embedding_mixin.py

import os
import sqlite3
import networkx as nx
import numpy as np
import torch
from tqdm.auto import tqdm
from typing import Callable, Dict, List, Optional, Set
from torch.utils.data import DataLoader
from torch.utils.data import Dataset as TorchDataset
from transformers import PreTrainedTokenizerBase


class Pooler:
    def __init__(self, pooling_types: List[str]) -> None:
        self.pooling_types = pooling_types
        self.pooling_options: Dict[str, Callable] = {
            'mean': self.mean_pooling,
            'max': self.max_pooling,
            'norm': self.norm_pooling,
            'median': self.median_pooling,
            'std': self.std_pooling,
            'var': self.var_pooling,
            'cls': self.cls_pooling,
            'parti': self._pool_parti,
        }

    def _create_pooled_matrices_across_layers(self, attentions: torch.Tensor) -> torch.Tensor:
        assert isinstance(attentions, torch.Tensor)
        maxed_attentions = torch.max(attentions, dim=1)[0]
        return maxed_attentions

    def _page_rank(self, attention_matrix: np.ndarray, personalization: Optional[dict] = None, nstart: Optional[dict] = None, prune_type: str = "top_k_outdegree") -> Dict[int, float]:
        # Run PageRank on the attention matrix converted to a graph.
        # Raises exceptions if the graph doesn't match the token sequence or has no edges.
        # Returns the PageRank scores for each token node.
        G = self._convert_to_graph(attention_matrix)
        if G.number_of_nodes() != attention_matrix.shape[0]:
            raise Exception(
                f"The number of nodes in the graph should be equal to the number of tokens in sequence! You have {G.number_of_nodes()} nodes for {attention_matrix.shape[0]} tokens.")
        if G.number_of_edges() == 0:
            raise Exception(f"You don't seem to have any attention edges left in the graph.")

        return nx.pagerank(G, alpha=0.85, tol=1e-06, weight='weight', personalization=personalization, nstart=nstart, max_iter=100)

    def _convert_to_graph(self, matrix: np.ndarray) -> nx.DiGraph:
        # Convert a matrix (e.g., attention scores) to a directed graph using networkx.
        # Each element in the matrix represents a directed edge with a weight.
        G = nx.from_numpy_array(matrix, create_using=nx.DiGraph)
        return G

    def _calculate_importance_weights(self, dict_importance: Dict[int, float], attention_mask: Optional[torch.Tensor] = None) -> np.ndarray:
        # Remove keys where attention_mask is 0
        if attention_mask is not None:
            for k in list(dict_importance.keys()):
                if attention_mask[k] == 0:
                    del dict_importance[k]

        #dict_importance[0] # remove cls
        #dict_importance[-1] # remove eos
        total = sum(dict_importance.values())
        return np.array([v / total for _, v in dict_importance.items()])

    def _pool_parti(self, emb: torch.Tensor, attentions: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor: # (b, L, d) -> (b, d)
        maxed_attentions = self._create_pooled_matrices_across_layers(attentions).numpy()
        # emb is (b, L, d), maxed_attentions is (b, L, L)
        emb_pooled = []
        for e, a, mask in zip(emb, maxed_attentions, attention_mask):
            dict_importance = self._page_rank(a)
            importance_weights = self._calculate_importance_weights(dict_importance, mask)
            num_tokens = int(mask.sum().item())
            emb_pooled.append(np.average(e[:num_tokens], weights=importance_weights, axis=0))
        pooled = torch.tensor(np.array(emb_pooled))
        return pooled

    def mean_pooling(self, emb: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, **kwargs) -> torch.Tensor: # (b, L, d) -> (b, d)
        if attention_mask is None:
            return emb.mean(dim=1)
        else:
            attention_mask = attention_mask.unsqueeze(-1)
            return (emb * attention_mask).sum(dim=1) / attention_mask.sum(dim=1)

    def max_pooling(self, emb: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, **kwargs) -> torch.Tensor: # (b, L, d) -> (b, d)
        if attention_mask is None:
            return emb.max(dim=1).values
        else:
            attention_mask = attention_mask.unsqueeze(-1)
            return (emb * attention_mask).max(dim=1).values

    def norm_pooling(self, emb: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, **kwargs) -> torch.Tensor: # (b, L, d) -> (b, d)
        if attention_mask is None:
            return emb.norm(dim=1, p=2)
        else:
            attention_mask = attention_mask.unsqueeze(-1)
            return (emb * attention_mask).norm(dim=1, p=2)

    def median_pooling(self, emb: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, **kwargs) -> torch.Tensor: # (b, L, d) -> (b, d)
        if attention_mask is None:
            return emb.median(dim=1).values
        else:
            attention_mask = attention_mask.unsqueeze(-1)
            return (emb * attention_mask).median(dim=1).values
    
    def std_pooling(self, emb: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, **kwargs) -> torch.Tensor: # (b, L, d) -> (b, d)
        if attention_mask is None:
            return emb.std(dim=1)
        else:
            # Compute variance correctly over non-masked positions, then take sqrt
            var = self.var_pooling(emb, attention_mask, **kwargs)
            return torch.sqrt(var)
    
    def var_pooling(self, emb: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, **kwargs) -> torch.Tensor: # (b, L, d) -> (b, d)
        if attention_mask is None:
            return emb.var(dim=1)
        else:
            # Correctly compute variance over only non-masked positions
            attention_mask = attention_mask.unsqueeze(-1)  # (b, L, 1)
            # Compute mean over non-masked positions
            mean = (emb * attention_mask).sum(dim=1) / attention_mask.sum(dim=1)  # (b, d)
            mean = mean.unsqueeze(1)  # (b, 1, d)
            # Compute squared differences from mean, only over non-masked positions
            squared_diff = (emb - mean) ** 2  # (b, L, d)
            # Sum squared differences over non-masked positions and divide by count
            var = (squared_diff * attention_mask).sum(dim=1) / attention_mask.sum(dim=1)  # (b, d)
            return var

    def cls_pooling(self, emb: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, **kwargs) -> torch.Tensor: # (b, L, d) -> (b, d)
        return emb[:, 0, :]

    def __call__(
            self,
            emb: torch.Tensor,
            attention_mask: Optional[torch.Tensor] = None,
            attentions: Optional[torch.Tensor] = None
        ) -> torch.Tensor: # [mean, max]
        final_emb: List[torch.Tensor] = []
        for pooling_type in self.pooling_types:
            final_emb.append(self.pooling_options[pooling_type](emb=emb, attention_mask=attention_mask, attentions=attentions)) # (b, d)
        return torch.cat(final_emb, dim=-1) # (b, n_pooling_types * d)


class ProteinDataset(TorchDataset):
    """Simple dataset for protein sequences."""
    def __init__(self, sequences: List[str]) -> None:
        self.sequences = sequences

    def __len__(self) -> int:
        return len(self.sequences)

    def __getitem__(self, idx: int) -> str:
        return self.sequences[idx]


def build_collator(tokenizer: PreTrainedTokenizerBase) -> Callable[[List[str]], Dict[str, torch.Tensor]]:
    def _collate_fn(sequences: List[str]) -> Dict[str, torch.Tensor]:
        return tokenizer(sequences, return_tensors="pt", padding='longest')
    return _collate_fn


def parse_fasta(fasta_path: str) -> List[str]:
    assert os.path.exists(fasta_path), f"FASTA file does not exist: {fasta_path}"
    sequences = []
    current_seq = []
    with open(fasta_path, 'r') as f:
        for line in f:
            line = line.strip()
            if not line:
                continue
            if line.startswith('>'):
                if current_seq:
                    sequences.append(''.join(current_seq))
                    current_seq = []
            else:
                current_seq.append(line)
    if current_seq:
        sequences.append(''.join(current_seq))
    return sequences


class EmbeddingMixin:
    def _embed(self, input_ids: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
        raise NotImplementedError

    @property
    def device(self) -> torch.device:
        """Get the device of the model."""
        return next(self.parameters()).device

    def _read_sequences_from_db(self, db_path: str) -> Set[str]:
        """Read sequences from SQLite database."""
        sequences = []
        with sqlite3.connect(db_path) as conn:
            c = conn.cursor()
            c.execute("SELECT sequence FROM embeddings")
            while True:
                row = c.fetchone()
                if row is None:
                    break
                sequences.append(row[0])
        return set(sequences)

    def _ensure_embeddings_table(self, conn: sqlite3.Connection) -> None:
        cursor = conn.cursor()
        cursor.execute(
            "CREATE TABLE IF NOT EXISTS embeddings ("
            "sequence TEXT PRIMARY KEY, "
            "embedding BLOB NOT NULL, "
            "shape TEXT, "
            "dtype TEXT"
            ")"
        )
        cursor.execute("PRAGMA table_info(embeddings)")
        rows = cursor.fetchall()
        column_names = [row[1] for row in rows]
        if "shape" not in column_names:
            cursor.execute("ALTER TABLE embeddings ADD COLUMN shape TEXT")
        if "dtype" not in column_names:
            cursor.execute("ALTER TABLE embeddings ADD COLUMN dtype TEXT")
        conn.commit()

    def load_embeddings_from_pth(self, save_path: str) -> Dict[str, torch.Tensor]:
        assert os.path.exists(save_path), f"Embedding file does not exist: {save_path}"
        payload = torch.load(save_path, map_location="cpu", weights_only=True)
        assert isinstance(payload, dict), "Expected .pth embeddings file to contain a dictionary."
        for sequence, tensor in payload.items():
            assert isinstance(sequence, str), "Expected embedding dictionary keys to be sequences (str)."
            assert isinstance(tensor, torch.Tensor), "Expected embedding dictionary values to be tensors."
        return payload

    def load_embeddings_from_db(self, db_path: str, sequences: Optional[List[str]] = None) -> Dict[str, torch.Tensor]:
        assert os.path.exists(db_path), f"Embedding database does not exist: {db_path}"
        loaded: Dict[str, torch.Tensor] = {}
        with sqlite3.connect(db_path) as conn:
            self._ensure_embeddings_table(conn)
            cursor = conn.cursor()
            if sequences is None:
                cursor.execute("SELECT sequence, embedding, shape, dtype FROM embeddings")
            else:
                if len(sequences) == 0:
                    return loaded
                placeholders = ",".join(["?"] * len(sequences))
                cursor.execute(
                    f"SELECT sequence, embedding, shape, dtype FROM embeddings WHERE sequence IN ({placeholders})",
                    tuple(sequences),
                )

            rows = cursor.fetchall()
            for row in rows:
                sequence = row[0]
                embedding_bytes = row[1]
                shape_text = row[2]
                dtype_text = row[3]
                assert shape_text is not None, "Missing shape metadata in embeddings table."
                assert dtype_text is not None, "Missing dtype metadata in embeddings table."
                shape_values = [int(value) for value in shape_text.split(",") if len(value) > 0]
                assert len(shape_values) > 0, f"Invalid shape metadata for sequence: {sequence}"
                expected_size = int(np.prod(shape_values))
                np_dtype = np.dtype(dtype_text)
                array = np.frombuffer(embedding_bytes, dtype=np_dtype)
                assert array.size == expected_size, f"Shape mismatch while reading sequence: {sequence}"
                reshaped = array.copy().reshape(tuple(shape_values))
                loaded[sequence] = torch.from_numpy(reshaped)
        return loaded

    def embed_dataset(
        self,
        sequences: Optional[List[str]] = None,
        tokenizer: Optional[PreTrainedTokenizerBase] = None,
        batch_size: int = 2,
        max_len: int = 512,
        truncate: bool = True,
        full_embeddings: bool = False,
        embed_dtype: torch.dtype = torch.float32,
        pooling_types: List[str] = ['mean'],
        num_workers: int = 0,
        sql: bool = False,
        save: bool = True,
        sql_db_path: str = 'embeddings.db',
        save_path: str = 'embeddings.pth',
        fasta_path: Optional[str] = None,
        **kwargs,
    ) -> Optional[Dict[str, torch.Tensor]]:
        """
        Embed a dataset of protein sequences.

        Supports two modes:
        - Tokenizer mode (ESM2/ESM++): provide `tokenizer`, `_embed(input_ids, attention_mask)` is used.
        - Sequence mode (E1): pass `tokenizer=None`, `_embed(sequences, return_attention_mask=True, **kwargs)` is used.

        Sequences can be supplied as a list via `sequences`, parsed from a FASTA file via
        `fasta_path`, or both (the two sources are combined). At least one must be provided.
        """
        if fasta_path is not None:
            fasta_sequences = parse_fasta(fasta_path)
            sequences = list(sequences or []) + fasta_sequences
        assert sequences is not None and len(sequences) > 0, \
            "Must provide at least one sequence via `sequences` or `fasta_path`."
        sequences = list(set([seq[:max_len] if truncate else seq for seq in sequences]))
        sequences = sorted(sequences, key=len, reverse=True)
        hidden_size = self.config.hidden_size
        pooler = Pooler(pooling_types) if not full_embeddings else None
        tokenizer_mode = tokenizer is not None
        if tokenizer_mode:
            collate_fn = build_collator(tokenizer)
            device = self.device
        else:
            collate_fn = None
            device = None

        def get_embeddings(residue_embeddings: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
            assert isinstance(residue_embeddings, torch.Tensor)
            if full_embeddings or residue_embeddings.ndim == 2:
                return residue_embeddings
            return pooler(residue_embeddings, attention_mask)

        def iter_batches(to_embed: List[str]):
            if tokenizer_mode:
                assert collate_fn is not None
                assert device is not None
                dataset = ProteinDataset(to_embed)
                dataloader = DataLoader(dataset, batch_size=batch_size, num_workers=num_workers, collate_fn=collate_fn, shuffle=False)
                for i, batch in tqdm(enumerate(dataloader), total=len(dataloader), desc='Embedding batches'):
                    seqs = to_embed[i * batch_size:(i + 1) * batch_size]
                    input_ids = batch['input_ids'].to(device)
                    attention_mask = batch['attention_mask'].to(device)
                    residue_embeddings = self._embed(input_ids, attention_mask)
                    yield seqs, residue_embeddings, attention_mask
            else:
                for batch_start in tqdm(range(0, len(to_embed), batch_size), desc='Embedding batches'):
                    seqs = to_embed[batch_start:batch_start + batch_size]
                    batch_output = self._embed(seqs, return_attention_mask=True, **kwargs)
                    assert isinstance(batch_output, tuple), "Sequence mode _embed must return (last_hidden_state, attention_mask)."
                    assert len(batch_output) == 2, "Sequence mode _embed must return exactly two values."
                    residue_embeddings, attention_mask = batch_output
                    assert isinstance(attention_mask, torch.Tensor), "Sequence mode _embed must return attention_mask as a torch.Tensor."
                    yield seqs, residue_embeddings, attention_mask

        if sql:
            conn = sqlite3.connect(sql_db_path)
            self._ensure_embeddings_table(conn)
            c = conn.cursor()
            already_embedded = self._read_sequences_from_db(sql_db_path)
            to_embed = [seq for seq in sequences if seq not in already_embedded]
            print(f"Found {len(already_embedded)} already embedded sequences in {sql_db_path}")
            print(f"Embedding {len(to_embed)} new sequences")
            if len(to_embed) > 0:
                with torch.no_grad():
                    for i, (seqs, residue_embeddings, attention_mask) in enumerate(iter_batches(to_embed)):
                        embeddings = get_embeddings(residue_embeddings, attention_mask).to(embed_dtype)
                        for seq, emb, mask in zip(seqs, embeddings, attention_mask):
                            if full_embeddings:
                                emb = emb[mask.bool()].reshape(-1, hidden_size)
                            emb_np = emb.cpu().numpy()
                            emb_shape = ",".join([str(dim) for dim in emb_np.shape])
                            emb_dtype = str(emb_np.dtype)
                            c.execute(
                                "INSERT OR REPLACE INTO embeddings (sequence, embedding, shape, dtype) VALUES (?, ?, ?, ?)",
                                (seq, emb_np.tobytes(), emb_shape, emb_dtype),
                            )
                        if tokenizer_mode and (i + 1) % 100 == 0:
                            conn.commit()
                conn.commit()
            conn.close()
            return None

        embeddings_dict = {}
        if os.path.exists(save_path):
            embeddings_dict = self.load_embeddings_from_pth(save_path)
            to_embed = [seq for seq in sequences if seq not in embeddings_dict]
            print(f"Found {len(embeddings_dict)} already embedded sequences in {save_path}")
            print(f"Embedding {len(to_embed)} new sequences")
        else:
            to_embed = sequences
            print(f"Embedding {len(to_embed)} new sequences")

        if len(to_embed) > 0:
            with torch.no_grad():
                for seqs, residue_embeddings, attention_mask in iter_batches(to_embed):
                    embeddings = get_embeddings(residue_embeddings, attention_mask).to(embed_dtype)
                    for seq, emb, mask in zip(seqs, embeddings, attention_mask):
                        if full_embeddings:
                            emb = emb[mask.bool()].reshape(-1, hidden_size)
                        embeddings_dict[seq] = emb.cpu()

        if save:
            torch.save(embeddings_dict, save_path)

        return embeddings_dict


if __name__ == "__main__":
    # py -m pooler
    pooler = Pooler(pooling_types=['max', 'parti'])
    batch_size = 8
    seq_len = 64
    hidden_size = 128
    num_layers = 12
    emb = torch.randn(batch_size, seq_len, hidden_size)
    attentions = torch.randn(batch_size, num_layers, seq_len, seq_len)
    attention_mask = torch.ones(batch_size, seq_len)
    y = pooler(emb=emb, attention_mask=attention_mask, attentions=attentions)
    print(y.shape)