File size: 9,254 Bytes
23680f2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
278
279
280
"""Clean HyCoCLIP embedding provider (PyTorch) - no external hycoclip package.

This is a minimal reimplementation that loads HyCoCLIP weights directly.
Only depends on torch, timm, and numpy.

Architecture:
- ViT backbone (timm)
- Linear projection to embedding space
- Exponential map to hyperboloid (Lorentz model)

Checkpoints: https://huggingface.co/avik-pal/hycoclip

Requirements:
    uv sync --extra ml
"""

from __future__ import annotations

import math
from pathlib import Path
from typing import Any

import numpy as np

from hyperview.core.sample import Sample
from hyperview.embeddings.providers import (
    BaseEmbeddingProvider,
    ModelSpec,
    register_provider,
)

__all__ = ["HyCoCLIPProvider"]


HYCOCLIP_CHECKPOINTS: dict[str, str] = {
    "hycoclip_vit_s": "hf://avik-pal/hycoclip#hycoclip_vit_s.pth",
    "hycoclip_vit_b": "hf://avik-pal/hycoclip#hycoclip_vit_b.pth",
    "meru_vit_s": "hf://avik-pal/hycoclip#meru_vit_s.pth",
    "meru_vit_b": "hf://avik-pal/hycoclip#meru_vit_b.pth",
}


def _exp_map_lorentz(x: "torch.Tensor", c: float) -> "torch.Tensor":
    """Exponential map from tangent space at the hyperboloid vertex.

    Maps Euclidean tangent vectors at the origin onto the Lorentz (hyperboloid)
    model of hyperbolic space with curvature -c.

    Output is ordered as (t, x1, ..., xD) and satisfies:
        t^2 - ||x||^2 = 1/c

    This matches HyCoCLIP/MERU exp_map0 numerics by clamping the sinh input for
    stability and inferring the time component from the hyperboloid constraint.

    Args:
        x: Euclidean tangent vectors at the origin, shape (..., D).
        c: Positive curvature parameter (hyperbolic curvature is -c).

    Returns:
        Hyperboloid coordinates, shape (..., D + 1).
    """
    import torch

    if c <= 0:
        raise ValueError(f"curvature c must be > 0, got {c}")

    # Compute in float32 under AMP to avoid float16/bfloat16 overflow.
    if x.dtype in (torch.float16, torch.bfloat16):
        x = x.float()

    sqrt_c = math.sqrt(c)
    rc_xnorm = sqrt_c * torch.norm(x, dim=-1, keepdim=True)

    eps = 1e-8
    sinh_input = torch.clamp(rc_xnorm, min=eps, max=math.asinh(2**15))
    spatial = torch.sinh(sinh_input) * x / torch.clamp(rc_xnorm, min=eps)

    t = torch.sqrt((1.0 / c) + torch.sum(spatial * spatial, dim=-1, keepdim=True))
    return torch.cat([t, spatial], dim=-1)


def _create_encoder(
    embed_dim: int = 512,
    curvature: float = 0.1,
    vit_model: str = "vit_small_patch16_224",
) -> "nn.Module":
    """Create HyCoCLIP image encoder using timm ViT backbone."""
    import timm
    import torch.nn as nn

    class HyCoCLIPImageEncoder(nn.Module):
        def __init__(self) -> None:
            super().__init__()
            self.backbone = timm.create_model(vit_model, pretrained=False, num_classes=0)
            backbone_dim = int(getattr(self.backbone, "embed_dim"))
            self.proj = nn.Linear(backbone_dim, embed_dim, bias=False)
            self.curvature = curvature
            self.embed_dim = embed_dim

        def forward(self, x: "torch.Tensor") -> "torch.Tensor":
            features = self.backbone(x)
            spatial = self.proj(features)
            return _exp_map_lorentz(spatial, self.curvature)

    return HyCoCLIPImageEncoder()


def _load_encoder(checkpoint_path: str, device: str = "cpu") -> Any:
    """Load HyCoCLIP image encoder from checkpoint."""
    import torch

    ckpt = torch.load(checkpoint_path, map_location="cpu", weights_only=False)
    state = ckpt["model"]

    # Extract curvature (stored as log)
    curvature = torch.exp(state["curv"]).item()

    # Determine model variant from checkpoint
    proj_shape = state["visual_proj.weight"].shape
    embed_dim = proj_shape[0]
    backbone_dim = proj_shape[1]

    vit_models = {
        384: "vit_small_patch16_224",
        768: "vit_base_patch16_224",
        1024: "vit_large_patch16_224",
    }
    vit_model = vit_models.get(backbone_dim, "vit_small_patch16_224")

    model = _create_encoder(embed_dim=embed_dim, curvature=curvature, vit_model=vit_model)

    # Remap checkpoint keys
    new_state = {}
    for key, value in state.items():
        if key.startswith("visual."):
            new_state["backbone." + key[7:]] = value
        elif key == "visual_proj.weight":
            new_state["proj.weight"] = value

    model.load_state_dict(new_state, strict=False)
    return model.to(device).eval()


class HyCoCLIPProvider(BaseEmbeddingProvider):
    """Clean HyCoCLIP provider (PyTorch) - no hycoclip package dependency.

    Requires: torch, torchvision, timm (install via `uv sync --extra ml`)
    """

    def __init__(self) -> None:
        self._model: Any = None
        self._model_spec: ModelSpec | None = None
        self._device: Any = None
        self._transform: Any = None

    @property
    def provider_id(self) -> str:
        return "hycoclip"

    def _get_device(self) -> Any:
        import torch

        if self._device is None:
            self._device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        return self._device

    def _get_transform(self) -> Any:
        if self._transform is None:
            from torchvision import transforms

            self._transform = transforms.Compose([
                transforms.Resize(224, transforms.InterpolationMode.BICUBIC),
                transforms.CenterCrop(224),
                transforms.ToTensor(),
                transforms.Normalize(
                    mean=[0.48145466, 0.4578275, 0.40821073],
                    std=[0.26862954, 0.26130258, 0.27577711],
                ),
            ])
        return self._transform

    def _resolve_checkpoint(self, checkpoint: str) -> Path:
        """Resolve checkpoint path, downloading from HuggingFace if needed."""
        # Handle HuggingFace Hub URLs: hf://repo_id#filename
        if checkpoint.startswith("hf://"):
            from huggingface_hub import hf_hub_download

            path = checkpoint[5:]
            if "#" not in path:
                raise ValueError(f"HF checkpoint must include filename: {checkpoint}")
            repo_id, filename = path.split("#", 1)
            return Path(hf_hub_download(repo_id=repo_id, filename=filename)).resolve()

        # Local path
        path = Path(checkpoint).expanduser().resolve()
        if not path.exists():
            raise FileNotFoundError(f"Checkpoint not found: {path}")
        return path

    def _load_model(self, model_spec: ModelSpec) -> None:
        if self._model is not None and self._model_spec == model_spec:
            return

        # Auto-resolve checkpoint from model_id if not provided
        checkpoint = model_spec.checkpoint
        if not checkpoint:
            checkpoint = HYCOCLIP_CHECKPOINTS.get(model_spec.model_id)
            if not checkpoint:
                available = ", ".join(sorted(HYCOCLIP_CHECKPOINTS.keys()))
                raise ValueError(
                    f"Unknown HyCoCLIP model_id: '{model_spec.model_id}'. "
                    f"Known models: {available}. "
                    f"Or provide 'checkpoint' path explicitly."
                )

        checkpoint_path = self._resolve_checkpoint(checkpoint)
        self._model = _load_encoder(str(checkpoint_path), str(self._get_device()))
        self._model_spec = model_spec

    def compute_embeddings(
        self,
        samples: list["Sample"],
        model_spec: ModelSpec,
        batch_size: int = 32,
        show_progress: bool = True,
    ) -> np.ndarray:
        """Compute hyperboloid embeddings for samples."""
        import torch

        self._load_model(model_spec)
        assert self._model is not None

        device = self._get_device()
        transform = self._get_transform()

        if show_progress:
            print(f"Computing HyCoCLIP embeddings for {len(samples)} samples...")

        all_embeddings = []

        for i in range(0, len(samples), batch_size):
            batch_samples = samples[i : i + batch_size]

            images = []
            for sample in batch_samples:
                img = sample.load_image()
                if img.mode != "RGB":
                    img = img.convert("RGB")
                images.append(transform(img))

            batch_tensor = torch.stack(images).to(device)

            with torch.no_grad(), torch.amp.autocast(
                device_type=device.type, enabled=device.type == "cuda"
            ):
                embeddings = self._model(batch_tensor)

            all_embeddings.append(embeddings.cpu().numpy())

        return np.concatenate(all_embeddings, axis=0)

    def get_space_config(self, model_spec: ModelSpec, dim: int) -> dict[str, Any]:
        """Return embedding space configuration with curvature."""
        self._load_model(model_spec)
        assert self._model is not None

        return {
            "provider": self.provider_id,
            "model_id": model_spec.model_id,
            "geometry": "hyperboloid",
            "checkpoint": model_spec.checkpoint,
            "dim": dim,
            "curvature": self._model.curvature,
            "spatial_dim": self._model.embed_dim,
        }


# Auto-register on import
register_provider("hycoclip", HyCoCLIPProvider)