File size: 5,683 Bytes
9f5e507
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
ScGPT feature caches — Load pre-extracted scGPT features from HDF5.

ScGPTFeatureCache: per-cell encoder features (for feature_mode="encoder").
ScGPTAttnDeltaCache: per-cell attn@gene_emb features (for feature_mode="attention_delta").
    Stores raw per-cell features; computes delta (tgt - src) at lookup time.
"""

import h5py
import numpy as np
import torch


class ScGPTFeatureCache:
    """
    Loads pre-extracted per-gene scGPT features from HDF5 and provides
    batch lookup by cell name and gene indices.

    HDF5 layout:
        /features    (N, G_full, scgpt_dim) float16
        /norm_mean   (scgpt_dim,) float32
        /norm_var    (scgpt_dim,) float32
        /cell_names  (N,) string
    """

    def __init__(self, h5_path: str, target_std: float = 1.0):
        self.h5_path = h5_path
        self.target_std = target_std

        self.h5 = h5py.File(h5_path, "r")
        self.features = self.h5["features"]  # lazy dataset, shape (N, G, D)
        self.norm_mean = torch.from_numpy(self.h5["norm_mean"][:]).float()  # (D,)
        self.norm_var = torch.from_numpy(self.h5["norm_var"][:]).float()    # (D,)

        # Build cell_name -> row index mapping
        cell_names = self.h5["cell_names"].asstr()[:]
        self.name_to_idx = {name: i for i, name in enumerate(cell_names)}

    def lookup(self, cell_names, gene_indices, device=None) -> torch.Tensor:
        """
        Retrieve pre-extracted features for a batch.

        Args:
            cell_names: list of str, cell identifiers from batch
            gene_indices: (G_sub,) tensor, gene subset indices
            device: target torch device

        Returns:
            (B, G_sub, D) tensor, normalized features
        """
        # Map cell names to HDF5 row indices
        row_indices = np.array([self.name_to_idx[n] for n in cell_names])

        # h5py fancy indexing requires sorted, unique indices
        unique_indices, inverse = np.unique(row_indices, return_inverse=True)
        raw = self.features[unique_indices.tolist()]  # (U, G_full, D) as numpy

        # Map back to original batch order (handles duplicates)
        raw = raw[inverse]

        # Select gene subset
        gene_idx_np = gene_indices.cpu().numpy()
        raw = raw[:, gene_idx_np, :]      # (B, G_sub, D)

        z = torch.from_numpy(raw.astype(np.float32))

        # Normalize: (x - mean) / sqrt(var) * target_std
        eps = 1e-6
        z = (z - self.norm_mean) / (self.norm_var.sqrt() + eps)
        z = z * self.target_std

        if device is not None:
            z = z.to(device)

        return z

    def close(self):
        self.h5.close()

    def __del__(self):
        try:
            self.h5.close()
        except Exception:
            pass


class ScGPTAttnDeltaCache:
    """
    Attention-delta cache: stores per-cell attn@gene_emb features.
    At lookup time, computes delta = tgt_raw - src_raw, selects gene subset, normalizes.

    HDF5 layout (same as ScGPTFeatureCache):
        /features    (N, G_full, scgpt_dim) float16 — raw features per cell
        /norm_mean   (scgpt_dim,) float32 — mean of delta features
        /norm_var    (scgpt_dim,) float32 — var of delta features
        /cell_names  (N,) string
    """

    def __init__(self, h5_path: str, target_std: float = 1.0):
        self.h5_path = h5_path
        self.target_std = target_std

        self.h5 = h5py.File(h5_path, "r")
        self.features = self.h5["features"]  # lazy dataset, shape (N, G_full, D)
        self.norm_mean = torch.from_numpy(self.h5["norm_mean"][:]).float()  # (D,)
        self.norm_var = torch.from_numpy(self.h5["norm_var"][:]).float()    # (D,)

        # Build cell_name -> row index mapping
        cell_names = self.h5["cell_names"].asstr()[:]
        self.name_to_idx = {name: i for i, name in enumerate(cell_names)}

    def _lookup_raw(self, cell_names) -> np.ndarray:
        """Retrieve raw features for a batch of cells. Returns (B, G_full, D) numpy fp16."""
        row_indices = np.array([self.name_to_idx[n] for n in cell_names])

        # h5py fancy indexing requires sorted, unique indices
        unique_indices, inverse = np.unique(row_indices, return_inverse=True)
        raw = self.features[unique_indices.tolist()]  # (U, G_full, D)
        raw = raw[inverse]  # (B, G_full, D)
        return raw

    def lookup_delta(self, src_cell_names, tgt_cell_names, gene_indices, device=None):
        """
        Compute normalized attention delta for (src, tgt) cell pairs.

        Args:
            src_cell_names: list of str, control cell identifiers
            tgt_cell_names: list of str, perturbation cell identifiers
            gene_indices: (G_sub,) tensor, gene subset indices
            device: target torch device

        Returns:
            (B, G_sub, D) tensor, normalized delta features
        """
        src_raw = self._lookup_raw(src_cell_names)  # (B, G_full, D) fp16
        tgt_raw = self._lookup_raw(tgt_cell_names)  # (B, G_full, D) fp16

        # Delta in float32
        delta = tgt_raw.astype(np.float32) - src_raw.astype(np.float32)

        # Select gene subset
        gene_idx_np = gene_indices.cpu().numpy()
        delta = delta[:, gene_idx_np, :]  # (B, G_sub, D)

        z = torch.from_numpy(delta)

        # Normalize: (x - mean) / sqrt(var) * target_std
        eps = 1e-6
        z = (z - self.norm_mean) / (self.norm_var.sqrt() + eps)
        z = z * self.target_std

        if device is not None:
            z = z.to(device)

        return z

    def close(self):
        self.h5.close()

    def __del__(self):
        try:
            self.h5.close()
        except Exception:
            pass