File size: 18,008 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 | """
SparseTopkEmbCache — Sparse top-K attention delta @ gene_emb features.
Uses precomputed per-cell sparse attention (K=300) to compute delta top-K
weighted gene embeddings at training time. Filters 99.3% noise compared to
dense attention delta.
HDF5 layout (from precompute_sparse_attn.py):
/attn_values (N, G_full, K) float16 — top-K attention values per row
/attn_indices (N, G_full, K) int16 — column indices in G_full space
/cell_names (N,) string
/valid_gene_mask (G_full,) bool
"""
import h5py
import numpy as np
import torch
class SparseTopkEmbCache:
"""
Sparse top-K attention delta @ gene_emb cache.
At lookup time:
1. Read sparse attention for src/tgt cells from HDF5
2. Scatter to dense, compute delta = tgt - src
3. Take top-K (default 30) by |delta|
4. Weighted sum with gene embeddings → (B, G_sub, 512)
5. Normalize with pre-computed statistics
"""
def __init__(self, h5_path, gene_emb, top_k=30, target_std=1.0,
norm_n_pairs=2000, seed=42):
"""
Args:
h5_path: path to sparse attention HDF5 cache
gene_emb: (G_full, D) CPU tensor of gene embeddings
top_k: number of top delta entries to keep per gene
target_std: target standard deviation for normalization
norm_n_pairs: number of random cell pairs for norm statistics
seed: random seed for norm sampling
"""
self.h5_path = h5_path
self.top_k = top_k
self.target_std = target_std
self.h5 = h5py.File(h5_path, "r")
self.attn_values = self.h5["attn_values"] # (N, G_full, K)
self.attn_indices = self.h5["attn_indices"] # (N, G_full, K)
self.G_full = self.attn_values.shape[1]
self.K_sparse = self.attn_values.shape[2]
cell_names = self.h5["cell_names"].asstr()[:]
self.name_to_idx = {name: i for i, name in enumerate(cell_names)}
if "valid_gene_mask" in self.h5:
self.valid_gene_mask = torch.from_numpy(
self.h5["valid_gene_mask"][:].astype(bool))
else:
self.valid_gene_mask = torch.ones(self.G_full, dtype=torch.bool)
self.gene_emb = gene_emb # (G_full, D) CPU tensor
assert gene_emb.shape[0] == self.G_full, (
f"gene_emb shape {gene_emb.shape} != G_full {self.G_full}")
print(f" SparseTopkEmbCache: {len(self.name_to_idx)} cells, "
f"G_full={self.G_full}, K_sparse={self.K_sparse}, top_k={top_k}")
print(f" gene_emb shape: {gene_emb.shape}, "
f"valid genes: {self.valid_gene_mask.sum().item()}")
# Compute normalization statistics
self._compute_norm_stats(norm_n_pairs, seed)
def _compute_norm_stats(self, n_pairs, seed):
"""Sample random cell pairs and compute norm statistics for the 512-d output."""
rng = np.random.RandomState(seed)
cell_names_list = list(self.name_to_idx.keys())
n_cells = len(cell_names_list)
src_idx = rng.randint(0, n_cells, size=n_pairs)
tgt_idx = rng.randint(0, n_cells, size=n_pairs)
# Sample gene subset for faster norm estimation
valid_positions = torch.where(self.valid_gene_mask)[0]
n_sample_genes = min(500, len(valid_positions))
sample_perm = rng.choice(len(valid_positions), n_sample_genes, replace=False)
sample_gene_idx = valid_positions[sample_perm]
sample_gene_idx = torch.sort(sample_gene_idx)[0]
D = self.gene_emb.shape[1]
sum_x = torch.zeros(D, dtype=torch.float64)
sum_x2 = torch.zeros(D, dtype=torch.float64)
total_n = 0
batch_size = 64
print(f" Computing norm stats from {n_pairs} cell pairs, "
f"{n_sample_genes} sampled genes...")
for i in range(0, n_pairs, batch_size):
batch_end = min(i + batch_size, n_pairs)
batch_src = [cell_names_list[j] for j in src_idx[i:batch_end]]
batch_tgt = [cell_names_list[j] for j in tgt_idx[i:batch_end]]
feats = self._compute_features(
batch_src, batch_tgt, sample_gene_idx, torch.device("cpu"))
flat = feats.reshape(-1, D).double()
sum_x += flat.sum(dim=0)
sum_x2 += (flat ** 2).sum(dim=0)
total_n += flat.shape[0]
self.norm_mean = (sum_x / total_n).float()
self.norm_var = (sum_x2 / total_n - (sum_x / total_n) ** 2).float()
# Clamp variance to avoid division by zero
self.norm_var = self.norm_var.clamp(min=1e-8)
print(f" Norm stats computed: mean [{self.norm_mean.min():.4f}, "
f"{self.norm_mean.max():.4f}], "
f"std [{self.norm_var.sqrt().min():.4f}, "
f"{self.norm_var.sqrt().max():.4f}]")
def _compute_features(self, src_cell_names, tgt_cell_names, gene_indices, device):
"""
Compute raw (unnormalized) sparse topk @ gene_emb features.
Args:
src_cell_names: list of str, control cell names
tgt_cell_names: list of str, perturbation cell names
gene_indices: (G_sub,) tensor or None (all genes)
device: torch device for computation
Returns:
(B, G_sub, D) tensor of raw features
"""
B = len(src_cell_names)
D = self.gene_emb.shape[1]
if gene_indices is not None:
gene_idx_np = gene_indices.cpu().numpy()
G_sub = len(gene_idx_np)
else:
gene_idx_np = None
G_sub = self.G_full
# Step 1: Collect unique cells, read HDF5 once
seen = {}
unique_names = []
for n in src_cell_names + tgt_cell_names:
if n not in seen:
seen[n] = len(unique_names)
unique_names.append(n)
unique_h5_idx = [self.name_to_idx[n] for n in unique_names]
sorted_order = np.argsort(unique_h5_idx)
sorted_h5_idx = [unique_h5_idx[i] for i in sorted_order]
# Read from HDF5 (sorted for sequential access)
raw_vals = self.attn_values[sorted_h5_idx] # (U, G_full, K) float16
raw_idxs = self.attn_indices[sorted_h5_idx] # (U, G_full, K) int16
# Unsort to match unique_names order
unsort = np.argsort(sorted_order)
raw_vals = raw_vals[unsort]
raw_idxs = raw_idxs[unsort]
# Select gene subset in numpy (before GPU transfer)
if gene_idx_np is not None:
raw_vals = raw_vals[:, gene_idx_np, :] # (U, G_sub, K)
raw_idxs = raw_idxs[:, gene_idx_np, :]
# Step 2: Map to src/tgt batch order
src_map = [seen[n] for n in src_cell_names]
tgt_map = [seen[n] for n in tgt_cell_names]
# Step 3: Convert to torch and move to device
src_vals = torch.from_numpy(raw_vals[src_map].astype(np.float32)).to(device)
src_idxs = torch.from_numpy(raw_idxs[src_map].astype(np.int64)).to(device)
tgt_vals = torch.from_numpy(raw_vals[tgt_map].astype(np.float32)).to(device)
tgt_idxs = torch.from_numpy(raw_idxs[tgt_map].astype(np.int64)).to(device)
gene_emb_d = self.gene_emb.to(device)
# Step 4: Process in chunks (100 genes per chunk to limit memory)
chunk_size = 100
output = torch.zeros(B, G_sub, D, device=device)
for c_start in range(0, G_sub, chunk_size):
c_end = min(c_start + chunk_size, G_sub)
c_len = c_end - c_start
sv = src_vals[:, c_start:c_end, :] # (B, c_len, K)
si = src_idxs[:, c_start:c_end, :]
tv = tgt_vals[:, c_start:c_end, :]
ti = tgt_idxs[:, c_start:c_end, :]
# Scatter sparse entries to dense attention rows
src_dense = torch.zeros(B, c_len, self.G_full, device=device)
tgt_dense = torch.zeros(B, c_len, self.G_full, device=device)
src_dense.scatter_(-1, si, sv)
tgt_dense.scatter_(-1, ti, tv)
# Delta attention
delta = tgt_dense - src_dense # (B, c_len, G_full)
# Top-k by absolute delta value
_, topk_idx = delta.abs().topk(self.top_k, dim=-1) # (B, c_len, top_k)
topk_delta = delta.gather(-1, topk_idx) # (B, c_len, top_k)
# Gather gene embeddings at top-k positions
flat_idx = topk_idx.reshape(-1) # (B * c_len * top_k,)
topk_emb = gene_emb_d[flat_idx].reshape(
B, c_len, self.top_k, D) # (B, c_len, top_k, D)
# Weighted sum: delta_values * gene_emb → (B, c_len, D)
chunk_feat = (topk_delta.unsqueeze(-1) * topk_emb).sum(dim=2)
output[:, c_start:c_end, :] = chunk_feat
return output
def lookup_delta(self, src_cell_names, tgt_cell_names, gene_indices, device=None):
"""
Compute normalized sparse topk @ gene_emb features for (src, tgt) 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 features
"""
if device is None:
device = torch.device("cpu")
feats = self._compute_features(
src_cell_names, tgt_cell_names, gene_indices, device)
# Normalize: (x - mean) / sqrt(var) * target_std
eps = 1e-6
norm_mean = self.norm_mean.to(device)
norm_var = self.norm_var.to(device)
feats = (feats - norm_mean) / (norm_var.sqrt() + eps)
feats = feats * self.target_std
return feats
def close(self):
self.h5.close()
def __del__(self):
try:
self.h5.close()
except Exception:
pass
def _read_sparse_batch(h5_values, h5_indices, name_to_idx,
src_cell_names, tgt_cell_names, gene_idx_np=None):
"""
Shared HDF5 reading logic for sparse caches.
Returns:
src_vals, src_idxs, tgt_vals, tgt_idxs: numpy arrays (B, G_sub, K)
"""
seen = {}
unique_names = []
for n in src_cell_names + tgt_cell_names:
if n not in seen:
seen[n] = len(unique_names)
unique_names.append(n)
unique_h5_idx = [name_to_idx[n] for n in unique_names]
sorted_order = np.argsort(unique_h5_idx)
sorted_h5_idx = [unique_h5_idx[i] for i in sorted_order]
raw_vals = h5_values[sorted_h5_idx]
raw_idxs = h5_indices[sorted_h5_idx]
unsort = np.argsort(sorted_order)
raw_vals = raw_vals[unsort]
raw_idxs = raw_idxs[unsort]
if gene_idx_np is not None:
raw_vals = raw_vals[:, gene_idx_np, :]
raw_idxs = raw_idxs[:, gene_idx_np, :]
src_map = [seen[n] for n in src_cell_names]
tgt_map = [seen[n] for n in tgt_cell_names]
return raw_vals[src_map], raw_idxs[src_map], raw_vals[tgt_map], raw_idxs[tgt_map]
class SparsePCADeltaCache:
"""
Sparse PCA delta cache: topk filtering + PCA projection.
Same topk30 filtering as SparseTopkEmbCache, but projects through
precomputed PCA basis (d-dim) instead of gene_emb (512-dim):
1. scatter sparse K=300 → dense delta (G_full)
2. topk(top_k) by |delta|
3. delta_vals @ pca_basis[topk_idx] → (d,)
"""
def __init__(self, h5_path, top_k=30, target_std=1.0,
norm_n_pairs=2000, seed=42):
self.h5_path = h5_path
self.top_k = top_k
self.target_std = target_std
self.h5 = h5py.File(h5_path, "r")
self.attn_values = self.h5["attn_values"]
self.attn_indices = self.h5["attn_indices"]
self.G_full = self.attn_values.shape[1]
self.K_sparse = self.attn_values.shape[2]
cell_names = self.h5["cell_names"].asstr()[:]
self.name_to_idx = {name: i for i, name in enumerate(cell_names)}
if "valid_gene_mask" in self.h5:
self.valid_gene_mask = torch.from_numpy(
self.h5["valid_gene_mask"][:].astype(bool))
else:
self.valid_gene_mask = torch.ones(self.G_full, dtype=torch.bool)
# Load PCA basis
self.pca_basis = torch.from_numpy(self.h5["pca_basis"][:]).float() # (G_full, d)
self.pca_dim = self.pca_basis.shape[1]
print(f" SparsePCADeltaCache: {len(self.name_to_idx)} cells, "
f"G_full={self.G_full}, K_sparse={self.K_sparse}, "
f"top_k={top_k}, PCA dim={self.pca_dim}")
self._compute_norm_stats(norm_n_pairs, seed)
def _compute_norm_stats(self, n_pairs, seed):
"""Sample random cell pairs and compute norm statistics for PCA-d output."""
rng = np.random.RandomState(seed)
cell_names_list = list(self.name_to_idx.keys())
n_cells = len(cell_names_list)
src_idx = rng.randint(0, n_cells, size=n_pairs)
tgt_idx = rng.randint(0, n_cells, size=n_pairs)
valid_positions = torch.where(self.valid_gene_mask)[0]
n_sample_genes = min(500, len(valid_positions))
sample_perm = rng.choice(len(valid_positions), n_sample_genes, replace=False)
sample_gene_idx = valid_positions[sample_perm]
sample_gene_idx = torch.sort(sample_gene_idx)[0]
D = self.pca_dim
sum_x = torch.zeros(D, dtype=torch.float64)
sum_x2 = torch.zeros(D, dtype=torch.float64)
total_n = 0
batch_size = 64
print(f" Computing PCA norm stats from {n_pairs} cell pairs, "
f"{n_sample_genes} sampled genes...")
for i in range(0, n_pairs, batch_size):
batch_end = min(i + batch_size, n_pairs)
batch_src = [cell_names_list[j] for j in src_idx[i:batch_end]]
batch_tgt = [cell_names_list[j] for j in tgt_idx[i:batch_end]]
feats = self._compute_features(
batch_src, batch_tgt, sample_gene_idx, torch.device("cpu"))
flat = feats.reshape(-1, D).double()
sum_x += flat.sum(dim=0)
sum_x2 += (flat ** 2).sum(dim=0)
total_n += flat.shape[0]
self.norm_mean = (sum_x / total_n).float()
self.norm_var = (sum_x2 / total_n - (sum_x / total_n) ** 2).float()
self.norm_var = self.norm_var.clamp(min=1e-8)
print(f" Norm stats computed: mean [{self.norm_mean.min():.4f}, "
f"{self.norm_mean.max():.4f}], "
f"std [{self.norm_var.sqrt().min():.4f}, "
f"{self.norm_var.sqrt().max():.4f}]")
def _compute_features(self, src_cell_names, tgt_cell_names, gene_indices, device):
"""
Compute topk-filtered PCA-projected delta features.
Flow per chunk:
1. scatter sparse K=300 → dense (B, chunk, G_full)
2. delta = tgt_dense - src_dense
3. topk(top_k) by |delta|
4. delta_vals @ pca_basis[topk_idx] → (B, chunk, pca_dim)
Returns: (B, G_sub, pca_dim) tensor
"""
B = len(src_cell_names)
D = self.pca_dim
gene_idx_np = gene_indices.cpu().numpy() if gene_indices is not None else None
G_sub = len(gene_idx_np) if gene_idx_np is not None else self.G_full
sv_np, si_np, tv_np, ti_np = _read_sparse_batch(
self.attn_values, self.attn_indices, self.name_to_idx,
src_cell_names, tgt_cell_names, gene_idx_np)
src_vals = torch.from_numpy(sv_np.astype(np.float32)).to(device)
src_idxs = torch.from_numpy(si_np.astype(np.int64)).to(device)
tgt_vals = torch.from_numpy(tv_np.astype(np.float32)).to(device)
tgt_idxs = torch.from_numpy(ti_np.astype(np.int64)).to(device)
pca_d = self.pca_basis.to(device) # (G_full, d)
chunk_size = 100
output = torch.zeros(B, G_sub, D, device=device)
for c_start in range(0, G_sub, chunk_size):
c_end = min(c_start + chunk_size, G_sub)
c_len = c_end - c_start
sv = src_vals[:, c_start:c_end, :] # (B, c_len, K)
si = src_idxs[:, c_start:c_end, :]
tv = tgt_vals[:, c_start:c_end, :]
ti = tgt_idxs[:, c_start:c_end, :]
# Scatter sparse → dense
src_dense = torch.zeros(B, c_len, self.G_full, device=device)
tgt_dense = torch.zeros(B, c_len, self.G_full, device=device)
src_dense.scatter_(-1, si, sv)
tgt_dense.scatter_(-1, ti, tv)
# Delta + topk
delta = tgt_dense - src_dense # (B, c_len, G_full)
_, topk_idx = delta.abs().topk(self.top_k, dim=-1) # (B, c_len, top_k)
topk_delta = delta.gather(-1, topk_idx) # (B, c_len, top_k)
# Gather PCA basis at topk positions & weighted sum
flat_idx = topk_idx.reshape(-1)
topk_pca = pca_d[flat_idx].reshape(
B, c_len, self.top_k, D) # (B, c_len, top_k, d)
chunk_feat = (topk_delta.unsqueeze(-1) * topk_pca).sum(dim=2)
output[:, c_start:c_end, :] = chunk_feat
return output
def lookup_delta(self, src_cell_names, tgt_cell_names, gene_indices, device=None):
"""
Compute normalized PCA-projected delta features for (src, tgt) pairs.
Returns: (B, G_sub, pca_dim) tensor, normalized features
"""
if device is None:
device = torch.device("cpu")
feats = self._compute_features(
src_cell_names, tgt_cell_names, gene_indices, device)
eps = 1e-6
norm_mean = self.norm_mean.to(device)
norm_var = self.norm_var.to(device)
feats = (feats - norm_mean) / (norm_var.sqrt() + eps)
feats = feats * self.target_std
return feats
def close(self):
self.h5.close()
def __del__(self):
try:
self.h5.close()
except Exception:
pass
|