File size: 8,419 Bytes
fcb9b70
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Exact architecture and integrity-checked loader for YellowCab v0."""

from __future__ import annotations

import hashlib
import json
from collections import Counter
from pathlib import Path
from typing import Any, Mapping

import torch
from safetensors.torch import load_file
from torch import nn
from torchvision.models import efficientnet_b0


class CheckpointIntegrityError(RuntimeError):
    """Raised when a release checkpoint does not match its signed-off contract."""


class TemporalFusionHead(nn.Module):
    """The trained temporal and telemetry fusion head."""

    def __init__(
        self,
        *,
        image_feature_dim: int,
        telemetry_dim: int,
        num_classes: int,
        hidden_dim: int = 256,
        dropout: float = 0.25,
    ) -> None:
        super().__init__()
        self.image_feature_dim = image_feature_dim
        self.image_norm = nn.LayerNorm(image_feature_dim)
        self.temporal = nn.GRU(
            input_size=image_feature_dim,
            hidden_size=hidden_dim,
            batch_first=True,
        )
        self.telemetry = nn.Sequential(
            nn.LayerNorm(telemetry_dim),
            nn.Linear(telemetry_dim, 96),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(96, 96),
            nn.GELU(),
        )
        self.classifier = nn.Sequential(
            nn.LayerNorm(hidden_dim + 96),
            nn.Linear(hidden_dim + 96, hidden_dim),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim, num_classes),
        )

    def forward(
        self,
        image_features: torch.Tensor,
        telemetry: torch.Tensor,
    ) -> torch.Tensor:
        temporal_output, _ = self.temporal(self.image_norm(image_features))
        telemetry_output = self.telemetry(telemetry)
        fused = torch.cat((temporal_output[:, -1], telemetry_output), dim=-1)
        return self.classifier(fused)


class TaxiManeuverModel(nn.Module):
    """Frozen EfficientNet-B0 plus the trained v0 fusion head."""

    def __init__(self, head_config: Mapping[str, Any]) -> None:
        super().__init__()
        self.encoder = efficientnet_b0(weights=None)
        self.encoder.classifier = nn.Identity()
        self.head = TemporalFusionHead(**dict(head_config))

    def forward(
        self,
        images: torch.Tensor,
        telemetry: torch.Tensor,
    ) -> torch.Tensor:
        if images.ndim != 5 or images.shape[1] != 3 or images.shape[2] != 3:
            raise ValueError("images must have shape [batch, 3, 3, height, width]")
        batch_size, frame_count, channels, height, width = images.shape
        flat_images = images.reshape(
            batch_size * frame_count,
            channels,
            height,
            width,
        )
        image_features = self.encoder(flat_images).reshape(
            batch_size,
            frame_count,
            -1,
        )
        return self.head(image_features, telemetry)


def load_release_config(repository_path: str | Path) -> dict[str, Any]:
    """Load and minimally validate the one canonical release configuration."""

    root = Path(repository_path)
    path = root / "config.json"
    try:
        config = json.loads(path.read_text(encoding="utf-8"))
    except (OSError, json.JSONDecodeError) as exc:
        raise CheckpointIntegrityError("release configuration is unavailable") from exc
    if config.get("schema") != "gdc_taxi_maneuver_release_config_v1":
        raise CheckpointIntegrityError("unsupported release configuration schema")
    for identity_field in ("model_id", "model_name", "model_version"):
        identity_value = config.get(identity_field)
        if not isinstance(identity_value, str) or not identity_value.strip():
            raise CheckpointIntegrityError(
                f"release configuration {identity_field} is invalid"
            )
    labels = config.get("labels")
    feature_order = config.get("telemetry", {}).get("feature_order")
    if not isinstance(labels, list) or len(labels) != 5 or len(set(labels)) != 5:
        raise CheckpointIntegrityError("canonical label configuration is invalid")
    if (
        not isinstance(feature_order, list)
        or len(feature_order) != 12
        or len(set(feature_order)) != 12
    ):
        raise CheckpointIntegrityError("canonical telemetry configuration is invalid")
    return config


def sha256_file(path: str | Path) -> str:
    digest = hashlib.sha256()
    with Path(path).open("rb") as handle:
        for chunk in iter(lambda: handle.read(1024 * 1024), b""):
            digest.update(chunk)
    return digest.hexdigest()


def _verify_checkpoint_metadata(
    *,
    checkpoint_path: Path,
    weights: Mapping[str, torch.Tensor],
    model: nn.Module,
    checkpoint_config: Mapping[str, Any],
) -> None:
    expected_bytes = int(checkpoint_config["bytes"])
    expected_sha256 = str(checkpoint_config["sha256"])
    if checkpoint_path.stat().st_size != expected_bytes:
        raise CheckpointIntegrityError("checkpoint byte size does not match config.json")
    if sha256_file(checkpoint_path) != expected_sha256:
        raise CheckpointIntegrityError("checkpoint SHA-256 does not match config.json")

    expected_state = model.state_dict()
    actual_keys = set(weights)
    expected_keys = set(expected_state)
    missing = sorted(expected_keys - actual_keys)
    unexpected = sorted(actual_keys - expected_keys)
    if missing or unexpected:
        raise CheckpointIntegrityError(
            "checkpoint tensor keys do not exactly match the model architecture"
        )

    for name, expected_tensor in expected_state.items():
        actual_tensor = weights[name]
        if actual_tensor.shape != expected_tensor.shape:
            raise CheckpointIntegrityError(
                f"checkpoint tensor shape mismatch for {name}"
            )
        if actual_tensor.dtype != expected_tensor.dtype:
            raise CheckpointIntegrityError(
                f"checkpoint tensor dtype mismatch for {name}"
            )

    actual_tensor_count = len(weights)
    actual_state_numel = sum(tensor.numel() for tensor in weights.values())
    actual_dtype_counts = Counter(str(tensor.dtype) for tensor in weights.values())
    expected_dtype_counts = {
        str(name): int(count)
        for name, count in checkpoint_config["state_dtypes"].items()
    }
    if actual_tensor_count != int(checkpoint_config["state_tensor_count"]):
        raise CheckpointIntegrityError("checkpoint tensor count does not match config.json")
    if actual_state_numel != int(checkpoint_config["state_numel"]):
        raise CheckpointIntegrityError("checkpoint state size does not match config.json")
    if dict(actual_dtype_counts) != expected_dtype_counts:
        raise CheckpointIntegrityError("checkpoint dtype inventory does not match config.json")

    parameter_count = sum(parameter.numel() for parameter in model.parameters())
    if parameter_count != int(checkpoint_config["parameter_count"]):
        raise CheckpointIntegrityError(
            "model parameter count does not match config.json"
        )


def load_model(
    repository_path: str | Path,
    *,
    device: torch.device | str = "cpu",
) -> tuple[TaxiManeuverModel, dict[str, Any]]:
    """Load the checkpoint only after hash, keys, shapes and dtypes all match."""

    root = Path(repository_path)
    config = load_release_config(root)
    checkpoint_config = config["checkpoint"]
    checkpoint_path = root / str(checkpoint_config["file"])
    try:
        weights = load_file(str(checkpoint_path), device="cpu")
    except (OSError, ValueError) as exc:
        raise CheckpointIntegrityError("checkpoint cannot be read") from exc

    model = TaxiManeuverModel(config["architecture"]["head_config"])
    _verify_checkpoint_metadata(
        checkpoint_path=checkpoint_path,
        weights=weights,
        model=model,
        checkpoint_config=checkpoint_config,
    )
    try:
        model.load_state_dict(weights, strict=True)
    except RuntimeError as exc:
        raise CheckpointIntegrityError("strict checkpoint loading failed") from exc
    model.eval()
    model.requires_grad_(False)
    model.to(torch.device(device))
    return model, config


__all__ = [
    "CheckpointIntegrityError",
    "TaxiManeuverModel",
    "TemporalFusionHead",
    "load_model",
    "load_release_config",
    "sha256_file",
]