File size: 9,786 Bytes
44f5956
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b9efa38
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d15d8f2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
44f5956
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b9efa38
 
 
 
 
 
 
 
 
 
 
 
 
44f5956
 
 
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
import unittest

import torch
from torch import nn

from algorithms.dememwm.models.diffusion import Diffusion


class FakeDenoiser(nn.Module):
    def __init__(self, output_frames=None):
        super().__init__()
        self.output_frames = output_frames
        self.calls = []

    def forward(self, x, t, action_cond, **kwargs):
        self.calls.append({
            "x_shape": tuple(x.shape),
            "t_shape": tuple(t.shape),
            "kwargs": kwargs,
        })
        frames = self.output_frames if self.output_frames is not None else x.shape[1]
        return torch.zeros((x.shape[0], frames, *x.shape[2:]), device=x.device, dtype=x.dtype)


def _make_diffusion(output_frames=None):
    diffusion = Diffusion.__new__(Diffusion)
    nn.Module.__init__(diffusion)
    diffusion.x_shape = torch.Size((1, 1, 1))
    diffusion.timesteps = 4
    diffusion.sampling_timesteps = 4
    diffusion.is_ddim_sampling = False
    diffusion.objective = "pred_noise"
    diffusion.use_fused_snr = False
    diffusion.snr_clip = 5.0
    diffusion.cum_snr_decay = 0.9
    diffusion.ddim_sampling_eta = 0.0
    diffusion.clip_noise = 10.0
    diffusion.stabilization_level = 1
    diffusion.model = FakeDenoiser(output_frames=output_frames)

    betas = torch.tensor([0.05, 0.10, 0.15, 0.20], dtype=torch.float32)
    alphas = 1.0 - betas
    alphas_cumprod = torch.cumprod(alphas, dim=0)
    alphas_cumprod_prev = torch.nn.functional.pad(alphas_cumprod[:-1], (1, 0), value=1.0)
    posterior_variance = betas * (1.0 - alphas_cumprod_prev) / (1.0 - alphas_cumprod)
    snr = alphas_cumprod / (1.0 - alphas_cumprod)

    diffusion.register_buffer("betas", betas)
    diffusion.register_buffer("alphas_cumprod", alphas_cumprod)
    diffusion.register_buffer("alphas_cumprod_prev", alphas_cumprod_prev)
    diffusion.register_buffer("sqrt_alphas_cumprod", torch.sqrt(alphas_cumprod))
    diffusion.register_buffer("sqrt_one_minus_alphas_cumprod", torch.sqrt(1.0 - alphas_cumprod))
    diffusion.register_buffer("log_one_minus_alphas_cumprod", torch.log(1.0 - alphas_cumprod))
    diffusion.register_buffer("sqrt_recip_alphas_cumprod", torch.sqrt(1.0 / alphas_cumprod))
    diffusion.register_buffer("sqrt_recipm1_alphas_cumprod", torch.sqrt(1.0 / alphas_cumprod - 1.0))
    diffusion.register_buffer("posterior_variance", posterior_variance)
    diffusion.register_buffer("posterior_log_variance_clipped", torch.log(posterior_variance.clamp(min=1e-20)))
    diffusion.register_buffer("posterior_mean_coef1", betas * torch.sqrt(alphas_cumprod_prev) / (1.0 - alphas_cumprod))
    diffusion.register_buffer("posterior_mean_coef2", (1.0 - alphas_cumprod_prev) * torch.sqrt(alphas) / (1.0 - alphas_cumprod))
    diffusion.register_buffer("snr", snr)
    diffusion.register_buffer("clipped_snr", snr.clamp(max=diffusion.snr_clip))
    return diffusion


def _packed_inputs():
    x = torch.arange(5, dtype=torch.float32).view(5, 1, 1, 1, 1)
    action_cond = torch.zeros((5, 1, 3), dtype=torch.float32)
    noise_levels = torch.tensor([[1], [2], [0], [0], [0]], dtype=torch.long)
    segments = {"target": 2, "anchor": 1, "dynamic": 1, "revisit": 1}
    return x, action_cond, noise_levels, segments


class DeMemWMDiffusionTargetOnlyTests(unittest.TestCase):
    def test_forward_noises_packed_input_but_returns_target_loss(self):
        torch.manual_seed(0)
        diffusion = _make_diffusion(output_frames=2)
        x, action_cond, noise_levels, segments = _packed_inputs()

        x_pred, loss = diffusion(
            x,
            action_cond,
            None,
            noise_levels=noise_levels,
            reference_length=0,
            frame_memory_segments=segments,
        )

        self.assertEqual(diffusion.model.calls[0]["x_shape"], (1, 5, 1, 1, 1))
        self.assertEqual(tuple(x_pred.shape), (2, 1, 1, 1, 1))
        self.assertEqual(tuple(loss.shape), (2, 1, 1, 1, 1))

    def test_forward_without_frame_memory_keeps_full_length(self):
        torch.manual_seed(0)
        diffusion = _make_diffusion()
        x, action_cond, noise_levels, _ = _packed_inputs()

        x_pred, loss = diffusion(
            x,
            action_cond,
            None,
            noise_levels=noise_levels,
            reference_length=0,
        )

        self.assertEqual(diffusion.model.calls[0]["x_shape"], (1, 5, 1, 1, 1))
        self.assertEqual(tuple(x_pred.shape), tuple(x.shape))
        self.assertEqual(tuple(loss.shape), tuple(x.shape))

    def test_padded_frame_memory_masks_keep_target_prediction_and_loss_shapes(self):
        torch.manual_seed(0)
        diffusion = _make_diffusion(output_frames=2)
        x, action_cond, noise_levels, segments = _packed_inputs()
        masks = {
            "target": torch.ones((1, 2), dtype=torch.bool),
            "anchor": torch.ones((1, 1), dtype=torch.bool),
            "dynamic": torch.zeros((1, 1), dtype=torch.bool),
            "revisit": torch.ones((1, 1), dtype=torch.bool),
        }

        x_pred, loss = diffusion(
            x,
            action_cond,
            None,
            noise_levels=noise_levels,
            reference_length=0,
            frame_memory_segments=segments,
            frame_memory_masks=masks,
        )

        self.assertIs(diffusion.model.calls[0]["kwargs"]["frame_memory_masks"], masks)
        self.assertEqual(tuple(x_pred.shape), (2, 1, 1, 1, 1))
        self.assertEqual(tuple(loss.shape), (2, 1, 1, 1, 1))

    def test_frame_memory_pose_is_batch_first_and_separate_from_pose_cond(self):
        torch.manual_seed(0)
        diffusion = _make_diffusion(output_frames=2)
        x = torch.zeros((5, 2, 1, 1, 1), dtype=torch.float32)
        action_cond = torch.zeros((5, 2, 3), dtype=torch.float32)
        noise_levels = torch.zeros((5, 2), dtype=torch.long)
        pose_cond = torch.full((5, 2, 5), 99.0, dtype=torch.float32)
        frame_memory_pose = torch.arange(5 * 2 * 5, dtype=torch.float32).view(5, 2, 5)
        image_hw = torch.tensor([[360, 640], [720, 1280]], dtype=torch.long)
        segments = {"target": 2, "anchor": 1, "dynamic": 1, "revisit": 1}

        diffusion(
            x,
            action_cond,
            pose_cond,
            noise_levels=noise_levels,
            reference_length=0,
            frame_memory_segments=segments,
            frame_memory_pose=frame_memory_pose,
            image_hw=image_hw,
        )

        kwargs = diffusion.model.calls[0]["kwargs"]
        self.assertIsNone(kwargs["pose_cond"])
        self.assertEqual(tuple(kwargs["frame_memory_pose"].shape), (2, 5, 5))
        self.assertTrue(torch.equal(kwargs["frame_memory_pose"], frame_memory_pose.permute(1, 0, 2)))
        self.assertIs(kwargs["image_hw"], image_hw)

    def test_baseline_pose_cond_still_reaches_dit_batch_first(self):
        torch.manual_seed(0)
        diffusion = _make_diffusion()
        x = torch.zeros((5, 2, 1, 1, 1), dtype=torch.float32)
        action_cond = torch.zeros((5, 2, 3), dtype=torch.float32)
        noise_levels = torch.zeros((5, 2), dtype=torch.long)
        pose_cond = torch.arange(5 * 2 * 5, dtype=torch.float32).view(5, 2, 5)

        diffusion(
            x,
            action_cond,
            pose_cond,
            noise_levels=noise_levels,
            reference_length=0,
        )

        kwargs = diffusion.model.calls[0]["kwargs"]
        self.assertTrue(torch.equal(kwargs["pose_cond"], pose_cond.permute(1, 0, 2)))
        self.assertNotIn("frame_memory_pose", kwargs)

    def test_posterior_and_sample_steps_return_target_frames_for_frame_memory(self):
        torch.manual_seed(0)
        x, action_cond, _, segments = _packed_inputs()
        curr = torch.tensor([[2], [1], [-1], [-1], [-1]], dtype=torch.long)
        next_level = torch.tensor([[1], [0], [-1], [-1], [-1]], dtype=torch.long)

        diffusion = _make_diffusion(output_frames=2)
        mean, variance, log_variance = diffusion.p_mean_variance(
            x,
            curr,
            action_cond=action_cond,
            pose_cond=None,
            reference_length=0,
            frame_memory_segments=segments,
        )
        self.assertEqual(tuple(mean.shape), (2, 1, 1, 1, 1))
        self.assertEqual(tuple(variance.shape), (2, 1, 1, 1, 1))
        self.assertEqual(tuple(log_variance.shape), (2, 1, 1, 1, 1))

        diffusion = _make_diffusion(output_frames=2)
        ddpm = diffusion.ddpm_sample_step(
            x,
            action_cond,
            None,
            curr_noise_level=curr,
            reference_length=0,
            frame_memory_segments=segments,
        )
        self.assertEqual(diffusion.model.calls[0]["x_shape"], (1, 5, 1, 1, 1))
        self.assertEqual(tuple(ddpm.shape), (2, 1, 1, 1, 1))

        diffusion = _make_diffusion(output_frames=2)
        ddim = diffusion.ddim_sample_step(
            x,
            action_cond,
            None,
            curr_noise_level=curr,
            next_noise_level=next_level,
            reference_length=0,
            frame_memory_segments=segments,
        )
        self.assertEqual(diffusion.model.calls[0]["x_shape"], (1, 5, 1, 1, 1))
        self.assertEqual(tuple(ddim.shape), (2, 1, 1, 1, 1))

        diffusion = _make_diffusion(output_frames=2)
        sample = diffusion.sample_step(
            x,
            action_cond,
            None,
            curr_noise_level=torch.tensor([[3], [2], [0], [0], [0]], dtype=torch.long),
            next_noise_level=torch.tensor([[2], [1], [0], [0], [0]], dtype=torch.long),
            reference_length=0,
            frame_memory_segments=segments,
        )
        self.assertEqual(diffusion.model.calls[0]["x_shape"], (1, 5, 1, 1, 1))
        self.assertEqual(tuple(sample.shape), (2, 1, 1, 1, 1))


if __name__ == "__main__":
    unittest.main()