File size: 7,301 Bytes
411f238
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from __future__ import annotations
import hashlib
import json
import math
import re
from pathlib import Path
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F

VARIANTS = {
    "small_cnn": {"width": 32, "depth": 3, "chunk": 1, "dropout": 0.05},
    "base_cnn": {"width": 64, "depth": 4, "chunk": 1, "dropout": 0.10},
    "chunk_cnn": {"width": 48, "depth": 4, "chunk": 8, "dropout": 0.10},
    "ensemble_cnn": {"width": 64, "depth": 5, "chunk": 8, "dropout": 0.15},
}

class VisionBackbone(nn.Module):
    def __init__(self, width, depth):
        super().__init__()
        layers = [nn.Conv2d(3, width, 5, 2, 2), nn.GroupNorm(max(1, width // 8), width), nn.SiLU()]
        channels = width
        for index in range(depth - 1):
            next_channels = min(width * (2 ** min(index + 1, 2)), 256)
            layers += [nn.Conv2d(channels, next_channels, 3, 2, 1),
                       nn.GroupNorm(max(1, next_channels // 8), next_channels), nn.SiLU()]
            channels = next_channels
        layers += [nn.AdaptiveAvgPool2d((1, 1)), nn.Flatten(), nn.Linear(channels, 384), nn.SiLU()]
        self.net = nn.Sequential(*layers)
    def forward(self, images):
        return self.net(images)

class PolicyNet(nn.Module):
    def __init__(self, config):
        super().__init__()
        variant = config["variant"]
        cfg = VARIANTS[variant]
        self.chunk = int(cfg["chunk"])
        self.vision = VisionBackbone(int(cfg["width"]), int(cfg["depth"]))
        self.proprio = nn.Sequential(nn.Linear(int(config["proprio_dim"]), 128), nn.LayerNorm(128), nn.SiLU())
        self.text = nn.Sequential(nn.Linear(int(config["text_dim"]), 128), nn.LayerNorm(128), nn.SiLU())
        self.task = nn.Embedding(int(config["task_count"]), 32)
        self.difficulty = nn.Embedding(int(config["difficulty_count"]), 16)
        self.head = nn.Sequential(
            nn.Linear(384 + 128 + 128 + 32 + 16, 512), nn.LayerNorm(512), nn.SiLU(),
            nn.Dropout(float(cfg["dropout"])), nn.Linear(512, 256), nn.SiLU(),
            nn.Linear(256, self.chunk * 7))
    def forward(self, images, proprio, text, task_id, difficulty_id):
        if images.ndim != 4:
            raise ValueError("image batch must have four dimensions")
        if images.shape[-1] == 3:
            images = images.permute(0, 3, 1, 2)
        images = images.float()
        if images.max() > 2:
            images = images / 255.0
        images = F.interpolate(images, size=(96, 96), mode="bilinear", align_corners=False)
        fused = torch.cat([
            self.vision(images), self.proprio(proprio.float()), self.text(text.float()),
            self.task(task_id.clamp(0, self.task.num_embeddings - 1)),
            self.difficulty(difficulty_id.clamp(0, self.difficulty.num_embeddings - 1)),
        ], dim=-1)
        return torch.tanh(self.head(fused)).reshape(-1, self.chunk, 7)

def _text_vector(text, dim):
    result = np.zeros(dim, dtype=np.float32)
    for token in re.findall(r"[a-z0-9_]+", str(text).lower()):
        value = int.from_bytes(hashlib.blake2b(token.encode(), digest_size=8).digest(), "little")
        result[value % dim] += 1.0 if value & 1 else -1.0
    norm = float(np.linalg.norm(result))
    return result / norm if norm else result

def _proprio(obs, config):
    value = np.asarray(obs.get("proprio", np.zeros(25, dtype=np.float32)), dtype=np.float32).reshape(-1)
    if value.size == 21:
        value = np.pad(value, (0, 4))
    if value.size != 25:
        raise ValueError(f"proprio must have 25 values, got {value.size}")
    step = float(obs.get("step", 0))
    horizon = float(obs.get("horizon", 320) or 320)
    value = np.concatenate([value, np.asarray([step / max(horizon, 1.0)], dtype=np.float32)])
    if value.size < int(config["proprio_dim"]):
        value = np.pad(value, (0, int(config["proprio_dim"]) - value.size))
    value = value[: int(config["proprio_dim"])]
    mean = np.asarray(config["proprio_mean"], dtype=np.float32)
    std = np.asarray(config["proprio_std"], dtype=np.float32)
    return ((value - mean) / np.maximum(std, 1e-4)).astype(np.float32)

class Policy:
    def __init__(self, model, config, device):
        self.model, self.config, self.device = model, config, device
        self.chunk = int(config["chunk"])
        self.last_step = None
        self.chunks = {}
        self.last_action = np.zeros(7, dtype=np.float32)
        self.task_to_id = config["task_to_id"]
        self.difficulty_to_id = config["difficulty_to_id"]
    @torch.inference_mode()
    def act(self, obs):
        image = np.asarray(obs.get("image", np.zeros((96, 96, 3), dtype=np.uint8)))
        if image.ndim == 2:
            image = np.repeat(image[..., None], 3, axis=-1)
        if image.shape[-1] == 4:
            image = image[..., :3]
        if image.shape[-1] != 3:
            raise ValueError("image must have 3 or 4 channels")
        step = int(obs.get("step", 0))
        if self.last_step is None or step == 0 or step != self.last_step + 1:
            self.chunks = {}
        if step == self.last_step:
            return self.last_action.copy()
        task = str(obs.get("task", ""))
        difficulty = str(obs.get("difficulty", "") or "")
        text = f"task {task} difficulty {difficulty} instruction {obs.get('instruction', '')}"
        text_feature = _text_vector(text, int(self.config["text_dim"]))
        proprio = _proprio(obs, self.config)
        task_id = self.task_to_id.get(task, len(self.task_to_id))
        difficulty_id = self.difficulty_to_id.get(difficulty, len(self.difficulty_to_id))
        raw = self.model(
            torch.from_numpy(image).unsqueeze(0).to(self.device),
            torch.from_numpy(proprio).unsqueeze(0).to(self.device),
            torch.from_numpy(text_feature).unsqueeze(0).to(self.device),
            torch.tensor([task_id], device=self.device),
            torch.tensor([difficulty_id], device=self.device),
        )[0].float().cpu().numpy()
        self.chunks[step] = raw
        candidates, weights = [], []
        for start, chunk in list(self.chunks.items()):
            offset = step - start
            if 0 <= offset < self.chunk:
                candidates.append(chunk[offset])
                weights.append(math.exp(-0.55 * offset))
            else:
                del self.chunks[start]
        action = np.average(np.asarray(candidates), axis=0, weights=np.asarray(weights))
        action = np.clip(action, -1.0, 1.0).astype(np.float32)
        if action[6] > 0.12:
            action[6] = 1.0
        elif action[6] < -0.12:
            action[6] = -1.0
        self.last_step, self.last_action = step, action.copy()
        return action

def load_policy(model_dir: str, device: str, dtype: str):
    root = Path(model_dir)
    config = json.loads((root / "vla_config.json").read_text())
    selected = torch.device(device if str(device).startswith("cuda") and torch.cuda.is_available() else "cpu")
    model = PolicyNet(config).to(selected)
    checkpoint = torch.load(root / "model.pt", map_location=selected, weights_only=True)
    state = checkpoint["state_dict"] if "state_dict" in checkpoint else checkpoint
    model.load_state_dict(state, strict=True)
    model.eval()
    return Policy(model, config, selected)