File size: 17,228 Bytes
f3ff4f1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
from typing import *
import torch
import numpy as np
from tqdm import tqdm
from easydict import EasyDict as edict
from .base import Sampler
from .classifier_free_guidance_mixin import ClassifierFreeGuidanceSamplerMixin
from .guidance_interval_mixin import GuidanceIntervalSamplerMixin
import trellis.modules.sparse as sp
from trellis.modules.spatial import patchify, unpatchify

class FlowEulerSampler(Sampler):
    """
    Generate samples from a flow-matching model using Euler sampling.

    Args:
        sigma_min: The minimum scale of noise in flow.
    """
    def __init__(
        self,
        sigma_min: float,
    ):
        self.sigma_min = sigma_min

    def _eps_to_xstart(self, x_t, t, eps):
        assert x_t.shape == eps.shape
        return (x_t - (self.sigma_min + (1 - self.sigma_min) * t) * eps) / (1 - t)
    
    def _xstart_to_x_t(self, x_0, t, eps):
        assert x_0.shape == eps.shape
        return (1-t) * x_0 + (self.sigma_min + (1 - self.sigma_min) * t) * eps
    
    def _xstart_to_x_t(self, x_0, t, eps):
        assert x_0.shape == eps.shape
        return (1-t) * x_0 + (self.sigma_min + (1 - self.sigma_min) * t) * eps

    def _xstart_to_eps(self, x_t, t, x_0):
        assert x_t.shape == x_0.shape
        return (x_t - (1 - t) * x_0) / (self.sigma_min + (1 - self.sigma_min) * t)

    def _v_to_xstart_eps(self, x_t, t, v):
        assert x_t.shape == v.shape
        eps = (1 - t) * v + x_t
        x_0 = (1 - self.sigma_min) * x_t - (self.sigma_min + (1 - self.sigma_min) * t) * v
        return x_0, eps
        
    def _xstart_to_v(self, x_0, x_t, t):
        assert x_0.shape == x_t.shape
        return (x_t - (1 - self.sigma_min) * x_0) / (self.sigma_min + (1 - self.sigma_min) * t)


    def _inference_model(self, model, x_t, t, cond=None, **kwargs):
        t = torch.tensor([1000 * t] * x_t.shape[0], device=x_t.device, dtype=torch.float32)
        return model(x_t.to(torch.float32), t, cond, **kwargs)

    def _get_model_prediction(self, model, x_t, t, cond=None, **kwargs):
        param = kwargs.pop("parameterization", "v")
        if param == "v":
            pred_v = self._inference_model(model, x_t, t, cond, **kwargs)
            pred_x_0, pred_eps = self._v_to_xstart_eps(x_t=x_t, t=t, v=pred_v)
        elif param == "x0":
            pred_x_0 = self._inference_model(model, x_t, t, cond, **kwargs)
            pred_v = self._xstart_to_v(x_0=pred_x_0, x_t=x_t, t=t)
        return pred_x_0, None, pred_v

    def _get_model_gt(self, x_0, t, noise):
        gt_x_t = self._xstart_to_x_t(x_0, t, noise)
        gt_v = self._xstart_to_v(x_0, gt_x_t, t)
        return gt_x_t, gt_v

    @torch.no_grad()
    def sample_once(
        self,
        model,
        x_t,
        t: float,
        t_prev: float,
        cond: Optional[Any] = None,
        **kwargs
    ):
        """
        Sample x_{t-1} from the model using Euler method.
        
        Args:
            model: The model to sample from.
            x_t: The [N x C x ...] tensor of noisy inputs at time t.
            t: The current timestep.
            t_prev: The previous timestep.
            cond: conditional information.
            **kwargs: Additional arguments for model inference.

        Returns:
            a dict containing the following
            - 'pred_x_prev': x_{t-1}.
            - 'pred_x_0': a prediction of x_0.
        """
        pred_x_0, pred_eps, pred_v = self._get_model_prediction(model, x_t, t, cond, **kwargs)
        pred_x_prev = x_t - (t - t_prev) * pred_v
        return edict({"pred_x_prev": pred_x_prev, "pred_x_0": pred_x_0, "pred_eps": pred_eps})
    
    @torch.no_grad()
    def sample_once_featurevolume(
        self,
        model,
        cond_model,
        x_t,
        t: float,
        t_prev: float,
        cond: Optional[Any] = None,
        **kwargs
    ):
        """
        Sample x_{t-1} from the model using Euler method.
        
        Args:
            model: The model to sample from.
            x_t: The [N x C x ...] tensor of noisy inputs at time t.
            t: The current timestep.
            t_prev: The previous timestep.
            cond: conditional information.
            **kwargs: Additional arguments for model inference.

        Returns:
            a dict containing the following
            - 'pred_x_prev': x_{t-1}.
            - 'pred_x_0': a prediction of x_0.
        """
        if isinstance(cond, sp.SparseTensor):
            t_tmp = torch.tensor([1000 * t] * x_t.shape[0], device=x_t.device, dtype=x_t.dtype)
            t_embed = model.t_embedder(t_tmp).to(x_t.dtype)
            for block in cond_model:
                cond = block(cond, t_embed)
            if model.pe_mode == "ape":
                cond = cond + model.pos_embedder(cond.coords[:, 1:]).to(x_t.dtype)
            if 'neg_cond' in kwargs.keys():
                neg_cond = kwargs['neg_cond']
                for block in cond_model:
                    neg_cond = block(neg_cond, t_embed)
                if model.pe_mode == "ape":
                    neg_cond = neg_cond + model.pos_embedder(neg_cond.coords[:, 1:]).to(x_t.dtype)
                kwargs['neg_cond'] = neg_cond
        else:
            for block in cond_model:
                cond = block(cond)
            cond = patchify(cond, model.patch_size)
            cond = cond.view(*cond.shape[:2], -1).permute(0, 2, 1).contiguous()
            cond = cond + model.pos_emb[None].type(model.dtype)
            if 'neg_cond' in kwargs.keys():
                neg_cond = kwargs['neg_cond']
                for block in cond_model:
                    neg_cond = block(neg_cond)
                neg_cond = patchify(neg_cond, model.patch_size)
                neg_cond = neg_cond.view(*neg_cond.shape[:2], -1).permute(0, 2, 1).contiguous()
                neg_cond = neg_cond + model.pos_emb[None].type(model.dtype)
                kwargs['neg_cond'] = neg_cond
        pred_x_0, pred_eps, pred_v = self._get_model_prediction(model, x_t, t, cond, **kwargs)
        pred_x_prev = x_t - (t - t_prev) * pred_v
        return edict({"pred_x_prev": pred_x_prev, "pred_x_0": pred_x_0, "pred_eps": pred_eps})

    @torch.no_grad()
    def sample_featurevolume(
        self,
        model,
        cond_model,
        noise,
        cond: Optional[Any] = None,
        steps: int = 50,
        rescale_t: float = 1.0,
        verbose: bool = True,
        **kwargs
    ):
        """
        Generate samples from the model using Euler method.
        
        Args:
            model: The model to sample from.
            noise: The initial noise tensor.
            cond: conditional information.
            steps: The number of steps to sample.
            rescale_t: The rescale factor for t.
            verbose: If True, show a progress bar.
            **kwargs: Additional arguments for model_inference.

        Returns:
            a dict containing the following
            - 'samples': the model samples.
            - 'pred_x_t': a list of prediction of x_t.
            - 'pred_x_0': a list of prediction of x_0.
        """
        sample = noise
        t_seq = np.linspace(1, 0, steps + 1)
        t_seq = rescale_t * t_seq / (1 + (rescale_t - 1) * t_seq)
        t_pairs = list((t_seq[i], t_seq[i + 1]) for i in range(steps))
        ret = edict({"samples": None, "pred_x_t": [], "pred_x_0": []})
        for t, t_prev in tqdm(t_pairs, desc="Sampling", disable=not verbose):
            out = self.sample_once_featurevolume(model, cond_model, sample, t, t_prev, cond, **kwargs)
            sample = out.pred_x_prev
            ret.pred_x_t.append(out.pred_x_prev)
            ret.pred_x_0.append(out.pred_x_0)
        ret.samples = sample
        return ret
    
    @torch.no_grad()
    def sample(
        self,
        model,
        noise,
        cond: Optional[Any] = None,
        steps: int = 50,
        rescale_t: float = 1.0,
        verbose: bool = True,
        **kwargs
    ):
        """
        Generate samples from the model using Euler method.
        
        Args:
            model: The model to sample from.
            noise: The initial noise tensor.
            cond: conditional information.
            steps: The number of steps to sample.
            rescale_t: The rescale factor for t.
            verbose: If True, show a progress bar.
            **kwargs: Additional arguments for model_inference.

        Returns:
            a dict containing the following
            - 'samples': the model samples.
            - 'pred_x_t': a list of prediction of x_t.
            - 'pred_x_0': a list of prediction of x_0.
        """
        sample = noise
        t_seq = np.linspace(1, 0, steps + 1)
        t_seq = rescale_t * t_seq / (1 + (rescale_t - 1) * t_seq)
        t_pairs = list((t_seq[i], t_seq[i + 1]) for i in range(steps))
        ret = edict({"samples": None, "pred_x_t": [], "pred_x_0": []})
        for t, t_prev in tqdm(t_pairs, desc="Sampling", disable=not verbose):
            out = self.sample_once(model, sample, t, t_prev, cond, **kwargs)
            sample = out.pred_x_prev
            ret.pred_x_t.append(out.pred_x_prev)
            ret.pred_x_0.append(out.pred_x_0)
        ret.samples = sample
        return ret


class FlowMatchingSampler(FlowEulerSampler):
    """
    Implementation of Flow Matching using Euler sampling.
    Inherits from FlowEulerSampler and modifies key methods for flow matching.
    """
    def __init__(self, sigma_min: float = 0.0):
        super().__init__(sigma_min=sigma_min)

    def _compute_velocity(self, x_t: torch.Tensor, x_0: torch.Tensor, t: float) -> torch.Tensor:
        return ((1 - self.sigma_min) * x_t - x_0 ) / (self.sigma_min + (1 - self.sigma_min) * t)

    def _get_model_gt(self, x_1: torch.Tensor, t: float, x_0: torch.Tensor = None):
        # TODO: Implement this method
        pass
        # """
        # Get ground truth for training.
        # Args:
        #     x_1: Target endpoint
        #     t: Time point
        #     noise: Initial noise to use as x_0
        # """
        # x_t = (1 - t) * x_0 + t * x_1
        # v = self._compute_velocity(x_t, x_0, t)
        # eps = x_t + (1 - t) * v  # Convert velocity to noise
        # return x_t, eps, v

    def _v_to_xstart_eps(self, x_t: torch.Tensor, t: float, v: torch.Tensor):
        """Convert velocity to x_0 and noise predictions"""
        eps = x_t + (1 - t) * v
        x_0 = self._eps_to_xstart(x_t, t, eps)
        return x_0, eps

    @torch.no_grad()
    def sample(
        self,
        model,
        x_1: torch.Tensor,
        cond: Optional[Any] = None,
        steps: int = 50,
        rescale_t: float = 1.0,
        verbose: bool = True,
        **kwargs
    ) -> Dict[str, torch.Tensor]:
        """
        Generate samples by following the flow from noise to x_1.
        Args:
            model: The model to sample from
            x_1: Target endpoint
            cond: Conditional information
            steps: Number of sampling steps
            rescale_t: Time rescaling factor
            verbose: Whether to show progress bar
            **kwargs: Additional model arguments
        Returns:
            Dictionary containing sampling trajectory and predictions
        """
        # Initialize with noise as x_0
        noise = torch.randn_like(x_1)
        current_x = noise
        
        t_seq = np.linspace(1, 0, steps + 1)
        t_seq = rescale_t * t_seq / (1 + (rescale_t - 1) * t_seq)
        t_pairs = list(zip(t_seq[:-1], t_seq[1:]))
        
        ret = edict({
            "samples": None,
            "pred_x_t": [],
            "pred_x_0": []
        })
        
        for t, t_prev in tqdm(t_pairs, desc="Sampling", disable=not verbose):
            out = self.sample_once(model, current_x, t, t_prev, cond, **kwargs)
            current_x = out.pred_x_prev
            ret.pred_x_t.append(out.pred_x_prev)
            ret.pred_x_0.append(out.pred_x_0)
            
        ret.samples = current_x
        return ret

    def sample_once(
        self,
        model,
        x_t: torch.Tensor,
        t: float,
        t_prev: float,
        cond: Optional[Any] = None,
        **kwargs
    ) -> Dict:
        """
        Sample x_{t-1} from the model using Euler method.
        Args:
            model: The model to sample from
            x_t: Current state
            t: Current time
            t_prev: Next time step
            cond: Conditional information
            **kwargs: Additional model arguments
        Returns:
            Dictionary containing predictions
        """
        pred_x_0, pred_eps, pred_v = self._get_model_prediction(model, x_t, t, cond, **kwargs)
        pred_x_prev = x_t + (t_prev - t) * pred_v
        return edict({
            "pred_x_prev": pred_x_prev,
            "pred_x_0": pred_x_0,
            "pred_eps": pred_eps
        })

class FlowEulerCfgSampler(ClassifierFreeGuidanceSamplerMixin, FlowEulerSampler):
    """
    Generate samples from a flow-matching model using Euler sampling with classifier-free guidance.
    """
    @torch.no_grad()
    def sample(
        self,
        model,
        noise,
        cond,
        neg_cond,
        steps: int = 50,
        rescale_t: float = 1.0,
        cfg_strength: float = 3.0,
        verbose: bool = True,
        **kwargs
    ):
        """
        Generate samples from the model using Euler method.
        
        Args:
            model: The model to sample from.
            noise: The initial noise tensor.
            cond: conditional information.
            neg_cond: negative conditional information.
            steps: The number of steps to sample.
            rescale_t: The rescale factor for t.
            cfg_strength: The strength of classifier-free guidance.
            verbose: If True, show a progress bar.
            **kwargs: Additional arguments for model_inference.

        Returns:
            a dict containing the following
            - 'samples': the model samples.
            - 'pred_x_t': a list of prediction of x_t.
            - 'pred_x_0': a list of prediction of x_0.
        """
        return super().sample(model, noise, cond, steps, rescale_t, verbose, neg_cond=neg_cond, cfg_strength=cfg_strength, **kwargs)


class FlowEulerGuidanceIntervalSampler(GuidanceIntervalSamplerMixin, FlowEulerSampler):
    """
    Generate samples from a flow-matching model using Euler sampling with classifier-free guidance and interval.
    """
    @torch.no_grad()
    def sample(
        self,
        model,
        noise,
        cond,
        neg_cond,
        steps: int = 50,
        rescale_t: float = 1.0,
        cfg_strength: float = 3.0,
        cfg_interval: Tuple[float, float] = (0.0, 1.0),
        verbose: bool = True,
        **kwargs
    ):
        """
        Generate samples from the model using Euler method.
        
        Args:
            model: The model to sample from.
            noise: The initial noise tensor.
            cond: conditional information.
            neg_cond: negative conditional information.
            steps: The number of steps to sample.
            rescale_t: The rescale factor for t.
            cfg_strength: The strength of classifier-free guidance.
            cfg_interval: The interval for classifier-free guidance.
            verbose: If True, show a progress bar.
            **kwargs: Additional arguments for model_inference.

        Returns:
            a dict containing the following
            - 'samples': the model samples.
            - 'pred_x_t': a list of prediction of x_t.
            - 'pred_x_0': a list of prediction of x_0.
        """
        return super().sample(model, noise, cond, steps, rescale_t, verbose, neg_cond=neg_cond, cfg_strength=cfg_strength, cfg_interval=cfg_interval, **kwargs)

    @torch.no_grad()
    def sample_featurevolume(
        self,
        model,
        cond_model,
        noise,
        cond,
        neg_cond,
        steps: int = 50,
        rescale_t: float = 1.0,
        cfg_strength: float = 3.0,
        cfg_interval: Tuple[float, float] = (0.0, 1.0),
        verbose: bool = True,
        **kwargs
    ):
        """
        Generate samples from the model using Euler method.
        
        Args:
            model: The model to sample from.
            noise: The initial noise tensor.
            cond: conditional information.
            neg_cond: negative conditional information.
            steps: The number of steps to sample.
            rescale_t: The rescale factor for t.
            cfg_strength: The strength of classifier-free guidance.
            cfg_interval: The interval for classifier-free guidance.
            verbose: If True, show a progress bar.
            **kwargs: Additional arguments for model_inference.

        Returns:
            a dict containing the following
            - 'samples': the model samples.
            - 'pred_x_t': a list of prediction of x_t.
            - 'pred_x_0': a list of prediction of x_0.
        """
        return super().sample_featurevolume(model, cond_model, noise, cond, steps, rescale_t, verbose, neg_cond=neg_cond, cfg_strength=cfg_strength, cfg_interval=cfg_interval, **kwargs)