File size: 20,392 Bytes
37171a0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""
WorldModel: Joint Environment-Genome Embedding for Productivity Prediction.

Architecture:
  - Encoder_E: Environment MLP (env_dim -> 128 -> latent_dim)
  - Encoder_P: PFAM Module MLP (pfam_dim -> 256 -> 128 -> latent_dim)
  - Predictor: Productivity head (latent_dim -> 64 -> 3)

Training:
  Loss = VICReg(z_env, z_pfam) + alpha * MSE(Predictor(z_env), bio_targets)

Inference (environment-only):
  env -> Encoder_E -> z_env -> Predictor -> productivity (chl-a, POC, NFLH)

Designed for:
  - 1,810 ocean samples with 24 environmental variables, 20 PFAM modules, 3 bio targets
  - Spatial block CV (leave-one-basin-out)
  - VICReg non-contrastive alignment (Bardes et al., ICLR 2022)

Author: World Model RALPH Loop
Date: 2026-01-27
"""

import sys
import os
import torch
import torch.nn as nn

# Import VICReg loss from sibling module
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from vicreg_loss import VICRegLoss


class EncoderE(nn.Module):
    """Environment encoder MLP.

    Architecture: env_dim -> 128 -> latent_dim
    Each layer: Linear -> BatchNorm1d -> ReLU -> Dropout

    Parameters
    ----------
    env_dim : int
        Number of environment input features (default 24).
    latent_dim : int
        Latent embedding dimension (default 16).
    dropout : float
        Dropout probability (default 0.3).
    """

    def __init__(self, env_dim=24, latent_dim=16, dropout=0.3):
        super().__init__()
        self.env_dim = env_dim
        self.latent_dim = latent_dim

        self.layers = nn.Sequential(
            # Block 1: env_dim -> 128
            nn.Linear(env_dim, 128),
            nn.BatchNorm1d(128),
            nn.ReLU(),
            nn.Dropout(dropout),
            # Block 2: 128 -> latent_dim
            nn.Linear(128, latent_dim),
            nn.BatchNorm1d(latent_dim),
            nn.ReLU(),
            nn.Dropout(dropout),
        )

    def forward(self, x):
        """
        Parameters
        ----------
        x : torch.Tensor, shape (N, env_dim)
            Standardized environment features.

        Returns
        -------
        z_env : torch.Tensor, shape (N, latent_dim)
            Environment embedding.
        """
        return self.layers(x)


class EncoderP(nn.Module):
    """PFAM module encoder MLP.

    Architecture: pfam_dim -> 256 -> 128 -> latent_dim
    Each layer: Linear -> BatchNorm1d -> ReLU -> Dropout
    Deeper than EncoderE because PFAM modules encode richer combinatorial
    information.

    Parameters
    ----------
    pfam_dim : int
        Number of PFAM module input features (default 20).
    latent_dim : int
        Latent embedding dimension (default 16).
    dropout : float
        Dropout probability (default 0.3).
    """

    def __init__(self, pfam_dim=20, latent_dim=16, dropout=0.3):
        super().__init__()
        self.pfam_dim = pfam_dim
        self.latent_dim = latent_dim

        self.layers = nn.Sequential(
            # Block 1: pfam_dim -> 256
            nn.Linear(pfam_dim, 256),
            nn.BatchNorm1d(256),
            nn.ReLU(),
            nn.Dropout(dropout),
            # Block 2: 256 -> 128
            nn.Linear(256, 128),
            nn.BatchNorm1d(128),
            nn.ReLU(),
            nn.Dropout(dropout),
            # Block 3: 128 -> latent_dim
            nn.Linear(128, latent_dim),
            nn.BatchNorm1d(latent_dim),
            nn.ReLU(),
            nn.Dropout(dropout),
        )

    def forward(self, x):
        """
        Parameters
        ----------
        x : torch.Tensor, shape (N, pfam_dim)
            Standardized PFAM module features.

        Returns
        -------
        z_pfam : torch.Tensor, shape (N, latent_dim)
            PFAM module embedding.
        """
        return self.layers(x)


class Predictor(nn.Module):
    """Productivity prediction head.

    Architecture: input_dim -> 64 -> bio_dim
    Simple head: Linear -> ReLU -> Linear (no BatchNorm/Dropout).

    Parameters
    ----------
    input_dim : int
        Input dimension (latent_dim for env-only, 2*latent_dim for joint).
    bio_dim : int
        Number of bio-response targets (default 3: chl-a, POC, NFLH).
    """

    def __init__(self, input_dim=16, bio_dim=3):
        super().__init__()
        self.input_dim = input_dim
        self.bio_dim = bio_dim

        self.layers = nn.Sequential(
            nn.Linear(input_dim, 64),
            nn.ReLU(),
            nn.Linear(64, bio_dim),
        )

    def forward(self, z):
        """
        Parameters
        ----------
        z : torch.Tensor, shape (N, input_dim)
            Latent embedding (z_env or [z_env, z_pfam]).

        Returns
        -------
        y_pred : torch.Tensor, shape (N, bio_dim)
            Predicted productivity (chl-a, POC, NFLH).
        """
        return self.layers(z)


class WorldModel(nn.Module):
    """Joint Environment-Genome Embedding Model.

    Wraps Encoder_E, Encoder_P, Predictor, and VICRegLoss into a single
    module for training and inference.

    Training flow:
        env -> Encoder_E -> z_env --|
                                     |--> VICReg(z_env, z_pfam)
        pfam -> Encoder_P -> z_pfam--|
                                     |--> Predictor(z_env) -> y_pred
                                          MSE(y_pred, bio_targets)

    Inference flow (env-only):
        env -> Encoder_E -> z_env -> Predictor -> productivity

    Parameters
    ----------
    env_dim : int
        Number of environment input features (default 24).
    pfam_dim : int
        Number of PFAM module input features (default 20).
    bio_dim : int
        Number of bio-response targets (default 3).
    latent_dim : int
        Latent embedding dimension (default 16).
    dropout : float
        Dropout probability (default 0.3).
    lambda_inv : float
        VICReg invariance weight (default 25.0).
    lambda_var : float
        VICReg variance weight (default 25.0).
    lambda_cov : float
        VICReg covariance weight (default 1.0).
    pred_alpha : float
        Weight for productivity prediction loss (default 1.0).
    """

    def __init__(self, env_dim=24, pfam_dim=20, bio_dim=3, latent_dim=16,
                 dropout=0.3, lambda_inv=25.0, lambda_var=25.0,
                 lambda_cov=1.0, pred_alpha=1.0):
        super().__init__()

        self.env_dim = env_dim
        self.pfam_dim = pfam_dim
        self.bio_dim = bio_dim
        self.latent_dim = latent_dim
        self.pred_alpha = pred_alpha

        # Sub-modules
        self.encoder_e = EncoderE(env_dim, latent_dim, dropout)
        self.encoder_p = EncoderP(pfam_dim, latent_dim, dropout)
        self.predictor = Predictor(latent_dim, bio_dim)
        self.vicreg = VICRegLoss(lambda_inv, lambda_var, lambda_cov)

        # Store config for serialization
        self.config = {
            'env_dim': env_dim,
            'pfam_dim': pfam_dim,
            'bio_dim': bio_dim,
            'latent_dim': latent_dim,
            'dropout': dropout,
            'lambda_inv': lambda_inv,
            'lambda_var': lambda_var,
            'lambda_cov': lambda_cov,
            'pred_alpha': pred_alpha,
        }

    def forward(self, env, pfam, bio_targets=None, bio_valid=None):
        """Full training forward pass.

        Parameters
        ----------
        env : torch.Tensor, shape (N, env_dim)
            Standardized environment features.
        pfam : torch.Tensor, shape (N, pfam_dim)
            Standardized PFAM module features.
        bio_targets : torch.Tensor or None, shape (N, bio_dim)
            Standardized bio-response targets. If None, skip pred loss.
        bio_valid : torch.Tensor or None, shape (N,)
            Boolean mask: True where all bio targets are valid.
            If None and bio_targets given, assume all valid.

        Returns
        -------
        result : dict
            'z_env': (N, latent_dim) environment embedding
            'z_pfam': (N, latent_dim) PFAM module embedding
            'y_pred': (N, bio_dim) predicted productivity
            'total_loss': scalar total loss
            'vicreg_loss': scalar VICReg loss
            'pred_loss': scalar prediction MSE loss (0 if no targets)
            'vicreg_components': dict of individual VICReg terms
        """
        # Encode both modalities
        z_env = self.encoder_e(env)
        z_pfam = self.encoder_p(pfam)

        # Predict productivity from environment embedding
        y_pred = self.predictor(z_env)

        # Compute VICReg alignment loss
        vicreg_loss, vicreg_components = self.vicreg(z_env, z_pfam)

        # Compute prediction loss (only on bio_valid samples)
        pred_loss = torch.tensor(0.0, device=env.device)
        if bio_targets is not None:
            if bio_valid is not None:
                valid_mask = bio_valid.bool()
                if valid_mask.sum() > 0:
                    pred_loss = nn.functional.mse_loss(
                        y_pred[valid_mask], bio_targets[valid_mask]
                    )
            else:
                pred_loss = nn.functional.mse_loss(y_pred, bio_targets)

        # Total loss
        total_loss = vicreg_loss + self.pred_alpha * pred_loss

        return {
            'z_env': z_env,
            'z_pfam': z_pfam,
            'y_pred': y_pred,
            'total_loss': total_loss,
            'vicreg_loss': vicreg_loss,
            'pred_loss': pred_loss,
            'vicreg_components': vicreg_components,
        }

    def encode_env(self, env):
        """Encode environment features to latent space.

        Parameters
        ----------
        env : torch.Tensor, shape (N, env_dim)

        Returns
        -------
        z_env : torch.Tensor, shape (N, latent_dim)
        """
        return self.encoder_e(env)

    def encode_pfam(self, pfam):
        """Encode PFAM module features to latent space.

        Parameters
        ----------
        pfam : torch.Tensor, shape (N, pfam_dim)

        Returns
        -------
        z_pfam : torch.Tensor, shape (N, latent_dim)
        """
        return self.encoder_p(pfam)

    def inference(self, env):
        """Environment-only inference path.

        Parameters
        ----------
        env : torch.Tensor, shape (N, env_dim)
            Standardized environment features.

        Returns
        -------
        y_pred : torch.Tensor, shape (N, bio_dim)
            Predicted productivity.
        """
        z_env = self.encoder_e(env)
        return self.predictor(z_env)

    def count_parameters(self):
        """Count total trainable parameters.

        Returns
        -------
        int
            Total number of trainable parameters.
        """
        return sum(p.numel() for p in self.parameters() if p.requires_grad)

    def count_parameters_by_component(self):
        """Count trainable parameters per sub-module.

        Returns
        -------
        dict
            {'encoder_e': int, 'encoder_p': int, 'predictor': int, 'total': int}
        """
        counts = {}
        for name, module in [('encoder_e', self.encoder_e),
                             ('encoder_p', self.encoder_p),
                             ('predictor', self.predictor)]:
            counts[name] = sum(p.numel() for p in module.parameters()
                               if p.requires_grad)
        counts['total'] = sum(counts.values())
        return counts


def self_test():
    """Run comprehensive self-tests for WorldModel. Returns True if all pass."""
    tests_passed = 0
    tests_total = 0

    def check(name, condition):
        nonlocal tests_passed, tests_total
        tests_total += 1
        if condition:
            tests_passed += 1
            print(f"  PASS: {name}")
        else:
            print(f"  FAIL: {name}")

    print("=" * 70)
    print("WorldModel Self-Tests")
    print("=" * 70)

    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    print(f"Device: {device}")

    # ── Test 1: Instantiation with default parameters ──
    print("\nTest 1: Instantiation with default parameters")
    model = WorldModel(env_dim=24, pfam_dim=20, bio_dim=3,
                       latent_dim=16, dropout=0.3).to(device)
    params = model.count_parameters()
    param_detail = model.count_parameters_by_component()
    print(f"  Total parameters: {params:,}")
    print(f"  Encoder_E: {param_detail['encoder_e']:,}")
    print(f"  Encoder_P: {param_detail['encoder_p']:,}")
    print(f"  Predictor: {param_detail['predictor']:,}")
    check("model instantiates", model is not None)
    check("total params > 0", params > 0)
    check("param counts sum correctly",
          param_detail['total'] == params)

    # ── Test 2: Forward pass shapes ──
    print("\nTest 2: Forward pass shapes")
    N = 64
    env = torch.randn(N, 24, device=device)
    pfam = torch.randn(N, 20, device=device)
    bio = torch.randn(N, 3, device=device)
    bio_valid = torch.ones(N, dtype=torch.bool, device=device)

    model.train()
    result = model(env, pfam, bio, bio_valid)
    check("z_env shape", result['z_env'].shape == (N, 16))
    check("z_pfam shape", result['z_pfam'].shape == (N, 16))
    check("y_pred shape", result['y_pred'].shape == (N, 3))
    check("total_loss is scalar", result['total_loss'].dim() == 0)
    check("vicreg_loss is scalar", result['vicreg_loss'].dim() == 0)
    check("pred_loss is scalar", result['pred_loss'].dim() == 0)
    check("vicreg_components present",
          all(k in result['vicreg_components']
              for k in ['invariance', 'variance_a', 'variance_b',
                        'covariance_a', 'covariance_b', 'total']))

    # ── Test 3: Forward without bio targets (VICReg-only mode) ──
    print("\nTest 3: Forward without bio targets (VICReg-only)")
    result_no_bio = model(env, pfam, bio_targets=None)
    check("works without bio targets", result_no_bio['total_loss'].item() > 0)
    check("pred_loss is zero", result_no_bio['pred_loss'].item() == 0.0)

    # ── Test 4: Forward with partial bio_valid mask ──
    print("\nTest 4: Forward with partial bio_valid mask")
    partial_valid = torch.zeros(N, dtype=torch.bool, device=device)
    partial_valid[:32] = True  # Only half valid
    result_partial = model(env, pfam, bio, partial_valid)
    check("works with partial bio_valid", result_partial['total_loss'].item() > 0)
    check("pred_loss computed on valid subset",
          result_partial['pred_loss'].item() >= 0)

    # Forward with all-invalid bio_valid mask
    all_invalid = torch.zeros(N, dtype=torch.bool, device=device)
    result_novalid = model(env, pfam, bio, all_invalid)
    check("works with all-invalid mask",
          result_novalid['pred_loss'].item() == 0.0)

    # ── Test 5: Gradient flow ──
    print("\nTest 5: Gradient flow")
    model.zero_grad()
    result = model(env, pfam, bio, bio_valid)
    result['total_loss'].backward()
    all_params = list(model.named_parameters())
    grad_count = sum(1 for _, p in all_params
                     if p.grad is not None and p.grad.abs().sum() > 0)
    check(f"all {len(all_params)} param tensors receive gradients",
          grad_count == len(all_params))
    no_nan = all(not torch.isnan(p.grad).any()
                 for _, p in all_params if p.grad is not None)
    check("no NaN in any gradient", no_nan)

    # ── Test 6: Inference mode (env-only) ──
    print("\nTest 6: Inference mode (env-only)")
    model.eval()
    with torch.no_grad():
        y_pred_inf = model.inference(env)
    check("inference returns correct shape", y_pred_inf.shape == (N, 3))
    check("no NaN in inference output", not torch.isnan(y_pred_inf).any())

    # ── Test 7: Training convergence (50 steps) ──
    print("\nTest 7: Training convergence (50 steps)")
    model.train()
    optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
    losses = []
    for step in range(50):
        optimizer.zero_grad()
        result = model(env, pfam, bio, bio_valid)
        result['total_loss'].backward()
        optimizer.step()
        losses.append(result['total_loss'].item())
    reduction = (losses[0] - losses[-1]) / losses[0] * 100
    print(f"  Loss: {losses[0]:.2f} -> {losses[-1]:.2f} ({reduction:.1f}% reduction)")
    check("loss decreases over 50 steps", losses[-1] < losses[0])
    check("no NaN in loss", all(not (l != l) for l in losses))

    # ── Test 8: Different latent dimensions ──
    print("\nTest 8: Different latent dimensions {16, 32, 64}")
    for ld in [16, 32, 64]:
        m = WorldModel(env_dim=24, pfam_dim=20, latent_dim=ld).to(device)
        m.train()
        r = m(env, pfam, bio, bio_valid)
        check(f"latent_dim={ld}: z_env shape ({N},{ld})",
              r['z_env'].shape == (N, ld))
        check(f"latent_dim={ld}: valid loss",
              r['total_loss'].item() > 0 and not torch.isnan(r['total_loss']))

    # ── Test 9: Custom VICReg configs ──
    print("\nTest 9: Custom VICReg configurations")
    configs = {
        'default': dict(lambda_inv=25.0, lambda_var=25.0, lambda_cov=1.0),
        'high_variance': dict(lambda_inv=10.0, lambda_var=50.0, lambda_cov=1.0),
        'high_covariance': dict(lambda_inv=25.0, lambda_var=25.0, lambda_cov=10.0),
    }
    for name, cfg in configs.items():
        m = WorldModel(env_dim=24, pfam_dim=20, **cfg).to(device)
        m.train()
        r = m(env, pfam, bio, bio_valid)
        check(f"{name}: valid loss",
              r['total_loss'].item() > 0 and not torch.isnan(r['total_loss']))

    # ── Test 10: Minimum batch size (N=2) ──
    print("\nTest 10: Minimum batch size (N=2)")
    env_small = torch.randn(2, 24, device=device)
    pfam_small = torch.randn(2, 20, device=device)
    bio_small = torch.randn(2, 3, device=device)
    valid_small = torch.ones(2, dtype=torch.bool, device=device)
    model.train()
    r_small = model(env_small, pfam_small, bio_small, valid_small)
    check("batch size 2 works", not torch.isnan(r_small['total_loss']))

    # ── Test 11: Standalone encoder methods ──
    print("\nTest 11: Standalone encoder methods")
    model.eval()
    with torch.no_grad():
        ze = model.encode_env(env)
        zp = model.encode_pfam(pfam)
    check("encode_env shape", ze.shape == (N, 16))
    check("encode_pfam shape", zp.shape == (N, 16))

    # ── Test 12: GPU computation (if available) ──
    print("\nTest 12: GPU computation")
    if torch.cuda.is_available():
        m_gpu = WorldModel(env_dim=24, pfam_dim=20).to('cuda')
        m_gpu.train()
        e_gpu = torch.randn(32, 24, device='cuda')
        p_gpu = torch.randn(32, 20, device='cuda')
        b_gpu = torch.randn(32, 3, device='cuda')
        v_gpu = torch.ones(32, dtype=torch.bool, device='cuda')
        r_gpu = m_gpu(e_gpu, p_gpu, b_gpu, v_gpu)
        r_gpu['total_loss'].backward()
        check("GPU forward + backward succeeded",
              not torch.isnan(r_gpu['total_loss']))
    else:
        print("  SKIP: CUDA not available")
        tests_total += 1
        tests_passed += 1

    # ── Test 13: Model serialization (save/load) ──
    print("\nTest 13: Model serialization (save/load)")
    import tempfile
    model.eval()
    with torch.no_grad():
        y_before = model.inference(env)

    checkpoint = {
        'model_state_dict': model.state_dict(),
        'config': model.config,
    }
    with tempfile.NamedTemporaryFile(suffix='.pt', delete=False) as f:
        tmp_path = f.name
        torch.save(checkpoint, f)

    # Load into fresh model
    loaded = torch.load(tmp_path, map_location=device, weights_only=False)
    model2 = WorldModel(**loaded['config']).to(device)
    model2.load_state_dict(loaded['model_state_dict'])
    model2.eval()
    with torch.no_grad():
        y_after = model2.inference(env)

    max_diff = (y_before - y_after).abs().max().item()
    print(f"  Max prediction diff after save/load: {max_diff:.2e}")
    check("save/load produces identical predictions", max_diff < 1e-6)

    os.unlink(tmp_path)

    # ── Summary ──
    print(f"\n{'=' * 70}")
    print(f"Results: {tests_passed}/{tests_total} tests passed")
    print(f"{'=' * 70}")

    return tests_passed == tests_total


if __name__ == '__main__':
    success = self_test()
    sys.exit(0 if success else 1)