code/spec_rag/embeddings.py

from __future__ import annotations

from dataclasses import dataclass
from typing import Iterable, List, Optional

import numpy as np
from tqdm import tqdm
from .io import chunked


def l2_normalize(x: np.ndarray, eps: float = 1e-8) -> np.ndarray:
    norm = np.linalg.norm(x, axis=-1, keepdims=True)
    return x / (norm + eps)


@dataclass
class SMILESEmbedder:
    model_name: str
    device: str = "cpu"
    pooling: str = "cls"  # "cls" or "mean"
    batch_size: int = 64
    max_length: int = 256
    normalize: bool = True

    def _load(self):
        from transformers import AutoModel, AutoTokenizer

        tokenizer = AutoTokenizer.from_pretrained(self.model_name)
        # Use safetensors to avoid torch.load CVE (CVE-2025-32434) when torch < 2.6
        try:
            model = AutoModel.from_pretrained(self.model_name, use_safetensors=True)
        except Exception as e:
            raise RuntimeError(
                "Loading ChemBERTa failed (transformers require torch>=2.6 or safetensors). "
                "Upgrade with: pip install 'torch>=2.6', or ensure the model has .safetensors on the Hub."
            ) from e
        import torch
        if self.device == "cuda" and torch.cuda.device_count() > 1:
            model = torch.nn.DataParallel(model, device_ids=list(range(torch.cuda.device_count())))
        model.to(self.device)
        model.eval()
        return tokenizer, model

    def encode(self, smiles: Iterable[str]) -> np.ndarray:
        smiles_list = list(smiles)
        if not smiles_list:
            return np.zeros((0, 0), dtype=np.float32)
        tokenizer, model = self._load()
        outputs: List[np.ndarray] = []
        import torch

        n_batches = (len(smiles_list) + self.batch_size - 1) // self.batch_size
        with torch.no_grad():
            for _, batch in tqdm(
                chunked(smiles_list, self.batch_size),
                total=n_batches,
                desc="Encoding SMILES",
                unit="batch",
            ):
                toks = tokenizer(
                    batch,
                    padding=True,
                    truncation=True,
                    max_length=self.max_length,
                    return_tensors="pt",
                ).to(self.device)
                h = model(**toks).last_hidden_state
                if self.pooling == "mean":
                    mask = toks["attention_mask"].unsqueeze(-1).float()
                    pooled = (h * mask).sum(dim=1) / mask.sum(dim=1).clamp(min=1.0)
                else:
                    pooled = h[:, 0]
                outputs.append(pooled.detach().cpu().numpy())
        arr = np.concatenate(outputs, axis=0).astype(np.float32)
        if self.normalize:
            arr = l2_normalize(arr)
        return arr


@dataclass
class SpectrumEmbedder:
    specbridge_ckpt: str
    dreams_ckpt: Optional[str] = None
    d_out: int = 512
    mapper_hidden: int = 512
    n_blocks: int = 4
    chemberta_model: str = "seyonec/ChemBERTa-zinc-base-v1"
    device: str = "cpu"
    normalize: bool = True
    use_lightweight: bool = False

    def _load(self):
        import torch
        from argparse import Namespace
        from pathlib import PosixPath
        from pathlib import Path
        import sys
        import types

        specbridge_root = Path(__file__).resolve().parents[2] / "SpecBridge"
        dreams_root = specbridge_root / "DreaMS"
        for p in (specbridge_root, dreams_root):
            if p.exists() and str(p) not in sys.path:
                sys.path.insert(0, str(p))
        try:
            torch.serialization.add_safe_globals([Namespace, PosixPath])
        except Exception:
            pass

        from specbridge.adapters.dreams_adapter import load_dreams_encoder

        try:
            state = torch.load(self.specbridge_ckpt, map_location="cpu", weights_only=False)
        except TypeError:
            state = torch.load(self.specbridge_ckpt, map_location="cpu")
        model_state = state.get("model", state)
        ckpt_args = state.get("args", {}) if isinstance(state, dict) else {}

        d_out = self.d_out
        mapper_hidden = self.mapper_hidden
        n_blocks = self.n_blocks
        chemberta_model = self.chemberta_model
        spec_bins = 2048
        if isinstance(ckpt_args, dict):
            if ckpt_args.get("cond_dim") is not None:
                d_out = int(ckpt_args["cond_dim"])
            if ckpt_args.get("mapper_hidden") is not None:
                mapper_hidden = int(ckpt_args["mapper_hidden"])
            if ckpt_args.get("n_blocks") is not None:
                n_blocks = int(ckpt_args["n_blocks"])
            if ckpt_args.get("chemberta_model"):
                chemberta_model = str(ckpt_args["chemberta_model"])
            if ckpt_args.get("spec_bins") is not None:
                spec_bins = int(ckpt_args["spec_bins"])

        dreams_d_out = 1024
        if isinstance(model_state, dict) and "spec.proj.0.0.weight" in model_state:
            dreams_d_out = int(model_state["spec.proj.0.0.weight"].shape[1])
        dreams = load_dreams_encoder(self.dreams_ckpt, d_in=spec_bins, d_out=dreams_d_out)
        self._dreams_is_dummy = dreams.__class__.__name__ == "DummyDreams"

        if self.use_lightweight:
            if isinstance(model_state, dict):
                if "spec.proj.0.0.weight" in model_state:
                    d_out = int(model_state["spec.proj.0.0.weight"].shape[0])
                elif "mapB.W.weight" in model_state:
                    d_out = int(model_state["mapB.W.weight"].shape[1])
                if "mapB.blocks.0.fc1.weight" in model_state:
                    mapper_hidden = int(model_state["mapB.blocks.0.fc1.weight"].shape[0])
                block_ids = set()
                for key in model_state.keys():
                    if key.startswith("mapB.blocks."):
                        parts = key.split(".")
                        if len(parts) > 2 and parts[2].isdigit():
                            block_ids.add(int(parts[2]))
                if block_ids:
                    n_blocks = max(block_ids) + 1

            from transformers import AutoConfig
            from specbridge.adapters.dreams_adapter import DreamsAdapter
            from specbridge.models.mapper import ProcrustesResidualMapper

            hid = int(AutoConfig.from_pretrained(chemberta_model).hidden_size)
            spec = DreamsAdapter(
                dreams_encoder=dreams,
                d_out=d_out,
                hidden=mapper_hidden,
                freeze_backbone=True,
            )
            mapB = ProcrustesResidualMapper(
                d_in=d_out,
                d_out=hid,
                n_blocks=n_blocks,
                hidden=mapper_hidden,
                gaussian=True,
                random_init=False,
            )

            class _LightSpecBridge(torch.nn.Module):
                def __init__(self, spec, mapB):
                    super().__init__()
                    self.spec = spec
                    self.mapB = mapB

            model = _LightSpecBridge(spec, mapB)
            model._dreams_is_dummy = self._dreams_is_dummy
            model.load_state_dict(model_state, strict=False)
            model.to(self.device)
            model.eval()
            return model

        from specbridge.models.mapper import DreamsToMolCondition

        model = DreamsToMolCondition(
            dreams_encoder=dreams,
            d_out=d_out,
            mapper_hidden=mapper_hidden,
            gaussian=True,
            mol_space="chemberta",
            chemberta_model=chemberta_model,
            args=type("Args", (), {"n_blocks": n_blocks, "random_mapper_init": False})(),
            freeze_backbone=True,
        )
        model.load_state_dict(model_state, strict=False)
        model._dreams_is_dummy = self._dreams_is_dummy
        model.to(self.device)
        model.eval()
        return model

    def encode(self, spectra_binned, meta: dict, batch_size: int | None = None) -> np.ndarray:
        import torch

        model = self._load()
        if not isinstance(spectra_binned, torch.Tensor):
            spectra_binned = torch.tensor(spectra_binned, dtype=torch.float32)
        total = spectra_binned.shape[0]
        if batch_size is None or batch_size <= 0:
            batch_size = total
        outputs = []
        use_peaks = not getattr(model, "_dreams_is_dummy", False)
        with torch.no_grad():
            for start in tqdm(range(0, total, batch_size)):
                end = min(total, start + batch_size)
                batch = spectra_binned[start:end].to(self.device)
                batch_meta = meta
                if isinstance(meta, dict):
                    batch_meta = {}
                    for k, v in meta.items():
                        if k == "peaks" and not use_peaks:
                            continue
                        if isinstance(v, torch.Tensor) and v.shape[0] == total:
                            batch_meta[k] = v[start:end].to(self.device)
                        else:
                            batch_meta[k] = v
                z_s = model.spec(batch, batch_meta)
                mu_s, _ = model.mapB(z_s)
                outputs.append(mu_s.detach().cpu().numpy().astype(np.float32))
        emb = np.concatenate(outputs, axis=0) if outputs else np.zeros((0, 0), dtype=np.float32)
        if self.normalize:
            emb = l2_normalize(emb)
        return emb

    def encode_spec_only(
        self, spectra_binned, meta: dict, batch_size: int | None = None
    ) -> np.ndarray:
        """Return E_mist(spec): spectrum embedding before mapper (d_spec), for mapper training."""
        import torch

        model = self._load()
        if not isinstance(spectra_binned, torch.Tensor):
            spectra_binned = torch.tensor(spectra_binned, dtype=torch.float32)
        total = spectra_binned.shape[0]
        if batch_size is None or batch_size <= 0:
            batch_size = total
        outputs = []
        use_peaks = not getattr(model, "_dreams_is_dummy", False)
        with torch.no_grad():
            for start in range(0, total, batch_size):
                end = min(total, start + batch_size)
                batch = spectra_binned[start:end].to(self.device)
                batch_meta = meta
                if isinstance(meta, dict):
                    batch_meta = {}
                    for k, v in meta.items():
                        if k == "peaks" and not use_peaks:
                            continue
                        if isinstance(v, torch.Tensor) and v.shape[0] == total:
                            batch_meta[k] = v[start:end].to(self.device)
                        else:
                            batch_meta[k] = v
                z_s = model.spec(batch, batch_meta)
                outputs.append(z_s.detach().cpu().numpy().astype(np.float32))
        return (
            np.concatenate(outputs, axis=0)
            if outputs
            else np.zeros((0, 0), dtype=np.float32)
        )