File size: 12,923 Bytes
4cde5e8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Wrist-only RGB-D Stereo-ACT-cross_relbias trainer for RoboFactory.

One shared policy sees only one local Panda wrist RGB-D stream at a time. RGB
uses frozen DINOv3-B/16; native ManiSkill depth (stored in millimetres) uses
frozen DeFM-S/14. Both become 30x40 patches before two region-aligned
cross_relbias fusion blocks and the standard ACT 4/7 CVAE policy.
"""
from __future__ import annotations

import argparse
import json
import os
from collections import Counter, OrderedDict, defaultdict
from pathlib import Path

import h5py
import numpy as np
import torch
import torch.nn as nn
from torch.utils.data import DataLoader, Dataset

from rgbd_patch_fusion import RGBDPatchFusion
from train_act import ACT, EpisodeBlockBatchSampler, _stats, _trajectories, seed_everything
from five_task_contract import hierarchical_item_weights


DEPTH_MM_TO_M = 0.001


class WristRGBDACTDataset(Dataset):
    def __init__(self, trajectories, horizon, stats, train, *, preload=True, cache_limit=0):
        self.horizon, self.stats, self.cache_limit = horizon, stats, int(cache_limit)
        kept = [item for index, item in enumerate(trajectories) if (index % 10 != 0) == train]
        self.items = [(path, key, t, arm, task) for path, key, n, present, task in kept
                      for arm in present for t in range(n)]
        self.item_tasks = [task for _, _, _, _, task in self.items]
        self.item_weights = hierarchical_item_weights(kept, self.items)
        self.stream_indices = defaultdict(list)
        for index, (path, key, _t, arm, task) in enumerate(self.items):
            self.stream_indices[(path, key, arm, task)].append(index)
        self.cache = OrderedDict()
        if preload:
            for path, key, _, present, _ in kept:
                for arm in present:
                    self._episode(path, key, arm)

    def __len__(self):
        return len(self.items)

    def _episode(self, path, key, arm):
        tag = (path, key, arm)
        if tag not in self.cache:
            with h5py.File(path, "r") as h5:
                tr = h5[key]
                sensor = tr["obs"]["sensor_data"][f"head_camera_agent{arm}"]
                if "depth" not in sensor:
                    raise ValueError(f"Stereo-ACT needs rgbd corpus; missing depth in {path}:{key}")
                rgb, depth = sensor["rgb"][:], sensor["depth"][:]
                if tuple(rgb.shape[1:]) != (480, 640, 3) or tuple(depth.shape[1:]) != (480, 640, 1):
                    raise ValueError(
                        f"strict 640x480 RGB-D required; {path}:{key}:panda-{arm} has "
                        f"rgb={tuple(rgb.shape[1:])}, depth={tuple(depth.shape[1:])}"
                    )
                self.cache[tag] = (
                    rgb, depth,
                    tr["obs"]["agent"][f"panda-{arm}"]["qpos"][:].astype(np.float32),
                    tr["actions"][f"panda-{arm}"][:].astype(np.float32),
                )
            if self.cache_limit > 0:
                while len(self.cache) > self.cache_limit:
                    self.cache.popitem(last=False)
        else:
            self.cache.move_to_end(tag)
        return self.cache[tag]

    def __getitem__(self, index):
        path, key, t, arm, _ = self.items[index]
        rgb, depth, qpos, actions = self._episode(path, key, arm)
        future = actions[t:t + self.horizon]
        valid = len(future)
        padded = np.empty((self.horizon, actions.shape[1]), np.float32)
        padded[:valid], padded[valid:] = future, future[-1]
        mask = np.zeros(self.horizon, np.bool_); mask[:valid] = True
        return (
            torch.from_numpy(rgb[t]).permute(2, 0, 1).contiguous(),
            torch.from_numpy(depth[t]).permute(2, 0, 1).contiguous(),
            torch.from_numpy((qpos[t] - self.stats["q_mean"]) / self.stats["q_std"]),
            torch.from_numpy((padded - self.stats["a_mean"]) / self.stats["a_std"]),
            torch.from_numpy(mask),
        )


class StereoACT(ACT):
    def __init__(self, state_dim, action_dim, horizon=100, d_model=384, enc_layers=4, dec_layers=7,
                 dino_model="facebook/dinov3-vitb16-pretrain-lvd1689m", defm_model="defm_vit_s14"):
        super().__init__(state_dim, action_dim, horizon, d_model, enc_layers, dec_layers,
                         vision_backbone="dinov3_vitb16_frozen", dino_model=dino_model)
        from defm.model_factory import create_defm_model
        self.defm_model_name = defm_model
        # Use an explicitly versioned local checkpoint when supplied.  This
        # prevents an interrupted Hub download from silently changing the
        # depth encoder during a formal run.
        checkpoint = os.environ.get("DEFM_CHECKPOINT")
        self.defm = create_defm_model(
            defm_model, pretrained=True, pretrained_path=checkpoint
        ).eval()
        self.defm.requires_grad_(False)
        self.depth_proj = nn.Linear(384, d_model)
        self.fusion = RGBDPatchFusion(d_model=d_model, heads=8, grid_h=30, grid_w=40,
                                      layers=2, ffn_dim=d_model * 4)
        self.fusion_pos = nn.Parameter(torch.randn(1, 30 * 40, d_model) * 0.02)

    def train(self, mode=True):
        super().train(mode)
        self.vision.eval(); self.defm.eval()
        return self

    def _rgbd_tokens(self, rgb, depth_mm):
        # Parent RGB method preserves native 640x480 -> 30x40 DINO patches.
        rgb_tokens = super()._vision_tokens(rgb)
        from defm.utils.utils import preprocess_depth_batch
        depth_m = depth_mm.float().squeeze(1).mul(DEPTH_MM_TO_M)
        prepared = preprocess_depth_batch(depth_m, target_size=(420, 560), patch_size=14, device=rgb.device)
        self.defm.eval()
        with torch.no_grad():
            spatial, _ = self.defm.to(rgb.device).get_intermediate_layers(
                prepared.float(), n=1, reshape=True, return_class_token=True
            )[0]
        depth_tokens = self.depth_proj(spatial.flatten(2).transpose(1, 2).to(dtype=rgb_tokens.dtype))
        if rgb_tokens.shape[1] != 30 * 40 or depth_tokens.shape[1] != 30 * 40:
            raise ValueError(
                f"strict aligned 30x40 tokens required, got RGB={rgb_tokens.shape[1]} "
                f"and depth={depth_tokens.shape[1]}"
            )
        return self.fusion(rgb_tokens, depth_tokens, self.fusion_pos.to(dtype=rgb_tokens.dtype))

    def forward(self, image, depth_mm, qpos, actions=None):
        x = self._rgbd_tokens(image, depth_mm)
        state = self.state(qpos).unsqueeze(1)
        if actions is not None:
            h = self.posterior(self.action(actions) + self.pos)
            mu, logvar = self.latent(h.mean(1)).chunk(2, -1)
            z = mu + torch.randn_like(mu) * torch.exp(.5 * logvar)
        else:
            mu = logvar = None
            z = torch.zeros((image.shape[0], self.z_proj.in_features), device=image.device)
        memory = torch.cat((state, self.z_proj(z).unsqueeze(1), x), dim=1)
        return self.out(self.decoder(self.query.expand(image.shape[0], -1, -1), memory)), mu, logvar


def loss(model, rgb, depth, qpos, actions, mask, beta):
    rgb = rgb.float().div_(255)
    with torch.autocast("cuda", dtype=torch.bfloat16):
        pred, mu, logvar = model(rgb, depth, qpos, actions)
        mse = ((pred - actions).square().mean(-1) * mask).sum() / mask.sum().clamp_min(1)
        kl = -.5 * (1 + logvar - mu.square() - logvar.exp()).sum(-1).mean()
    return mse + beta * kl, mse, kl


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--data", required=True)
    parser.add_argument("--shared-arms", default="0,1")
    parser.add_argument("--output", required=True)
    # Verified on the target 32 GB RTX 5090 with a real 640×480 RGB-D canary:
    # batch=32 completes DINO+DeFM plus two 30×40 fusion blocks' backward pass.
    parser.add_argument("--batch-size", type=int, default=32)
    parser.add_argument("--updates", type=int, default=60000)
    parser.add_argument("--save-updates", default="20000,40000,60000")
    parser.add_argument("--lr", type=float, default=2e-4)
    parser.add_argument("--beta", type=float, default=1e-3)
    parser.add_argument("--workers", type=int, default=0)
    parser.add_argument("--cache-episodes", type=int, default=0,
                        help="Bound decoded episode cache; 0 preloads all trajectories.")
    parser.add_argument("--episode-block-updates", type=int, default=64)
    parser.add_argument("--task-balanced", action="store_true")
    parser.add_argument("--seed", type=int, default=20260724)
    args = parser.parse_args()
    import glob
    arms = tuple(int(item) for item in args.shared_arms.split(","))
    paths = sorted({path for item in args.data.split(",") for path in glob.glob(item)})
    trajectories = _trajectories(paths, arms)
    if len(trajectories) < 10:
        raise ValueError("need at least 10 successful RGB-D demonstrations")
    seed_everything(args.seed); torch.backends.cudnn.benchmark = True
    stats = _stats(trajectories, arms)
    lazy_cache = args.cache_episodes > 0
    if lazy_cache and args.workers:
        raise ValueError("bounded RGB-D cache requires --workers 0 to keep a single cache")
    train = WristRGBDACTDataset(trajectories, 100, stats, True, preload=not lazy_cache,
                                cache_limit=args.cache_episodes)
    valid = WristRGBDACTDataset(trajectories, 100, stats, False, preload=False,
                                cache_limit=args.cache_episodes)
    counts = Counter(train.item_tasks)
    sampler = None
    if lazy_cache:
        sampler = EpisodeBlockBatchSampler(train, args.batch_size, args.updates,
                                           args.episode_block_updates, args.seed, args.task_balanced)
        loader = DataLoader(train, batch_sampler=sampler, num_workers=0, pin_memory=True)
    elif args.task_balanced and len(counts) > 1:
        weights = torch.as_tensor(train.item_weights, dtype=torch.double)
        sampler = torch.utils.data.WeightedRandomSampler(weights, num_samples=len(weights), replacement=True)
    if not lazy_cache:
        loader = DataLoader(train, batch_size=args.batch_size, shuffle=sampler is None, sampler=sampler,
                            drop_last=True, num_workers=args.workers, pin_memory=True,
                            persistent_workers=args.workers > 0)
    device = torch.device("cuda:0")
    sample = train[0]
    model = StereoACT(len(sample[2]), len(sample[3][0])).to(device)
    optimizer = torch.optim.AdamW((p for p in model.parameters() if p.requires_grad), lr=args.lr, weight_decay=1e-4)
    scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, args.updates)
    output = Path(args.output); output.mkdir(parents=True, exist_ok=True)
    config = vars(args) | {"horizon": 100, "enc_layers": 4, "dec_layers": 7, "d_model": 384,
        "vision_backbone": "stereo_act_cross_relbias",
        "dino_model": "facebook/dinov3-vitb16-pretrain-lvd1689m", "defm_model": model.defm_model_name,
        "defm_checkpoint": os.environ.get("DEFM_CHECKPOINT"),
        "camera_width": 640, "camera_height": 480, "patch_grid": [30, 40], "fusion_layers": 2,
        "depth_storage_unit": "millimeters", "depth_to_meters_scale": DEPTH_MM_TO_M,
        "arms": arms, "state_dim": len(sample[2]), "action_dim": len(sample[3][0]),
        "files": paths, "episodes": len(trajectories), "train_task_item_counts": dict(counts)}
    (output / "config.json").write_text(json.dumps(config, indent=2))
    np.savez(output / "normalization.npz", **stats)
    milestones, updates = {int(x) for x in args.save_updates.split(",") if x}, 0
    while updates < args.updates:
        model.train(); totals = {"loss": 0.0, "mse": 0.0, "kl": 0.0, "n": 0}
        for rgb, depth, qpos, actions, mask in loader:
            rgb, depth, qpos, actions, mask = (item.to(device, non_blocking=True) for item in (rgb, depth, qpos, actions, mask))
            optimizer.zero_grad(set_to_none=True)
            total, mse, kl = loss(model, rgb, depth, qpos, actions, mask, args.beta)
            total.backward(); torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
            optimizer.step(); scheduler.step(); updates += 1
            batch = len(rgb)
            for name, value in (("loss", total), ("mse", mse), ("kl", kl)):
                totals[name] += float(value.detach()) * batch
            totals["n"] += batch
            if updates in milestones:
                torch.save({"model": model.state_dict(), "stats": stats, "config": config, "update": updates},
                           output / f"checkpoint_{updates:06d}.pt")
            if updates % 100 == 0:
                print(json.dumps({"update": updates, **{key: value / totals["n"] for key, value in totals.items() if key != "n"}}), flush=True)
            if updates >= args.updates:
                break


if __name__ == "__main__":
    main()