File size: 11,014 Bytes
4d2eebd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Unified multi-skeleton motion dataset for TopoSlots (Scheme C).

Motion representation:
  Slot token input : per-joint [local_pos(3) + velocity(3)] = 6D  (cross-skeleton)
  Decoder GT       : per-joint local_rotations_6d               (skeleton-specific, FK supervision)
  Root track       : root_position(3) + root_velocity(3)         (separate)
  Auxiliary        : foot_contact(4), bone_lengths, accelerations (losses)
"""

import numpy as np
import torch
from torch.utils.data import Dataset
from pathlib import Path
from typing import Optional


class UnifiedMotionDataset(Dataset):
    """
    Multi-skeleton motion dataset with unified format.

    Each sample returns:
        - motion_features: [T, J, D] padded to max_joints
        - skeleton_features: [J, D_skel] padded to max_joints
        - joint_mask: [J] boolean mask (True = valid joint)
        - adjacency: [J, J] padded adjacency matrix
        - geodesic_dist: [J, J] padded geodesic distances
        - text: str (empty if unavailable)
        - metadata: dict
    """

    def __init__(
        self,
        data_dirs: list[str | Path],
        split: str = 'train',
        max_joints: int = 128,
        max_frames: int = 196,
        target_fps: float = 20.0,
        motion_dim: int = 6,  # local_pos (3) + velocity (3)
    ):
        self.data_dirs = [Path(d) for d in data_dirs]
        self.split = split
        self.max_joints = max_joints
        self.max_frames = max_frames
        self.target_fps = target_fps
        self.motion_dim = motion_dim

        # Load all samples
        self.samples = []
        self.skeletons = {}  # skeleton_id -> skeleton data
        self.stats = {}  # skeleton_id -> normalization stats

        for data_dir in self.data_dirs:
            self._load_data_source(data_dir)

        print(f"UnifiedMotionDataset [{split}]: {len(self.samples)} motions, "
              f"{len(self.skeletons)} skeleton types")

    def _load_data_source(self, data_dir: Path):
        """Load one data source (e.g., processed/humanml3d)."""
        if not data_dir.exists():
            print(f"  Warning: {data_dir} not found, skipping")
            return

        # Load skeleton
        skel_path = data_dir / 'skeleton.npz'
        if skel_path.exists():
            skel_data = dict(np.load(skel_path, allow_pickle=True))
            skeleton_id = data_dir.name
            self.skeletons[skeleton_id] = skel_data

        # Load stats
        stats_path = data_dir / 'stats.npz'
        if stats_path.exists():
            self.stats[data_dir.name] = dict(np.load(stats_path))

        # Load split
        split_path = data_dir / 'splits' / f'{self.split}.txt'
        if not split_path.exists():
            # Fall back to all.txt
            split_path = data_dir / 'splits' / 'all.txt'
            if not split_path.exists():
                print(f"  Warning: no split file for {data_dir.name}, skipping")
                return

        motion_ids = []
        with open(split_path, 'r') as f:
            for line in f:
                line = line.strip()
                if line:
                    motion_ids.append(line)

        for mid in motion_ids:
            motion_path = data_dir / 'motions' / f'{mid}.npz'
            if motion_path.exists():
                self.samples.append({
                    'motion_path': str(motion_path),
                    'motion_id': mid,
                    'data_source': data_dir.name,
                    'skeleton_id': data_dir.name,
                })

    def __len__(self) -> int:
        return len(self.samples)

    def __getitem__(self, idx: int) -> dict:
        sample_info = self.samples[idx]

        # Load motion data
        data = dict(np.load(sample_info['motion_path'], allow_pickle=True))

        # Get skeleton info
        skeleton_id = sample_info['skeleton_id']
        skel_data = self.skeletons.get(skeleton_id, {})

        # Extract motion features
        local_pos = data['local_positions']  # [T, J, 3]
        velocities = data['velocities']       # [T, J, 3]
        T, J, _ = local_pos.shape

        # Normalize if stats available
        if skeleton_id in self.stats:
            stats = self.stats[skeleton_id]
            local_pos = (local_pos - stats['local_pos_mean']) / stats['local_pos_std']
            velocities = (velocities - stats['velocity_mean']) / stats['velocity_std']

        # Concatenate motion features: [T, J, 6]
        motion_features = np.concatenate([local_pos, velocities], axis=-1)

        # Crop/pad temporal dimension
        if T > self.max_frames:
            # Random crop during training
            if self.split == 'train':
                start = np.random.randint(0, T - self.max_frames)
            else:
                start = 0
            motion_features = motion_features[start:start + self.max_frames]
            actual_frames = self.max_frames
        else:
            actual_frames = T
            # Pad with zeros
            pad = np.zeros(
                (self.max_frames - T, J, self.motion_dim),
                dtype=np.float32,
            )
            motion_features = np.concatenate([motion_features, pad], axis=0)

        # Pad joint dimension
        padded_motion = np.zeros(
            (self.max_frames, self.max_joints, self.motion_dim),
            dtype=np.float32,
        )
        padded_motion[:, :J, :] = motion_features

        # Joint mask
        joint_mask = np.zeros(self.max_joints, dtype=np.bool_)
        joint_mask[:J] = True

        # Frame mask
        frame_mask = np.zeros(self.max_frames, dtype=np.bool_)
        frame_mask[:actual_frames] = True

        # Skeleton features
        skeleton_features = np.zeros(
            (self.max_joints, 9), dtype=np.float32
        )
        if 'joint_names' in skel_data:
            from .skeleton_graph import SkeletonGraph
            sg = SkeletonGraph.from_dict(skel_data)
            skel_feats = sg.get_joint_features()  # [J, 9]
            skeleton_features[:J] = skel_feats

        # Adjacency and geodesic distance matrices
        adjacency = np.zeros(
            (self.max_joints, self.max_joints), dtype=np.float32
        )
        geodesic_dist = np.zeros(
            (self.max_joints, self.max_joints), dtype=np.float32
        )
        if 'adjacency' in skel_data:
            adj = skel_data['adjacency']
            adjacency[:J, :J] = adj
        if 'geodesic_dist' in skel_data:
            gdist = skel_data['geodesic_dist']
            geodesic_dist[:J, :J] = gdist

        # Text
        text = ''
        if 'texts' in data:
            texts_str = str(data['texts'])
            if texts_str:
                text_list = texts_str.split('|||')
                if text_list and text_list[0]:
                    # Random text during training
                    if self.split == 'train':
                        text = text_list[np.random.randint(len(text_list))]
                    else:
                        text = text_list[0]

        # --- Root track (separate from slot tokens) ---
        root_pos = data.get('root_position', np.zeros((T, 3), dtype=np.float32))
        root_vel = data.get('root_velocity', np.zeros((T, 3), dtype=np.float32))
        padded_root_pos = np.zeros((self.max_frames, 3), dtype=np.float32)
        padded_root_vel = np.zeros((self.max_frames, 3), dtype=np.float32)
        padded_root_pos[:actual_frames] = root_pos[:actual_frames]
        padded_root_vel[:actual_frames] = root_vel[:actual_frames]

        # --- Foot contact: [T, 4] (l_heel, l_toe, r_heel, r_toe) ---
        fc_raw = data.get('foot_contact', np.zeros((T, 4), dtype=np.float32))
        if fc_raw.shape[-1] == 2:
            # Legacy 2-channel → duplicate into 4-channel
            fc_4ch = np.zeros((fc_raw.shape[0], 4), dtype=np.float32)
            fc_4ch[:, 0] = fc_4ch[:, 1] = fc_raw[:, 0]
            fc_4ch[:, 2] = fc_4ch[:, 3] = fc_raw[:, 1]
            fc_raw = fc_4ch
        padded_contact = np.zeros((self.max_frames, 4), dtype=np.float32)
        padded_contact[:actual_frames] = fc_raw[:actual_frames]

        # --- Decoder GT: local rotations 6D (skeleton-specific, for FK supervision) ---
        rot_6d = data.get('local_rotations_6d', None)
        if rot_6d is not None:
            # [T, J-1, 6] → pad to [T_max, J_max, 6]
            Jr = rot_6d.shape[1]  # J-1 (non-root)
            padded_rot = np.zeros((self.max_frames, self.max_joints, 6), dtype=np.float32)
            T_rot = min(rot_6d.shape[0], actual_frames)
            padded_rot[:T_rot, :Jr, :] = rot_6d[:T_rot]
            has_rotations = True
        else:
            padded_rot = np.zeros((self.max_frames, self.max_joints, 6), dtype=np.float32)
            has_rotations = False

        # --- Bone lengths [T, J] ---
        bone_raw = data.get('bone_lengths', np.zeros((T, J), dtype=np.float32))
        padded_bones = np.zeros((self.max_frames, self.max_joints), dtype=np.float32)
        padded_bones[:actual_frames, :J] = bone_raw[:actual_frames]

        return {
            # Slot token input: per-joint [local_pos(3) + velocity(3)] = 6D
            'motion_features': torch.from_numpy(padded_motion),         # [T, J_max, 6]
            # Skeleton graph
            'skeleton_features': torch.from_numpy(skeleton_features),   # [J_max, 9]
            'joint_mask': torch.from_numpy(joint_mask),                 # [J_max]
            'frame_mask': torch.from_numpy(frame_mask),                 # [T_max]
            'adjacency': torch.from_numpy(adjacency),                   # [J_max, J_max]
            'geodesic_dist': torch.from_numpy(geodesic_dist),           # [J_max, J_max]
            # Root track (separate)
            'root_position': torch.from_numpy(padded_root_pos),         # [T_max, 3]
            'root_velocity': torch.from_numpy(padded_root_vel),         # [T_max, 3]
            # Decoder GT (skeleton-specific)
            'local_rotations_6d': torch.from_numpy(padded_rot),         # [T_max, J_max, 6]
            'has_rotations': has_rotations,
            # Auxiliary
            'foot_contact': torch.from_numpy(padded_contact),           # [T_max, 4]
            'bone_lengths': torch.from_numpy(padded_bones),             # [T_max, J_max]
            # Metadata
            'text': text,
            'num_joints': J,
            'num_frames': actual_frames,
            'skeleton_id': skeleton_id,
            'motion_id': sample_info['motion_id'],
        }


def collate_fn(batch: list[dict]) -> dict:
    """Custom collate function for variable-length text."""
    result = {}
    for key in batch[0]:
        if key == 'text':
            result[key] = [b[key] for b in batch]
        elif isinstance(batch[0][key], torch.Tensor):
            result[key] = torch.stack([b[key] for b in batch])
        elif isinstance(batch[0][key], (int, float)):
            result[key] = torch.tensor([b[key] for b in batch])
        else:
            result[key] = [b[key] for b in batch]
    return result