File size: 6,918 Bytes
b47a1ce
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1dae740
b47a1ce
1dae740
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
import types

import pytest
import torch
from dememwm_import_helper import install_dememwm_namespace

install_dememwm_namespace()
from algorithms.worldmem.dememwm.algorithm import MemoryDiTMixin
from algorithms.worldmem.dememwm.compression import CausalConv3DDynamicCompressor
from algorithms.worldmem.dememwm.schedules import (
    EVAL_ABLATION_BRANCHES,
    EVAL_ABLATION_BRANCH_TO_ID,
    EVAL_CORRUPTION_BRANCHES,
    normalize_eval_ablation_branch,
)


WAVE9_BRANCHES = (
    "memory_off",
    "A_only",
    "D_only",
    "A_plus_D",
    "A_plus_D_plus_R_normal",
    "R_forced_off",
    "R_forced_on",
    "wrong_pose",
    "time_shuffle",
    "source_matched_random",
    "pose_shuffle",
    "wrong_video",
    "local_context_overlap_fake_revisit",
)


def _device():
    return torch.device("cpu")


def test_wave9_branch_registry_is_exact_and_validated():
    assert EVAL_ABLATION_BRANCHES == WAVE9_BRANCHES
    assert EVAL_ABLATION_BRANCH_TO_ID["memory_off"] == 0
    assert EVAL_ABLATION_BRANCH_TO_ID["local_context_overlap_fake_revisit"] == len(WAVE9_BRANCHES) - 1
    assert EVAL_CORRUPTION_BRANCHES == WAVE9_BRANCHES[7:]
    assert normalize_eval_ablation_branch(None) == "A_plus_D_plus_R_normal"
    assert normalize_eval_ablation_branch("wrong_pose") == "wrong_pose"
    with pytest.raises(ValueError):
        normalize_eval_ablation_branch("ratio_sweep")


class ConstantGate(torch.nn.Module):
    def __init__(self, value: float):
        super().__init__()
        self.value = float(value)

    def forward(self, *, valid_revisit_mask, best_selected_fov_overlap, best_selected_plucker_overlap, selected_gap_frames):
        del valid_revisit_mask, best_selected_plucker_overlap, selected_gap_frames
        return torch.full_like(best_selected_fov_overlap, self.value, dtype=torch.float32)


class DummyDeMemWM(MemoryDiTMixin):
    def __init__(self, branch: str, device: torch.device):
        self.cfg = types.SimpleNamespace(
            dememwm=types.SimpleNamespace(
                enabled=True,
                training_stage="stage_2",
                token_patch_size=2,
                curriculum=types.SimpleNamespace(enabled=False),
                anchor=types.SimpleNamespace(
                    enabled=True,
                    anchor_indices=[0, 1],
                    allow_generated_as_anchor=False,
                    diverse_selection=False,
                    compress=types.SimpleNamespace(pool_h=1, pool_w=1),
                ),
                dynamic=types.SimpleNamespace(
                    enabled=True,
                    exclude_latest_local_frames=2,
                    recent_frames=4,
                    conv_kernel_t=3,
                    conv_stride_t=2,
                ),
                revisit=types.SimpleNamespace(
                    enabled=True,
                    deterministic_pose_retrieval=True,
                    fov_overlap_threshold=0.0,
                    plucker_weight=0.1,
                    max_frames=2,
                    compress=types.SimpleNamespace(pool_h=1, pool_w=1),
                ),
                eval_ablation=types.SimpleNamespace(enabled=True, branch=branch),
                generated_history_proxy=types.SimpleNamespace(enabled=False),
                injection=types.SimpleNamespace(dit_hidden_size=8, anchor_gate=1.0, dynamic_gate=1.0, revisit_gate=1.0),
                cache=types.SimpleNamespace(enabled=False),
                checkpoint=types.SimpleNamespace(strict_dememwm_eval_load=True),
            ),
            weight_decay=0.0,
            optimizer_beta=(0.9, 0.999),
        )
        self.global_step = 0
        self.x_stacked_shape = (1, 4, 4)
        self.dememwm_anchor_proj = torch.nn.Linear(4, 8, bias=False).to(device)
        self.dememwm_revisit_proj = torch.nn.Linear(4, 8, bias=False).to(device)
        self.dememwm_dynamic_compressor = CausalConv3DDynamicCompressor(
            latent_channels=1,
            dit_hidden_size=8,
            patch_size=2,
            conv_kernel_t=3,
            conv_stride_t=2,
            max_source_frames=4,
            exclude_latest_local_frames=2,
        ).to(device)
        self.dememwm_revisit_gate = ConstantGate(0.25).to(device)


def _streams(branch: str):
    device = _device()
    model = DummyDeMemWM(branch, device)
    latents = torch.arange(12 * 1 * 1 * 4 * 4, device=device, dtype=torch.float32).reshape(12, 1, 1, 4, 4) / 100.0
    source_frames = torch.arange(12, device=device).reshape(12, 1)
    target_frames = torch.tensor([[8], [12]], device=device)
    pose = torch.zeros((12, 1, 5), device=device, dtype=torch.float32)
    target_pose = torch.zeros((2, 1, 5), device=device, dtype=torch.float32)
    return model.build_memory_streams(
        latents,
        source_frames,
        target_frame_indices=target_frames,
        pose=pose,
        target_pose=target_pose,
        action=None,
        target_action=None,
    )


def test_eval_ablation_stream_enable_branches_control_masks_and_gates():
    memory_off = _streams("memory_off")
    assert memory_off.anchor_gate == 0.0
    assert memory_off.dynamic_gate == 0.0
    assert torch.count_nonzero(memory_off.revisit_gate).item() == 0
    assert not memory_off.anchor_mask.any()
    assert not memory_off.dynamic_mask.any()
    assert not memory_off.revisit_mask.any()

    a_only = _streams("A_only")
    assert a_only.anchor_gate == 1.0
    assert a_only.anchor_mask.any()
    assert not a_only.dynamic_mask.any()
    assert not a_only.revisit_mask.any()

    d_only = _streams("D_only")
    assert d_only.dynamic_gate == 1.0
    assert d_only.dynamic_mask.any()
    assert d_only.anchor_gate == 0.0
    assert not d_only.anchor_mask.any()
    assert not d_only.revisit_mask.any()

    a_plus_d = _streams("A_plus_D")
    assert a_plus_d.anchor_mask.any()
    assert a_plus_d.dynamic_mask.any()
    assert not a_plus_d.revisit_mask.any()
    assert torch.count_nonzero(a_plus_d.revisit_gate).item() == 0


def test_eval_ablation_forced_revisit_controls_are_isolated_to_eval_branch():
    normal = _streams("A_plus_D_plus_R_normal")
    forced_off = _streams("R_forced_off")
    forced_on = _streams("R_forced_on")
    assert normal.valid_revisit_mask.all()
    assert torch.allclose(normal.revisit_gate, torch.full_like(normal.revisit_gate, 0.25))
    assert torch.count_nonzero(forced_off.revisit_gate).item() == 0
    assert torch.equal(forced_on.revisit_gate, forced_on.valid_revisit_mask.to(dtype=forced_on.revisit_gate.dtype))


def test_eval_ablation_corruption_branch_marks_corrupted_revisit_without_zeroing_gate():
    wrong_pose = _streams("wrong_pose")
    assert wrong_pose.valid_revisit_mask.all()
    assert wrong_pose.revisit_mask.any()
    assert torch.allclose(wrong_pose.revisit_gate, torch.full_like(wrong_pose.revisit_gate, 0.25))
    assert wrong_pose.revisit_best_selected_fov_overlap.shape == wrong_pose.valid_revisit_mask.shape