File size: 17,862 Bytes
2af0e94
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Tests for SafeConvTranspose3d — verifies mathematical equivalence with nn.ConvTranspose3d.

Tests cover:
1. Forward pass: output correctness (V1: ~5e-7 precision, V2: bit-for-bit)
2. Backward pass: identical gradients w.r.t. input, weight, and bias
3. Checkpoint loading: weight shapes match nn.ConvTranspose3d
4. Various channel configurations matching the codebase usage
5. torch.autograd.gradcheck for numerical Jacobian verification
"""

import torch
import torch.nn as nn
import torch.nn.functional as F
import sys
import os

sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..'))
from Diffusion.safe_conv_transpose import (
    SafeConvTranspose3d,
    SafeConvTranspose3d_v2,
    replace_conv_transpose3d,
)


def _make_pair(in_c, out_c, kernel_size=4, stride=2, padding=1, bias=True):
    """Create nn.ConvTranspose3d and both Safe variants with identical weights."""
    torch.manual_seed(42)
    ref = nn.ConvTranspose3d(in_c, out_c, kernel_size, stride, padding, bias=bias)

    safe1 = SafeConvTranspose3d(in_c, out_c, kernel_size, stride, padding, bias=bias)
    safe1.weight.data.copy_(ref.weight.data)
    if bias:
        safe1.bias.data.copy_(ref.bias.data)

    safe2 = SafeConvTranspose3d_v2(in_c, out_c, kernel_size, stride, padding, bias=bias)
    safe2.weight.data.copy_(ref.weight.data)
    if bias:
        safe2.bias.data.copy_(ref.bias.data)

    return ref, safe1, safe2


# =============================================================================
# Basic shape tests
# =============================================================================

def test_weight_shape():
    """Weight and bias shapes must match nn.ConvTranspose3d exactly."""
    for in_c, out_c in [(16, 16), (32, 32), (64, 64), (128, 128), (256, 256), (16, 32)]:
        ref = nn.ConvTranspose3d(in_c, out_c, 4, 2, 1)
        s1 = SafeConvTranspose3d(in_c, out_c, 4, 2, 1)
        s2 = SafeConvTranspose3d_v2(in_c, out_c, 4, 2, 1)

        assert ref.weight.shape == s1.weight.shape == s2.weight.shape, \
            f"Weight shape mismatch for {in_c}->{out_c}"
        assert ref.bias.shape == s1.bias.shape == s2.bias.shape, \
            f"Bias shape mismatch for {in_c}->{out_c}"
    print("PASS: test_weight_shape")


def test_output_shape():
    """Output shape must be [B, C_out, 2*D, 2*H, 2*W] for stride=2."""
    for in_size in [2, 4, 8, 16]:
        ref = nn.ConvTranspose3d(16, 16, 4, 2, 1)
        safe1 = SafeConvTranspose3d(16, 16, 4, 2, 1)
        safe2 = SafeConvTranspose3d_v2(16, 16, 4, 2, 1)

        x = torch.randn(1, 16, in_size, in_size, in_size)
        expected = (1, 16, 2 * in_size, 2 * in_size, 2 * in_size)
        assert ref(x).shape == expected
        assert safe1(x).shape == expected
        assert safe2(x).shape == expected
    print("PASS: test_output_shape")


# =============================================================================
# Forward precision tests
# =============================================================================

def test_forward_v1():
    """V1 (decomposed) forward must be close to nn.ConvTranspose3d (~5e-7 precision)."""
    configs = [
        (16, 16, (2, 16, 4, 4, 4)),
        (32, 32, (1, 32, 8, 8, 8)),
        (64, 64, (1, 64, 4, 4, 4)),
        (128, 128, (1, 128, 2, 2, 2)),
        (256, 256, (1, 256, 2, 2, 2)),
    ]
    for in_c, out_c, shape in configs:
        ref, safe1, _ = _make_pair(in_c, out_c)
        x = torch.randn(shape)
        with torch.no_grad():
            y_ref = ref(x)
            y_safe = safe1(x)
        max_diff = (y_ref - y_safe).abs().max().item()
        assert max_diff < 1e-5, f"V1 forward diff {max_diff} for {in_c}->{out_c}"
        print(f"  {in_c:3d}->{out_c:3d} input={shape}: max_diff={max_diff:.2e}")
    print("PASS: test_forward_v1")


def test_forward_v2():
    """V2 (custom autograd) forward must be bit-for-bit identical."""
    configs = [
        (16, 16, (2, 16, 4, 4, 4)),
        (32, 32, (1, 32, 8, 8, 8)),
        (64, 64, (1, 64, 4, 4, 4)),
        (128, 128, (1, 128, 2, 2, 2)),
    ]
    for in_c, out_c, shape in configs:
        ref, _, safe2 = _make_pair(in_c, out_c)
        x = torch.randn(shape)
        with torch.no_grad():
            y_ref = ref(x)
            y_safe = safe2(x)
        max_diff = (y_ref - y_safe).abs().max().item()
        assert max_diff == 0.0, f"V2 forward should be bit-for-bit, got diff {max_diff}"
    print("PASS: test_forward_v2")


def test_forward_v1_precision_analysis():
    """Detailed precision analysis for V1 vs reference."""
    ref, safe1, _ = _make_pair(32, 32)
    x = torch.randn(2, 32, 8, 8, 8)
    with torch.no_grad():
        y_ref = ref(x)
        y_safe = safe1(x)
    diff = (y_ref - y_safe).abs()
    print(f"  32->32, [2,32,8,8,8]:")
    print(f"    max absolute diff:  {diff.max().item():.2e}")
    print(f"    mean absolute diff: {diff.mean().item():.2e}")
    print(f"    % elements > 1e-6:  {(diff > 1e-6).float().mean().item()*100:.2f}%")
    assert diff.max().item() < 1e-4
    print("PASS: test_forward_v1_precision_analysis")


# =============================================================================
# Backward tests
# =============================================================================

def _test_backward(version, label):
    """Test backward for grad_input, grad_weight, grad_bias with non-trivial upstream gradient."""
    for C_in, C_out, D_in, B in [(4, 4, 3, 2), (8, 4, 5, 1), (4, 8, 4, 3),
                                   (16, 16, 4, 2), (32, 32, 4, 1)]:
        torch.manual_seed(42)
        ct = nn.ConvTranspose3d(C_in, C_out, 4, 2, 1, bias=True)
        safe = (SafeConvTranspose3d if version == 1 else SafeConvTranspose3d_v2)(
            C_in, C_out, 4, 2, 1, bias=True
        )
        safe.weight.data.copy_(ct.weight.data)
        safe.bias.data.copy_(ct.bias.data)

        torch.manual_seed(123)
        x_ref = torch.randn(B, C_in, D_in, D_in, D_in, requires_grad=True)
        x_safe = x_ref.detach().clone().requires_grad_(True)

        torch.manual_seed(456)
        grad_y = torch.randn(B, C_out, 2 * D_in, 2 * D_in, 2 * D_in)

        ct(x_ref).backward(grad_y)
        safe(x_safe).backward(grad_y)

        dx = (x_ref.grad - x_safe.grad).abs().max().item()
        dw = (ct.weight.grad - safe.weight.grad).abs().max().item()
        db = (ct.bias.grad - safe.bias.grad).abs().max().item()

        assert dx < 1e-4, f"V{version} grad_input diff {dx} for {C_in}->{C_out}"
        assert dw < 1e-3, f"V{version} grad_weight diff {dw} for {C_in}->{C_out}"
        assert db < 1e-3, f"V{version} grad_bias diff {db} for {C_in}->{C_out}"
        print(f"  {C_in:2d}->{C_out:2d} D={D_in} B={B}: dx={dx:.2e} dw={dw:.2e} db={db:.2e}")

    print(f"PASS: test_backward_{label}")


def test_backward_v1():
    _test_backward(1, "v1")

def test_backward_v2():
    _test_backward(2, "v2")


def test_optimization_step():
    """Run 3 SGD steps and verify parameters stay close."""
    torch.manual_seed(42)
    ref = nn.ConvTranspose3d(16, 16, 4, 2, 1)

    safe1 = SafeConvTranspose3d(16, 16, 4, 2, 1)
    safe1.weight.data.copy_(ref.weight.data)
    safe1.bias.data.copy_(ref.bias.data)

    safe2 = SafeConvTranspose3d_v2(16, 16, 4, 2, 1)
    safe2.weight.data.copy_(ref.weight.data)
    safe2.bias.data.copy_(ref.bias.data)

    opt_ref = torch.optim.SGD(ref.parameters(), lr=0.01)
    opt_s1 = torch.optim.SGD(safe1.parameters(), lr=0.01)
    opt_s2 = torch.optim.SGD(safe2.parameters(), lr=0.01)

    for step in range(3):
        torch.manual_seed(step * 100)
        x = torch.randn(1, 16, 4, 4, 4)

        for opt, mod in [(opt_ref, ref), (opt_s1, safe1), (opt_s2, safe2)]:
            opt.zero_grad()
            mod(x).sum().backward()
            opt.step()

    w1 = (ref.weight.data - safe1.weight.data).abs().max().item()
    w2 = (ref.weight.data - safe2.weight.data).abs().max().item()
    print(f"  After 3 SGD steps: V1 drift={w1:.2e}, V2 drift={w2:.2e}")
    assert w1 < 1e-4
    assert w2 < 1e-4
    print("PASS: test_optimization_step")


# =============================================================================
# Checkpoint and replacement tests
# =============================================================================

def test_no_bias():
    """bias=False must work correctly."""
    ref, safe1, safe2 = _make_pair(16, 16, bias=False)
    x = torch.randn(1, 16, 4, 4, 4)
    with torch.no_grad():
        y_ref = ref(x)
        y_s1 = safe1(x)
        y_s2 = safe2(x)
    assert safe1.bias is None and safe2.bias is None
    assert (y_ref - y_s1).abs().max().item() < 1e-5
    assert (y_ref - y_s2).abs().max().item() == 0.0
    print("PASS: test_no_bias")


def test_checkpoint_loading():
    """state_dict from nn.ConvTranspose3d must load into Safe variants."""
    ref = nn.ConvTranspose3d(32, 32, 4, 2, 1)
    sd = ref.state_dict()

    safe1 = SafeConvTranspose3d(32, 32, 4, 2, 1)
    safe1.load_state_dict(sd)

    safe2 = SafeConvTranspose3d_v2(32, 32, 4, 2, 1)
    safe2.load_state_dict(sd)

    assert (safe1.weight.data - ref.weight.data).abs().max().item() == 0.0
    assert (safe2.weight.data - ref.weight.data).abs().max().item() == 0.0
    print("PASS: test_checkpoint_loading")


def test_replace_utility():
    """Test recursive replacement utility."""

    class Decoder(nn.Module):
        def __init__(self):
            super().__init__()
            self.up1 = nn.ConvTranspose3d(64, 32, 4, 2, 1)
            self.up2 = nn.ConvTranspose3d(32, 16, 4, 2, 1)
            self.conv = nn.Conv3d(16, 3, 3, 1, 1)  # should NOT be replaced

        def forward(self, x):
            return self.conv(self.up2(self.up1(x)))

    model = Decoder()
    x = torch.randn(1, 64, 4, 4, 4)
    with torch.no_grad():
        y_before = model(x).clone()

    replace_conv_transpose3d(model)
    assert isinstance(model.up1, SafeConvTranspose3d)
    assert isinstance(model.up2, SafeConvTranspose3d)
    assert isinstance(model.conv, nn.Conv3d)

    with torch.no_grad():
        y_after = model(x)
    max_diff = (y_before - y_after).abs().max().item()
    assert max_diff < 1e-4, f"Replace utility diff {max_diff}"
    print(f"  Replace utility: max diff = {max_diff:.2e}")
    print("PASS: test_replace_utility")


def test_replace_v2():
    """Replacement with V2 should be bit-for-bit in forward."""

    class Decoder(nn.Module):
        def __init__(self):
            super().__init__()
            self.up1 = nn.ConvTranspose3d(64, 32, 4, 2, 1)
            self.up2 = nn.ConvTranspose3d(32, 16, 4, 2, 1)

        def forward(self, x):
            return self.up2(self.up1(x))

    model = Decoder()
    x = torch.randn(1, 64, 4, 4, 4)
    with torch.no_grad():
        y_before = model(x).clone()

    replace_conv_transpose3d(model, target_cls=SafeConvTranspose3d_v2)
    assert isinstance(model.up1, SafeConvTranspose3d_v2)
    assert isinstance(model.up2, SafeConvTranspose3d_v2)

    with torch.no_grad():
        y_after = model(x)
    assert (y_before - y_after).abs().max().item() == 0.0
    print("PASS: test_replace_v2")


def test_asymmetric_channels():
    """in_channels != out_channels."""
    ref, safe1, safe2 = _make_pair(64, 32)
    x = torch.randn(1, 64, 4, 4, 4)
    with torch.no_grad():
        y_ref = ref(x)
        y_s1 = safe1(x)
        y_s2 = safe2(x)
    assert y_ref.shape == y_s1.shape == y_s2.shape
    assert (y_ref - y_s1).abs().max().item() < 1e-5
    assert (y_ref - y_s2).abs().max().item() == 0.0
    print("PASS: test_asymmetric_channels")


# =============================================================================
# Numerical gradient verification
# =============================================================================

def test_gradcheck_v1():
    """Numerical Jacobian check for V1."""
    safe1 = SafeConvTranspose3d(2, 2, 4, 2, 1, bias=True).double()
    x = torch.randn(1, 2, 3, 3, 3, dtype=torch.float64, requires_grad=True)
    result = torch.autograd.gradcheck(safe1, (x,), eps=1e-6, atol=1e-4, rtol=1e-3)
    assert result
    print("PASS: test_gradcheck_v1")


def test_gradcheck_v2():
    """Numerical Jacobian check for V2."""
    safe2 = SafeConvTranspose3d_v2(2, 2, 4, 2, 1, bias=True).double()
    x = torch.randn(1, 2, 3, 3, 3, dtype=torch.float64, requires_grad=True)
    result = torch.autograd.gradcheck(safe2, (x,), eps=1e-6, atol=1e-4, rtol=1e-3)
    assert result
    print("PASS: test_gradcheck_v2")


# =============================================================================
# Training loss equivalence test
# =============================================================================

def test_training_loss_equivalence():
    """Build a small encoder-decoder network with ConvTranspose3d layers,
    train for several steps, then replace with SafeConvTranspose3d and verify
    the loss values are identical (V2) or near-identical (V1)."""

    class MiniUNet(nn.Module):
        """Small UNet-like network with 3 ConvTranspose3d layers."""
        def __init__(self):
            super().__init__()
            self.enc1 = nn.Conv3d(1, 16, 4, 2, 1)
            self.enc2 = nn.Conv3d(16, 32, 4, 2, 1)
            self.enc3 = nn.Conv3d(32, 64, 4, 2, 1)
            self.dec3 = nn.ConvTranspose3d(64, 32, 4, 2, 1)
            self.dec2 = nn.ConvTranspose3d(32, 16, 4, 2, 1)
            self.dec1 = nn.ConvTranspose3d(16, 1, 4, 2, 1)
            self.act = nn.ReLU()

        def forward(self, x):
            e1 = self.act(self.enc1(x))
            e2 = self.act(self.enc2(e1))
            e3 = self.act(self.enc3(e2))
            d3 = self.act(self.dec3(e3))
            d2 = self.act(self.dec2(d3))
            d1 = self.dec1(d2)
            return d1

    import copy

    torch.manual_seed(42)
    model_ref = MiniUNet()

    # Create Safe V1 version by replacing ConvTranspose3d layers
    model_v1 = copy.deepcopy(model_ref)
    replace_conv_transpose3d(model_v1, target_cls=SafeConvTranspose3d)

    # Create Safe V2 version
    model_v2 = copy.deepcopy(model_ref)
    replace_conv_transpose3d(model_v2, target_cls=SafeConvTranspose3d_v2)

    # Verify the replacement happened
    assert isinstance(model_v1.dec1, SafeConvTranspose3d)
    assert isinstance(model_v2.dec1, SafeConvTranspose3d_v2)

    opt_ref = torch.optim.Adam(model_ref.parameters(), lr=1e-3)
    opt_v1 = torch.optim.Adam(model_v1.parameters(), lr=1e-3)
    opt_v2 = torch.optim.Adam(model_v2.parameters(), lr=1e-3)

    criterion = nn.MSELoss()
    n_steps = 5

    print(f"  Training {n_steps} steps, comparing loss at each step:")
    for step in range(n_steps):
        torch.manual_seed(step * 777)
        x = torch.randn(2, 1, 16, 16, 16)
        target = torch.randn(2, 1, 16, 16, 16)

        # Reference (ConvTranspose3d)
        opt_ref.zero_grad()
        loss_ref = criterion(model_ref(x), target)
        loss_ref.backward()
        opt_ref.step()

        # V1 (SafeConvTranspose3d)
        opt_v1.zero_grad()
        loss_v1 = criterion(model_v1(x), target)
        loss_v1.backward()
        opt_v1.step()

        # V2 (SafeConvTranspose3d_v2)
        opt_v2.zero_grad()
        loss_v2 = criterion(model_v2(x), target)
        loss_v2.backward()
        opt_v2.step()

        diff_v1 = abs(loss_ref.item() - loss_v1.item())
        diff_v2 = abs(loss_ref.item() - loss_v2.item())
        print(f"    step {step}: loss_ref={loss_ref.item():.6f}  "
              f"loss_v1={loss_v1.item():.6f} (diff={diff_v1:.2e})  "
              f"loss_v2={loss_v2.item():.6f} (diff={diff_v2:.2e})")

        assert diff_v1 < 1e-4, f"V1 loss diverged at step {step}: diff={diff_v1}"
        assert diff_v2 < 1e-6, f"V2 loss diverged at step {step}: diff={diff_v2}"

    # Check final weight divergence after training
    w_diff_v1 = max(
        (p1.data - p2.data).abs().max().item()
        for p1, p2 in zip(model_ref.parameters(), model_v1.parameters())
    )
    w_diff_v2 = max(
        (p1.data - p2.data).abs().max().item()
        for p1, p2 in zip(model_ref.parameters(), model_v2.parameters())
    )
    print(f"  After {n_steps} steps — max weight drift: V1={w_diff_v1:.2e}, V2={w_diff_v2:.2e}")
    assert w_diff_v1 < 1e-3, f"V1 weight drift too large: {w_diff_v1}"
    assert w_diff_v2 < 1e-4, f"V2 weight drift too large: {w_diff_v2}"
    print("PASS: test_training_loss_equivalence")


# =============================================================================
# Main
# =============================================================================

if __name__ == '__main__':
    print("=" * 70)
    print("Testing SafeConvTranspose3d implementations")
    print("=" * 70)

    tests = [
        ("Weight shapes", test_weight_shape),
        ("Output shapes", test_output_shape),
        ("Forward V1 (decomposed)", test_forward_v1),
        ("Forward V2 (custom autograd)", test_forward_v2),
        ("Forward V1 precision analysis", test_forward_v1_precision_analysis),
        ("Backward V1", test_backward_v1),
        ("Backward V2", test_backward_v2),
        ("Optimization step", test_optimization_step),
        ("No bias", test_no_bias),
        ("Checkpoint loading", test_checkpoint_loading),
        ("Replace utility (V1)", test_replace_utility),
        ("Replace utility (V2)", test_replace_v2),
        ("Asymmetric channels", test_asymmetric_channels),
        ("Gradcheck V1", test_gradcheck_v1),
        ("Gradcheck V2", test_gradcheck_v2),
        ("Training loss equivalence", test_training_loss_equivalence),
    ]

    failed = []
    for name, fn in tests:
        print(f"\n--- {name} ---")
        try:
            fn()
        except Exception as e:
            print(f"FAIL: {name}: {e}")
            failed.append(name)

    print("\n" + "=" * 70)
    if failed:
        print(f"FAILED ({len(failed)}/{len(tests)}): {', '.join(failed)}")
    else:
        print(f"ALL {len(tests)} TESTS PASSED")
    print("=" * 70)