
        t = torch.ones((bs,), dtype=x.dtype, device=dev) * t_curr
        pred = model(xt, t, **kwargs)

        dt = t_next - t_curr
        xt = xt + dt * pred

    return xt


class Euler(SamplerBase):
    def __repr__(self):
        return "Euler"

    def __call__(self, model, x, timesteps: list[float], progress=True, **kwargs):
        model_fn = partial(forward_with_cfg, model=model)
        return euler(model_fn, x, timesteps, progress=progress, **kwargs)


# ======================================================================================
# Patch Forcing samplers


class EulerPF(SamplerBase):
    """Default Euler sampler, ignores uncertainty"""

    def __init__(self, patch_size: int = 2):
        self.patch_size = patch_size

    def __repr__(self):
        return "EulerPF"

    def __call__(self, model, x, timesteps: list[float], progress=True, **kwargs):
        dev = x.device
        bs, c, h, w = x.shape
        f = (h // self.patch_size) * (w // self.patch_size)

        # prepare sample function
        sample_fn = partial(forward_with_cfg_and_uncertainty, model=model)

        xt = x
        for t_curr, t_next in tqdm(zip(timesteps[:-1], timesteps[1:]), disable=not progress, total=len(timesteps) - 1):
            t = torch.ones((bs,), dtype=x.dtype, device=dev) * t_curr
            # Here we broadcast to (b, n)
            t = einops.repeat(t, "b -> b f", f=f)
            pred = sample_fn(xt, t, **kwargs)
            pred = pred["vt"]

            dt = t_next - t_curr
            xt = xt + dt * pred

        return xt


# ======================================================================================
# Uncertainty-aware samplers


class DualLoopSampler(SamplerBase):
    def __init__(self, p: float = 0.7, n_inner: int = 4, mode: str = "mean", patch_size: int = 2):
        """
        Args:
            p: percentile for thresholding uncertainty. All patches with uncertainty
                lower than the p-th percentile will be considered certain. So lower
                p -> more restrictive (fewer certain patches), e.g. p=0.8 means 80%
                of patches are considered certain (20% uncertain).
            n_inner: Number of inner steps, per big step.
        """
        self.p = p
        self.n_inner = n_inner  # inner steps for uncertain patches
        self.patch_size = patch_size
        self.mode = mode
        assert 0.0 < p < 1.0, "p must be in (0, 1)"

    def __repr__(self):
        return f"DualLoop-p{self.p*100:.0f}-inner{self.n_inner}"

    def compute_mask(self, uq):
        uq_flat = uq.reshape(uq.shape[0], -1).double()
        thresh = torch.quantile(uq_flat, self.p, dim=-1)  # (bs,)
        thresh_exp = einops.repeat(thresh, "b -> b 1 1 1")
        uq_mask = uq < thresh_exp  # 0: uncertain (inner loop), 1: certain (forward)
        return uq_mask

    def __call__(self, model, x, timesteps: list[float], progress=True, **kwargs):
        dev = x.device
        bs, c, h, w = x.shape
        f = (h // self.patch_size) * (w // self.patch_size)

        # make denoising schedule
        num_steps = len(timesteps) - 1
        denoise_schedule = torch.linspace(0, 1, num_steps + 1)
        denoise_schedule = einops.repeat(denoise_schedule, "t -> t f", f=f).to(dev)
        assert denoise_schedule.shape[1] == f

        # prepare sample function
        sample_fn = partial(forward_with_cfg_and_uncertainty, model=model)

        # sampling loop
        xt = x
        for t_curr, t_next in tqdm(
            zip(denoise_schedule[:-1], denoise_schedule[1:]), total=len(denoise_schedule) - 1, disable=not progress
        ):
            t = einops.repeat(t_curr, "f -> b f", b=bs)

            model_out = sample_fn(xt, t, **kwargs)
            pred = model_out["vt"]

            dt = t_next - t_curr
            dt = torch.clamp(dt, min=0.0)
            dt = einops.repeat(dt, "f -> b f", b=bs)
            dt_grid = pad_v_like_x_patches(dt, pred, patch_size=self.patch_size)

            # x1 prediction from xt (not used during inference)
            # dt_x1 = (1 - t)
            # dt_x1_grid = pad_v_like_x_patches(dt_x1, pred, patch_size=self.patch_size)
            # x1_pred = xt + dt_x1_grid * pred

            # # update xt       # NORMALLY USE THIS
            # xt = xt + dt_grid * pred

            # ============================================= inner loop update xt
            # with mask
            uq = model_out["uq"].exp()
            uq = patch_reduce(uq, n=2, mode=self.mode)
            uq_mask = self.compute_mask(uq)

            dt_inner_grid = dt_grid / self.n_inner

            xt = xt + dt_grid * pred * uq_mask + dt_inner_grid * pred * (~uq_mask)

            t_grid = pad_v_like_x_patches(t, pred, patch_size=self.patch_size)
            t_grid = t_grid + dt_grid * uq_mask + dt_inner_grid * (~uq_mask)

            for _ in range(self.n_inner - 1):
                t_inp = patch_reduce_pool(t_grid, n=2, mode="mean")
                t_inp = einops.rearrange(t_inp, "b 1 h w -> b (h w)")
                model_out_inner = sample_fn(xt, t_inp, **kwargs)
                pred = model_out_inner["vt"]
                xt = xt + dt_inner_grid * pred * (~uq_mask)
                t_grid = t_grid + dt_inner_grid * (~uq_mask)

        return xt


class LookAheadSampler(SamplerBase):
    def __init__(self, p: float = 0.4, mode: str = "mean", patch_size: int = 2, context_t_ratio: int = 1.5):
        """
        Context-guidance on uncertain patches during sampling. For certain patches, use one-step prediction for better context for uncertain patches.
        """
        self.p = p
        self.patch_size = patch_size
        self.mode = mode
        self.context_t_ratio = context_t_ratio
        assert 0.0 < p < 1.0, "p must be in (0, 1)"

    def __repr__(self):
        return f"LookAheadSampler-p{self.p*100:.0f}-context{self.context_t_ratio:.2f}"

    def compute_mask(self, uq):
        uq_flat = uq.reshape(uq.shape[0], -1).double()
        thresh = torch.quantile(uq_flat, self.p, dim=-1)  # (bs,)
        thresh_exp = einops.repeat(thresh, "b -> b 1 1 1")
        uq_mask = uq < thresh_exp  # 0: uncertain (use context), 1: certain (use model)
        return uq_mask

    def __call__(self, model, x, timesteps: list[float], progress=True, **kwargs):
        dev = x.device
        bs, c, h, w = x.shape
        f = (h // self.patch_size) * (w // self.patch_size)

        # make denoising schedule
        num_steps = len(timesteps) - 1
        denoise_schedule = torch.linspace(0, 1, num_steps + 1)
        denoise_schedule = einops.repeat(denoise_schedule, "t -> t f", f=f).to(dev)
        assert denoise_schedule.shape[1] == f

        # prepare sample function
        sample_fn = partial(forward_with_cfg_and_uncertainty, model=model)

        # sampling loop
        xt = x
        for t_curr, t_next in tqdm(
            zip(denoise_schedule[:-1], denoise_schedule[1:]), total=len(denoise_schedule) - 1, disable=not progress
        ):
            t = einops.repeat(t_curr, "f -> b f", b=bs)

            # No CFG for context prediction
            model_out = sample_fn(xt, t, **kwargs)
            pred = model_out["vt"]
            pred_c = model_out["vt_c"]

            dt = t_next - t_curr
            dt = torch.clamp(dt, min=0.0)
            dt = einops.repeat(dt, "f -> b f", b=bs)
            dt_grid = pad_v_like_x_patches(dt, pred, patch_size=self.patch_size)

            # normal step, no context guidance
            if t_curr.mean() <= 0.05:
                xt = xt + dt_grid * pred_c
                continue

            # =============================================
            uq = model_out["uq"].exp()
            uq = patch_reduce(uq, n=2, mode=self.mode)
            low_uq_mask = self.compute_mask(uq)
            high_uq_mask = ~low_uq_mask
            low_uq_pool_mask = patch_reduce_pool(low_uq_mask.float(), n=self.patch_size, mode=self.mode).bool()

            # one step prediction for certain patches
            t_context = t_curr * self.context_t_ratio
            t_context = torch.clamp(t_context, max=1.0)
            dt_context = t_context - t_curr
            dt_context = einops.repeat(dt_context, "f -> b f", b=bs)
            dt_context_grid = pad_v_like_x_patches(dt_context, pred, patch_size=self.patch_size)

            pred_context = pred_c * low_uq_mask
            xt_context = xt + dt_context_grid * pred_context
            t_context = t + dt_context * low_uq_pool_mask.view(bs, -1)

            # context prediction
            pred_context = sample_fn(xt_context, t_context, **kwargs)["vt"]

            # update xt
            xt = xt + dt_grid * pred * low_uq_mask + dt_grid * pred_context * high_uq_mask

        return xt