File size: 28,205 Bytes
95ba5bc
 
 
 
 
 
 
 
 
 
2464d06
 
95ba5bc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e847c29
95ba5bc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2464d06
95ba5bc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
import torch
import torch.nn.functional as F
import numpy as np
import math

from src import utils
from src.egnn import Dynamics
from src.noise import GammaNetwork, PredefinedNoiseSchedule
from typing import Union

from tqdm import tqdm

from pdb import set_trace


class EDM(torch.nn.Module):
    def __init__(
            self,
            dynamics: Union[Dynamics],
            in_node_nf: int,
            n_dims: int,
            timesteps: int = 1000,
            noise_schedule='learned',
            noise_precision=1e-4,
            loss_type='vlb',
            norm_values=(1., 1., 1.),
            norm_biases=(None, 0., 0.),
    ):
        super().__init__()
        if noise_schedule == 'learned':
            assert loss_type == 'vlb', 'A noise schedule can only be learned with a vlb objective'
            self.gamma = GammaNetwork()
        else:
            self.gamma = PredefinedNoiseSchedule(noise_schedule, timesteps=timesteps, precision=noise_precision)

        self.dynamics = dynamics
        self.in_node_nf = in_node_nf
        self.n_dims = n_dims
        self.T = timesteps
        self.norm_values = norm_values
        self.norm_biases = norm_biases

    def forward(self, x, h, node_mask, fragment_mask, linker_mask, edge_mask, context=None):
        # Normalization and concatenation
        x, h = self.normalize(x, h)
        xh = torch.cat([x, h], dim=2)

        # Volume change loss term
        delta_log_px = self.delta_log_px(linker_mask).mean()

        # Sample t
        t_int = torch.randint(0, self.T + 1, size=(x.size(0), 1), device=x.device).float()
        s_int = t_int - 1
        t = t_int / self.T
        s = s_int / self.T

        # Masks for t=0 and t>0
        t_is_zero = (t_int == 0).squeeze().float()
        t_is_not_zero = 1 - t_is_zero

        # Compute gamma_t and gamma_s according to the noise schedule
        gamma_t = self.inflate_batch_array(self.gamma(t), x)
        gamma_s = self.inflate_batch_array(self.gamma(s), x)

        # Compute alpha_t and sigma_t from gamma
        alpha_t = self.alpha(gamma_t, x)
        sigma_t = self.sigma(gamma_t, x)

        # Sample noise
        # Note: only for linker
        eps_t = self.sample_combined_position_feature_noise(n_samples=x.size(0), n_nodes=x.size(1), mask=linker_mask)

        # Sample z_t given x, h for timestep t, from q(z_t | x, h)
        # Note: keep fragments unchanged
        z_t = alpha_t * xh + sigma_t * eps_t
        z_t = xh * fragment_mask + z_t * linker_mask

        # Neural net prediction
        eps_t_hat = self.dynamics.forward(
            xh=z_t,
            t=t,
            node_mask=node_mask,
            linker_mask=linker_mask,
            context=context,
            edge_mask=edge_mask,
        )
        eps_t_hat = eps_t_hat * linker_mask

        # Computing basic error (further used for computing NLL and L2-loss)
        error_t = self.sum_except_batch((eps_t - eps_t_hat) ** 2)

        # Computing L2-loss for t>0
        normalization = (self.n_dims + self.in_node_nf) * self.numbers_of_nodes(linker_mask)
        l2_loss = error_t / normalization
        l2_loss = l2_loss.mean()

        # The KL between q(z_T | x) and p(z_T) = Normal(0, 1) (should be close to zero)
        kl_prior = self.kl_prior(xh, linker_mask).mean()

        # Computing NLL middle term
        SNR_weight = (self.SNR(gamma_s - gamma_t) - 1).squeeze(1).squeeze(1)
        loss_term_t = self.T * 0.5 * SNR_weight * error_t
        loss_term_t = (loss_term_t * t_is_not_zero).sum() / t_is_not_zero.sum()

        # Computing noise returned by dynamics
        noise = torch.norm(eps_t_hat, dim=[1, 2])
        noise_t = (noise * t_is_not_zero).sum() / t_is_not_zero.sum()

        if t_is_zero.sum() > 0:
            # The _constants_ depending on sigma_0 from the
            # cross entropy term E_q(z0 | x) [log p(x | z0)]
            neg_log_constants = -self.log_constant_of_p_x_given_z0(x, linker_mask)

            # Computes the L_0 term (even if gamma_t is not actually gamma_0)
            # and selected only relevant via masking
            loss_term_0 = -self.log_p_xh_given_z0_without_constants(h, z_t, gamma_t, eps_t, eps_t_hat, linker_mask)
            loss_term_0 = loss_term_0 + neg_log_constants
            loss_term_0 = (loss_term_0 * t_is_zero).sum() / t_is_zero.sum()

            # Computing noise returned by dynamics
            noise_0 = (noise * t_is_zero).sum() / t_is_zero.sum()
        else:
            loss_term_0 = 0.
            noise_0 = 0.

        return delta_log_px, kl_prior, loss_term_t, loss_term_0, l2_loss, noise_t, noise_0

    @torch.no_grad()
    def sample_chain(self, x, h, node_mask, fragment_mask, linker_mask, edge_mask, context, keep_frames=None):
        n_samples = x.size(0)
        n_nodes = x.size(1)

        # Normalization and concatenation
        x, h, = self.normalize(x, h)
        xh = torch.cat([x, h], dim=2)

        # Initial linker sampling from N(0, I)
        z = self.sample_combined_position_feature_noise(n_samples, n_nodes, mask=linker_mask)
        z = xh * fragment_mask + z * linker_mask

        if keep_frames is None:
            keep_frames = self.T
        else:
            assert keep_frames <= self.T
        chain = torch.zeros((keep_frames,) + z.size(), device=z.device)

        # Sample p(z_s | z_t)
        for s in tqdm(reversed(range(0, self.T)), total=self.T):
            s_array = torch.full((n_samples, 1), fill_value=s, device=z.device)
            t_array = s_array + 1
            s_array = s_array / self.T
            t_array = t_array / self.T

            z = self.sample_p_zs_given_zt_only_linker(
                s=s_array,
                t=t_array,
                z_t=z,
                node_mask=node_mask,
                fragment_mask=fragment_mask,
                linker_mask=linker_mask,
                edge_mask=edge_mask,
                context=context,
            )
            write_index = (s * keep_frames) // self.T
            chain[write_index] = self.unnormalize_z(z)

        # Finally sample p(x, h | z_0)
        x, h = self.sample_p_xh_given_z0_only_linker(
            z_0=z,
            node_mask=node_mask,
            fragment_mask=fragment_mask,
            linker_mask=linker_mask,
            edge_mask=edge_mask,
            context=context,
        )
        chain[0] = torch.cat([x, h], dim=2)

        return chain

    def sample_p_zs_given_zt_only_linker(self, s, t, z_t, node_mask, fragment_mask, linker_mask, edge_mask, context):
        """Samples from zs ~ p(zs | zt). Only used during sampling. Samples only linker features and coords"""
        gamma_s = self.gamma(s)
        gamma_t = self.gamma(t)

        sigma2_t_given_s, sigma_t_given_s, alpha_t_given_s = self.sigma_and_alpha_t_given_s(gamma_t, gamma_s, z_t)
        sigma_s = self.sigma(gamma_s, target_tensor=z_t)
        sigma_t = self.sigma(gamma_t, target_tensor=z_t)

        # Neural net prediction.
        eps_hat = self.dynamics.forward(
            xh=z_t,
            t=t,
            node_mask=node_mask,
            linker_mask=linker_mask,
            context=context,
            edge_mask=edge_mask,
        )
        eps_hat = eps_hat * linker_mask

        # Compute mu for p(z_s | z_t)
        mu = z_t / alpha_t_given_s - (sigma2_t_given_s / alpha_t_given_s / sigma_t) * eps_hat

        # Compute sigma for p(z_s | z_t)
        sigma = sigma_t_given_s * sigma_s / sigma_t

        # Sample z_s given the parameters derived from zt
        z_s = self.sample_normal(mu, sigma, linker_mask)
        z_s = z_t * fragment_mask + z_s * linker_mask

        return z_s

    def sample_p_xh_given_z0_only_linker(self, z_0, node_mask, fragment_mask, linker_mask, edge_mask, context):
        """Samples x ~ p(x|z0). Samples only linker features and coords"""
        zeros = torch.zeros(size=(z_0.size(0), 1), device=z_0.device)
        gamma_0 = self.gamma(zeros)

        # Computes sqrt(sigma_0^2 / alpha_0^2)
        sigma_x = self.SNR(-0.5 * gamma_0).unsqueeze(1)
        eps_hat = self.dynamics.forward(
            t=zeros,
            xh=z_0,
            node_mask=node_mask,
            linker_mask=linker_mask,
            edge_mask=edge_mask,
            context=context
        )
        eps_hat = eps_hat * linker_mask

        mu_x = self.compute_x_pred(eps_t=eps_hat, z_t=z_0, gamma_t=gamma_0)
        xh = self.sample_normal(mu=mu_x, sigma=sigma_x, node_mask=linker_mask)
        xh = z_0 * fragment_mask + xh * linker_mask

        x, h = xh[:, :, :self.n_dims], xh[:, :, self.n_dims:]
        x, h = self.unnormalize(x, h)
        h = F.one_hot(torch.argmax(h, dim=2), self.in_node_nf) * node_mask

        return x, h

    def compute_x_pred(self, eps_t, z_t, gamma_t):
        """Computes x_pred, i.e. the most likely prediction of x."""
        sigma_t = self.sigma(gamma_t, target_tensor=eps_t)
        alpha_t = self.alpha(gamma_t, target_tensor=eps_t)
        x_pred = 1. / alpha_t * (z_t - sigma_t * eps_t)
        return x_pred

    def kl_prior(self, xh, mask):
        """
        Computes the KL between q(z1 | x) and the prior p(z1) = Normal(0, 1).
        This is essentially a lot of work for something that is in practice negligible in the loss.
        However, you compute it so that you see it when you've made a mistake in your noise schedule.
        """
        # Compute the last alpha value, alpha_T
        ones = torch.ones((xh.size(0), 1), device=xh.device)
        gamma_T = self.gamma(ones)
        alpha_T = self.alpha(gamma_T, xh)

        # Compute means
        mu_T = alpha_T * xh
        mu_T_x, mu_T_h = mu_T[:, :, :self.n_dims], mu_T[:, :, self.n_dims:]

        # Compute standard deviations (only batch axis for x-part, inflated for h-part)
        sigma_T_x = self.sigma(gamma_T, mu_T_x).view(-1)  # Remove inflate, only keep batch dimension for x-part
        sigma_T_h = self.sigma(gamma_T, mu_T_h)

        # Compute KL for h-part
        zeros, ones = torch.zeros_like(mu_T_h), torch.ones_like(sigma_T_h)
        kl_distance_h = self.gaussian_kl(mu_T_h, sigma_T_h, zeros, ones)

        # Compute KL for x-part
        zeros, ones = torch.zeros_like(mu_T_x), torch.ones_like(sigma_T_x)
        d = self.dimensionality(mask)
        kl_distance_x = self.gaussian_kl_for_dimension(mu_T_x, sigma_T_x, zeros, ones, d=d)

        return kl_distance_x + kl_distance_h

    def log_constant_of_p_x_given_z0(self, x, mask):
        batch_size = x.size(0)
        degrees_of_freedom_x = self.dimensionality(mask)
        zeros = torch.zeros((batch_size, 1), device=x.device)
        gamma_0 = self.gamma(zeros)

        # Recall that sigma_x = sqrt(sigma_0^2 / alpha_0^2) = SNR(-0.5 gamma_0)
        log_sigma_x = 0.5 * gamma_0.view(batch_size)

        return degrees_of_freedom_x * (- log_sigma_x - 0.5 * np.log(2 * np.pi))

    def log_p_xh_given_z0_without_constants(self, h, z_0, gamma_0, eps, eps_hat, mask, epsilon=1e-10):
        # Discrete properties are predicted directly from z_0
        z_h = z_0[:, :, self.n_dims:]

        # Take only part over x
        eps_x = eps[:, :, :self.n_dims]
        eps_hat_x = eps_hat[:, :, :self.n_dims]

        # Compute sigma_0 and rescale to the integer scale of the data
        sigma_0 = self.sigma(gamma_0, target_tensor=z_0) * self.norm_values[1]

        # Computes the error for the distribution N(x | 1 / alpha_0 z_0 + sigma_0/alpha_0 eps_0, sigma_0 / alpha_0),
        # the weighting in the epsilon parametrization is exactly '1'
        log_p_x_given_z_without_constants = -0.5 * self.sum_except_batch((eps_x - eps_hat_x) ** 2)

        # Categorical features
        # Compute delta indicator masks
        h = h * self.norm_values[1] + self.norm_biases[1]
        estimated_h = z_h * self.norm_values[1] + self.norm_biases[1]

        # Centered h_cat around 1, since onehot encoded
        centered_h = estimated_h - 1

        # Compute integrals from 0.5 to 1.5 of the normal distribution
        # N(mean=centered_h_cat, stdev=sigma_0_cat)
        log_p_h_proportional = torch.log(
            self.cdf_standard_gaussian((centered_h + 0.5) / sigma_0) -
            self.cdf_standard_gaussian((centered_h - 0.5) / sigma_0) +
            epsilon
        )

        # Normalize the distribution over the categories
        log_Z = torch.logsumexp(log_p_h_proportional, dim=2, keepdim=True)
        log_probabilities = log_p_h_proportional - log_Z

        # Select the log_prob of the current category using the onehot representation
        log_p_h_given_z = self.sum_except_batch(log_probabilities * h * mask)

        # Combine log probabilities for x and h
        log_p_xh_given_z = log_p_x_given_z_without_constants + log_p_h_given_z

        return log_p_xh_given_z

    def sample_combined_position_feature_noise(self, n_samples, n_nodes, mask):
        z_x = utils.sample_gaussian_with_mask(
            size=(n_samples, n_nodes, self.n_dims),
            device=mask.device,
            node_mask=mask
        )
        z_h = utils.sample_gaussian_with_mask(
            size=(n_samples, n_nodes, self.in_node_nf),
            device=mask.device,
            node_mask=mask
        )
        z = torch.cat([z_x, z_h], dim=2)
        return z

    def sample_normal(self, mu, sigma, node_mask):
        """Samples from a Normal distribution."""
        eps = self.sample_combined_position_feature_noise(mu.size(0), mu.size(1), node_mask)
        return mu + sigma * eps

    def normalize(self, x, h):
        new_x = x / self.norm_values[0]
        new_h = (h.float() - self.norm_biases[1]) / self.norm_values[1]
        return new_x, new_h

    def unnormalize(self, x, h):
        new_x = x * self.norm_values[0]
        new_h = h * self.norm_values[1] + self.norm_biases[1]
        return new_x, new_h

    def unnormalize_z(self, z):
        assert z.size(2) == self.n_dims + self.in_node_nf
        x, h = z[:, :, :self.n_dims], z[:, :, self.n_dims:]
        x, h = self.unnormalize(x, h)
        return torch.cat([x, h], dim=2)

    def delta_log_px(self, mask):
        return -self.dimensionality(mask) * np.log(self.norm_values[0])

    def dimensionality(self, mask):
        return self.numbers_of_nodes(mask) * self.n_dims

    def sigma(self, gamma, target_tensor):
        """Computes sigma given gamma."""
        return self.inflate_batch_array(torch.sqrt(torch.sigmoid(gamma)), target_tensor)

    def alpha(self, gamma, target_tensor):
        """Computes alpha given gamma."""
        return self.inflate_batch_array(torch.sqrt(torch.sigmoid(-gamma)), target_tensor)

    def SNR(self, gamma):
        """Computes signal to noise ratio (alpha^2/sigma^2) given gamma."""
        return torch.exp(-gamma)

    def sigma_and_alpha_t_given_s(self, gamma_t: torch.Tensor, gamma_s: torch.Tensor, target_tensor: torch.Tensor):
        """
        Computes sigma t given s, using gamma_t and gamma_s. Used during sampling.

        These are defined as:
            alpha t given s = alpha t / alpha s,
            sigma t given s = sqrt(1 - (alpha t given s) ^2 ).
        """
        sigma2_t_given_s = self.inflate_batch_array(
            -self.expm1(self.softplus(gamma_s) - self.softplus(gamma_t)),
            target_tensor
        )

        # alpha_t_given_s = alpha_t / alpha_s
        log_alpha2_t = F.logsigmoid(-gamma_t)
        log_alpha2_s = F.logsigmoid(-gamma_s)
        log_alpha2_t_given_s = log_alpha2_t - log_alpha2_s

        alpha_t_given_s = torch.exp(0.5 * log_alpha2_t_given_s)
        alpha_t_given_s = self.inflate_batch_array(alpha_t_given_s, target_tensor)
        sigma_t_given_s = torch.sqrt(sigma2_t_given_s)

        return sigma2_t_given_s, sigma_t_given_s, alpha_t_given_s

    @staticmethod
    def numbers_of_nodes(mask):
        return torch.sum(mask.squeeze(2), dim=1)

    @staticmethod
    def inflate_batch_array(array, target):
        """
        Inflates the batch array (array) with only a single axis (i.e. shape = (batch_size,),
        or possibly more empty axes (i.e. shape (batch_size, 1, ..., 1)) to match the target shape.
        """
        target_shape = (array.size(0),) + (1,) * (len(target.size()) - 1)
        return array.view(target_shape)

    @staticmethod
    def sum_except_batch(x):
        return x.view(x.size(0), -1).sum(-1)

    @staticmethod
    def expm1(x: torch.Tensor) -> torch.Tensor:
        return torch.expm1(x)

    @staticmethod
    def softplus(x: torch.Tensor) -> torch.Tensor:
        return F.softplus(x)

    @staticmethod
    def cdf_standard_gaussian(x):
        return 0.5 * (1. + torch.erf(x / math.sqrt(2)))

    @staticmethod
    def gaussian_kl(q_mu, q_sigma, p_mu, p_sigma):
        """
        Computes the KL distance between two normal distributions.
        Args:
            q_mu: Mean of distribution q.
            q_sigma: Standard deviation of distribution q.
            p_mu: Mean of distribution p.
            p_sigma: Standard deviation of distribution p.
        Returns:
            The KL distance, summed over all dimensions except the batch dim.
        """
        kl = torch.log(p_sigma / q_sigma) + 0.5 * (q_sigma ** 2 + (q_mu - p_mu) ** 2) / (p_sigma ** 2) - 0.5
        return EDM.sum_except_batch(kl)

    @staticmethod
    def gaussian_kl_for_dimension(q_mu, q_sigma, p_mu, p_sigma, d):
        """
        Computes the KL distance between two normal distributions taking the dimension into account.
        Args:
            q_mu: Mean of distribution q.
            q_sigma: Standard deviation of distribution q.
            p_mu: Mean of distribution p.
            p_sigma: Standard deviation of distribution p.
            d: dimension
        Returns:
            The KL distance, summed over all dimensions except the batch dim.
        """
        mu_norm_2 = EDM.sum_except_batch((q_mu - p_mu) ** 2)
        return d * torch.log(p_sigma / q_sigma) + 0.5 * (d * q_sigma ** 2 + mu_norm_2) / (p_sigma ** 2) - 0.5 * d


class InpaintingEDM(EDM):
    def forward(self, x, h, node_mask, fragment_mask, linker_mask, edge_mask, context=None):
        # Normalization and concatenation
        x, h = self.normalize(x, h)
        xh = torch.cat([x, h], dim=2)

        # Volume change loss term
        delta_log_px = self.delta_log_px(node_mask).mean()

        # Sample t
        t_int = torch.randint(0, self.T + 1, size=(x.size(0), 1), device=x.device).float()
        s_int = t_int - 1
        t = t_int / self.T
        s = s_int / self.T

        # Masks for t=0 and t>0
        t_is_zero = (t_int == 0).squeeze().float()
        t_is_not_zero = 1 - t_is_zero

        # Compute gamma_t and gamma_s according to the noise schedule
        gamma_t = self.inflate_batch_array(self.gamma(t), x)
        gamma_s = self.inflate_batch_array(self.gamma(s), x)

        # Compute alpha_t and sigma_t from gamma
        alpha_t = self.alpha(gamma_t, x)
        sigma_t = self.sigma(gamma_t, x)

        # Sample noise
        eps_t = self.sample_combined_position_feature_noise(n_samples=x.size(0), n_nodes=x.size(1), mask=node_mask)

        # Sample z_t given x, h for timestep t, from q(z_t | x, h)
        # Note: keep fragments unchanged
        z_t = alpha_t * xh + sigma_t * eps_t

        # Neural net prediction
        eps_t_hat = self.dynamics.forward(
            xh=z_t,
            t=t,
            node_mask=node_mask,
            linker_mask=None,
            context=context,
            edge_mask=edge_mask,
        )

        # Computing basic error (further used for computing NLL and L2-loss)
        error_t = self.sum_except_batch((eps_t - eps_t_hat) ** 2)

        # Computing L2-loss for t>0
        normalization = (self.n_dims + self.in_node_nf) * self.numbers_of_nodes(node_mask)
        l2_loss = error_t / normalization
        l2_loss = l2_loss.mean()

        # The KL between q(z_T | x) and p(z_T) = Normal(0, 1) (should be close to zero)
        kl_prior = self.kl_prior(xh, node_mask).mean()

        # Computing NLL middle term
        SNR_weight = (self.SNR(gamma_s - gamma_t) - 1).squeeze(1).squeeze(1)
        loss_term_t = self.T * 0.5 * SNR_weight * error_t
        loss_term_t = (loss_term_t * t_is_not_zero).sum() / t_is_not_zero.sum()

        # Computing noise returned by dynamics
        noise = torch.norm(eps_t_hat, dim=[1, 2])
        noise_t = (noise * t_is_not_zero).sum() / t_is_not_zero.sum()

        if t_is_zero.sum() > 0:
            # The _constants_ depending on sigma_0 from the
            # cross entropy term E_q(z0 | x) [log p(x | z0)]
            neg_log_constants = -self.log_constant_of_p_x_given_z0(x, node_mask)

            # Computes the L_0 term (even if gamma_t is not actually gamma_0)
            # and selected only relevant via masking
            loss_term_0 = -self.log_p_xh_given_z0_without_constants(h, z_t, gamma_t, eps_t, eps_t_hat, node_mask)
            loss_term_0 = loss_term_0 + neg_log_constants
            loss_term_0 = (loss_term_0 * t_is_zero).sum() / t_is_zero.sum()

            # Computing noise returned by dynamics
            noise_0 = (noise * t_is_zero).sum() / t_is_zero.sum()
        else:
            loss_term_0 = 0.
            noise_0 = 0.

        return delta_log_px, kl_prior, loss_term_t, loss_term_0, l2_loss, noise_t, noise_0

    @torch.no_grad()
    def sample_chain(self, x, h, node_mask, edge_mask, fragment_mask, linker_mask, context, keep_frames=None):
        n_samples = x.size(0)
        n_nodes = x.size(1)

        # Normalization and concatenation
        x, h, = self.normalize(x, h)
        xh = torch.cat([x, h], dim=2)

        # Sampling initial noise
        z = self.sample_combined_position_feature_noise(n_samples, n_nodes, node_mask)

        if keep_frames is None:
            keep_frames = self.T
        else:
            assert keep_frames <= self.T
        chain = torch.zeros((keep_frames,) + z.size(), device=z.device)

        # Sample p(z_s | z_t)
        for s in tqdm(reversed(range(0, self.T)), total=self.T):
            s_array = torch.full((n_samples, 1), fill_value=s, device=z.device)
            t_array = s_array + 1
            s_array = s_array / self.T
            t_array = t_array / self.T

            z_linker_only_sampled = self.sample_p_zs_given_zt(
                s=s_array,
                t=t_array,
                z_t=z,
                node_mask=node_mask,
                edge_mask=edge_mask,
                context=context,
            )
            z_fragments_only_sampled = self.sample_q_zs_given_zt_and_x(
                s=s_array,
                t=t_array,
                z_t=z,
                x=xh * fragment_mask,
                node_mask=fragment_mask,
            )
            z = z_linker_only_sampled * linker_mask + z_fragments_only_sampled * fragment_mask

            # Project down to avoid numerical runaway of the center of gravity
            z_x = utils.remove_mean_with_mask(z[:, :, :self.n_dims], node_mask)
            z_h = z[:, :, self.n_dims:]
            z = torch.cat([z_x, z_h], dim=2)

            # Saving step to the chain
            write_index = (s * keep_frames) // self.T
            chain[write_index] = self.unnormalize_z(z)

        # Finally sample p(x, h | z_0)
        x_out_linker, h_out_linker = self.sample_p_xh_given_z0(
            z_0=z,
            node_mask=node_mask,
            edge_mask=edge_mask,
            context=context,
        )
        x_out_fragments, h_out_fragments = self.sample_q_xh_given_z0_and_x(z_0=z, node_mask=node_mask)

        xh_out_linker = torch.cat([x_out_linker, h_out_linker], dim=2)
        xh_out_fragments = torch.cat([x_out_fragments, h_out_fragments], dim=2)
        xh_out = xh_out_linker * linker_mask + xh_out_fragments * fragment_mask

        # Overwrite last frame with the resulting x and h
        chain[0] = xh_out

        return chain

    def sample_p_zs_given_zt(self, s, t, z_t, node_mask, edge_mask, context):
        """Samples from zs ~ p(zs | zt). Only used during sampling"""
        gamma_s = self.gamma(s)
        gamma_t = self.gamma(t)
        sigma2_t_given_s, sigma_t_given_s, alpha_t_given_s = self.sigma_and_alpha_t_given_s(gamma_t, gamma_s, z_t)

        sigma_s = self.sigma(gamma_s, target_tensor=z_t)
        sigma_t = self.sigma(gamma_t, target_tensor=z_t)

        # Neural net prediction.
        eps_hat = self.dynamics.forward(
            xh=z_t,
            t=t,
            node_mask=node_mask,
            linker_mask=None,
            edge_mask=edge_mask,
            context=context
        )

        # Checking that epsilon is centered around linker COM
        utils.assert_mean_zero_with_mask(eps_hat[:, :, :self.n_dims], node_mask)

        # Compute mu for p(z_s | z_t)
        mu = z_t / alpha_t_given_s - (sigma2_t_given_s / alpha_t_given_s / sigma_t) * eps_hat

        # Compute sigma for p(z_s | z_t)
        sigma = sigma_t_given_s * sigma_s / sigma_t

        # Sample z_s given the parameters derived from z_t
        z_s = self.sample_normal(mu, sigma, node_mask)
        return z_s

    def sample_q_zs_given_zt_and_x(self, s, t, z_t, x, node_mask):
        """Samples from zs ~ q(zs | zt, x). Only used during sampling. Samples only linker features and coords"""
        gamma_s = self.gamma(s)
        gamma_t = self.gamma(t)
        sigma2_t_given_s, sigma_t_given_s, alpha_t_given_s = self.sigma_and_alpha_t_given_s(gamma_t, gamma_s, z_t)

        sigma_s = self.sigma(gamma_s, target_tensor=z_t)
        sigma_t = self.sigma(gamma_t, target_tensor=z_t)
        alpha_s = self.alpha(gamma_s, x)

        mu = (
            alpha_t_given_s * (sigma_s ** 2) / (sigma_t ** 2) * z_t +
            alpha_s * sigma2_t_given_s / (sigma_t ** 2) * x
        )

        # Compute sigma for p(zs | zt)
        sigma = sigma_t_given_s * sigma_s / sigma_t

        # Sample zs given the parameters derived from zt
        z_s = self.sample_normal(mu, sigma, node_mask)
        return z_s

    def sample_p_xh_given_z0(self, z_0, node_mask, edge_mask, context):
        """Samples x ~ p(x|z0). Samples only linker features and coords"""
        zeros = torch.zeros(size=(z_0.size(0), 1), device=z_0.device)
        gamma_0 = self.gamma(zeros)

        # Computes sqrt(sigma_0^2 / alpha_0^2)
        sigma_x = self.SNR(-0.5 * gamma_0).unsqueeze(1)
        eps_hat = self.dynamics.forward(
            xh=z_0,
            t=zeros,
            node_mask=node_mask,
            linker_mask=None,
            edge_mask=edge_mask,
            context=context
        )
        utils.assert_mean_zero_with_mask(eps_hat[:, :, :self.n_dims], node_mask)

        mu_x = self.compute_x_pred(eps_hat, z_0, gamma_0)
        xh = self.sample_normal(mu=mu_x, sigma=sigma_x, node_mask=node_mask)

        x, h = xh[:, :, :self.n_dims], xh[:, :, self.n_dims:]
        x, h = self.unnormalize(x, h)
        h = F.one_hot(torch.argmax(h, dim=2), self.in_node_nf) * node_mask

        return x, h

    def sample_q_xh_given_z0_and_x(self, z_0, node_mask):
        """Samples x ~ q(x|z0). Samples only linker features and coords"""
        zeros = torch.zeros(size=(z_0.size(0), 1), device=z_0.device)
        gamma_0 = self.gamma(zeros)
        alpha_0 = self.alpha(gamma_0, z_0)
        sigma_0 = self.sigma(gamma_0, z_0)

        eps = self.sample_combined_position_feature_noise(z_0.size(0), z_0.size(1), node_mask)

        xh = (1 / alpha_0) * z_0 - (sigma_0 / alpha_0) * eps

        x, h = xh[:, :, :self.n_dims], xh[:, :, self.n_dims:]
        x, h = self.unnormalize(x, h)
        h = F.one_hot(torch.argmax(h, dim=2), self.in_node_nf) * node_mask

        return x, h

    def sample_combined_position_feature_noise(self, n_samples, n_nodes, mask):
        z_x = utils.sample_center_gravity_zero_gaussian_with_mask(
            size=(n_samples, n_nodes, self.n_dims),
            device=mask.device,
            node_mask=mask
        )
        z_h = utils.sample_gaussian_with_mask(
            size=(n_samples, n_nodes, self.in_node_nf),
            device=mask.device,
            node_mask=mask
        )
        z = torch.cat([z_x, z_h], dim=2)
        return z

    def dimensionality(self, mask):
        return (self.numbers_of_nodes(mask) - 1) * self.n_dims