File size: 7,189 Bytes
bfd9ff4 | 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 | # From-scratch Stage 1 Cell-State JEPA fixed by RESEARCH_PROPOSAL.md.
# The encoder sees gene/value tokens only: no context, action, batch, or cell-line IDs.
import math
from copy import deepcopy
from pathlib import Path
import torch
from torch import nn
from torch.nn import functional as F
class CellEncoder(nn.Module):
"""Compact Perceiver encoder for unordered continuous-expression tokens."""
def __init__(
self,
vocab_size=3_000,
token_dim=192,
latent_queries=32,
blocks=3,
heads=6,
ffn_dim=768,
dropout=0.10,
cell_dim=256,
):
super().__init__()
self.padding_id = vocab_size
self.gene_embedding = nn.Embedding(vocab_size + 1, token_dim, padding_idx=vocab_size)
self.value_encoder = nn.Sequential(
nn.Linear(1, token_dim), nn.GELU(), nn.Linear(token_dim, token_dim)
)
self.token_norm = nn.LayerNorm(token_dim)
self.latents = nn.Parameter(torch.randn(latent_queries, token_dim) * 0.02)
self.query_norm = nn.LayerNorm(token_dim)
self.cross_attention = nn.MultiheadAttention(
token_dim, heads, dropout=dropout, batch_first=True
)
self.cross_ffn = nn.Sequential(
nn.LayerNorm(token_dim),
nn.Linear(token_dim, ffn_dim),
nn.GELU(),
nn.Dropout(dropout),
nn.Linear(ffn_dim, token_dim),
nn.Dropout(dropout),
)
layer = nn.TransformerEncoderLayer(
token_dim,
heads,
ffn_dim,
dropout,
activation="gelu",
batch_first=True,
norm_first=True,
)
self.blocks = nn.TransformerEncoder(
layer, blocks, nn.LayerNorm(token_dim), enable_nested_tensor=False
)
self.output = nn.Sequential(nn.LayerNorm(token_dim), nn.Linear(token_dim, cell_dim))
def forward(self, gene_ids, values, padding_mask):
tokens = self.token_norm(
self.gene_embedding(gene_ids) + self.value_encoder(values.unsqueeze(-1))
)
latents = self.latents.unsqueeze(0).expand(gene_ids.shape[0], -1, -1)
attended, _ = self.cross_attention(
self.query_norm(latents),
tokens,
tokens,
key_padding_mask=padding_mask,
need_weights=False,
)
latents = latents + attended
latents = latents + self.cross_ffn(latents)
return self.output(self.blocks(latents).mean(dim=1))
class CellJEPA(nn.Module):
"""Online encoder, frozen EMA teacher, and whole-cell latent predictor."""
def __init__(self, encoder, predictor_hidden=768):
super().__init__()
self.online = encoder
self.teacher = deepcopy(encoder).requires_grad_(False)
cell_dim = encoder.output[-1].out_features
self.predictor = nn.Sequential(
nn.LayerNorm(cell_dim),
nn.Linear(cell_dim, predictor_hidden),
nn.GELU(),
nn.Linear(predictor_hidden, cell_dim),
)
self.teacher.eval()
def train(self, mode=True):
super().train(mode)
self.teacher.eval()
return self
def forward(self, student_view, teacher_view):
context = self.online(**student_view)
with torch.no_grad():
target = self.teacher(**teacher_view).detach()
return self.predictor(context), target, context
@torch.no_grad()
def update_teacher(self, momentum):
for teacher, online in zip(self.teacher.parameters(), self.online.parameters()):
teacher.lerp_(online, 1.0 - momentum)
def load_frozen_teacher(path, vocab_size=3_000, device="cpu"):
"""Load the completed canonical EMA encoder without any trainable parameters."""
payload = torch.load(Path(path), map_location=device, weights_only=False)
config = payload["encoder_configuration"]
assert payload["format_version"] == 1 and payload["frozen"] is True
assert payload["training_state"]["complete"] is True
encoder = CellEncoder(
vocab_size=vocab_size,
token_dim=config["token_dim"],
latent_queries=config["latent_queries"],
blocks=config["blocks"],
heads=config["heads"],
ffn_dim=config["ffn_dim"],
dropout=config["dropout"],
cell_dim=config["cell_dim"],
).to(device)
encoder.load_state_dict(payload["teacher"], strict=True)
return encoder.requires_grad_(False).eval(), payload
def mask_gene_tokens(
gene_ids,
values,
padding_mask,
programs=(),
mask_range=(0.30, 0.60),
biological_probability=0.25,
minimum_program_genes=3,
padding_id=3_000,
generator=None,
):
"""Create a masked student view while leaving the complete teacher view untouched."""
masked_genes, masked_values, masked_padding = (
gene_ids.clone(),
values.clone(),
padding_mask.clone(),
)
for row in range(gene_ids.shape[0]):
visible = torch.nonzero(~padding_mask[row], as_tuple=False).flatten()
selected = None
if programs and torch.rand((), generator=generator) < biological_probability:
program_index = int(torch.randint(len(programs), (), generator=generator))
positions = visible[torch.isin(gene_ids[row, visible], programs[program_index])]
if positions.numel() >= minimum_program_genes:
selected = positions
if selected is None:
ratio = torch.empty(()).uniform_(*mask_range, generator=generator).item()
count = min(visible.numel() - 1, max(1, round(ratio * visible.numel())))
selected = visible[torch.randperm(visible.numel(), generator=generator)[:count]]
selected = selected[: max(0, visible.numel() - 1)]
masked_genes[row, selected] = padding_id
masked_values[row, selected] = 0
masked_padding[row, selected] = True
return {
"gene_ids": masked_genes,
"values": masked_values,
"padding_mask": masked_padding,
}
def jepa_loss(prediction, target, context, variance_weight=0.05, covariance_weight=0.005):
"""Normalized latent prediction plus light VICReg-style anti-collapse terms."""
prediction_loss = F.smooth_l1_loss(F.normalize(prediction), F.normalize(target))
variance_loss = F.relu(1.0 - torch.sqrt(context.var(dim=0, unbiased=False) + 1e-4)).mean()
centered = context - context.mean(dim=0)
covariance = centered.T @ centered / max(context.shape[0] - 1, 1)
covariance_loss = (
covariance[~torch.eye(covariance.shape[0], dtype=bool, device=context.device)].pow(2).mean()
)
total = prediction_loss + variance_weight * variance_loss + covariance_weight * covariance_loss
return {
"loss": total,
"prediction": prediction_loss,
"variance": variance_loss,
"covariance": covariance_loss,
}
def ema_momentum(step, total_steps, start=0.996, end=0.9999):
"""Cosine schedule the teacher momentum over optimizer updates."""
return end - (end - start) * (math.cos(math.pi * step / total_steps) + 1.0) / 2.0
|