File size: 10,338 Bytes
bbc7ddf |
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 |
import torch
import torch.nn as nn
import torch.nn.functional as F
from dataclasses import dataclass
from typing import Dict, List, Optional, Tuple
from transformers import PreTrainedModel
from transformers.utils import ModelOutput
from .configuration_roberta_zinc_compression_encoder import RZCompressionConfig
# pairwise cosine ----------------------------------------------------------------
def pairwise_cosine(x: torch.Tensor) -> torch.Tensor:
x = F.normalize(x, p=2, dim=-1)
return x @ x.t() # [B, B]
# remove diagonal -----------------------------------------------------------------
def drop_diag(M: torch.Tensor) -> torch.Tensor:
n = M.size(0)
return M.masked_select(~torch.eye(n, dtype=torch.bool, device=M.device)).view(n, n - 1)
# pearson row-wise ----------------------------------------------------------------
def rowwise_pearson(ref: torch.Tensor, comp: torch.Tensor, rm_diag: bool=True) -> torch.Tensor:
if rm_diag:
ref = drop_diag(ref)
comp = drop_diag(comp)
ref_z = F.normalize(ref - ref.mean(dim=1, keepdim=True), p=2, dim=1)
cmp_z = F.normalize(comp - comp.mean(dim=1, keepdim=True), p=2, dim=1)
return 1 - (ref_z * cmp_z).sum(dim=1).mean() # 0 = perfect corr
# aggregate loss ------------------------------------------------------------------
def compute_losses(
embedding: torch.Tensor, # (batch_size, d)
compressed: Dict[int, torch.Tensor], # Dict[size, (batch_size, size)]
recon_stack: torch.Tensor | None, # (batch_size, n_heads, d)
cfg,
) -> tuple[torch.Tensor, dict[str, float]]:
"""Return (total_loss, terms_dict)"""
device = embedding.device
loss_total = torch.zeros((), device=device)
terms: dict[str, float] = {}
# ---- base similarities (detach to save mem) ---------------------------
with torch.no_grad():
base_sims = pairwise_cosine(embedding)
ranks = base_sims.argsort(-1, descending=True)
# ======================================================================
# 1) encoder / compressed losses
# ======================================================================
for size, z in compressed.items():
tag = f"cmp{size}"
comp_sims = pairwise_cosine(z)
# plain MSE --------------------------------------------------------
if cfg.mse_loss_weight:
mse = F.mse_loss(drop_diag(base_sims), drop_diag(comp_sims))
loss_total += cfg.mse_loss_weight * mse
terms[f"{tag}_mse"] = mse.detach()
# top-k MSE --------------------------------------------------------
if cfg.mse_loss_weight and cfg.topk_values:
tk_vals = []
for k in cfg.topk_values:
idx = ranks[:, 1 : k + 1]
ref_k = torch.gather(base_sims, 1, idx)
cmp_k = torch.gather(comp_sims, 1, idx)
tk_mse = F.mse_loss(ref_k, cmp_k)
tk_vals.append(tk_mse)
terms[f"{tag}_top{k}"] = tk_mse.detach()
tk_agg = torch.stack(tk_vals).mean()
loss_total += cfg.topk_mse_loss_weight * tk_agg
terms[f"{tag}_topk_mean"] = tk_agg.detach()
# Pearson ----------------------------------------------------------
if cfg.pearson_loss_weight:
pr = rowwise_pearson(base_sims, comp_sims)
loss_total += cfg.pearson_loss_weight * pr
terms[f"{tag}_pearson"] = pr.detach()
if cfg.pearson_loss_weight and cfg.topk_values:
pr_vals = []
for k in cfg.topk_values:
idx = ranks[:, 1 : k + 1]
ref_k = torch.gather(base_sims, 1, idx)
cmp_k = torch.gather(comp_sims, 1, idx)
pr = rowwise_pearson(ref_k, cmp_k, rm_diag=False)
pr_vals.append(pr)
terms[f"{tag}_pearson_top{k}"] = pr.detach()
pr_agg = torch.stack(pr_vals).sum()
loss_total += cfg.pearson_loss_weight * pr_agg
# ======================================================================
# 2) decoder losses
# ======================================================================
if recon_stack is not None:
# cosine -----------------------------------------------------------
if cfg.decoder_cosine_weight:
cos_loss = 1 - F.cosine_similarity(
recon_stack,
embedding.unsqueeze(1).expand_as(recon_stack),
dim=-1,
).mean()
loss_total += cfg.decoder_cosine_weight * cos_loss
terms["dec_cosine"] = cos_loss.detach()
return loss_total, terms
# ─── basic blocks ───────────────────────────────────────────────
class FeedForward(nn.Module):
def __init__(self, d_in: int, d_out: int):
super().__init__()
self.fc1 = nn.Linear(d_in, d_out * 2)
self.fc2 = nn.Linear(d_out, d_out)
def forward(self, x):
x = self.fc1(x)
x1, x2 = x.chunk(2, dim=-1)
return self.fc2(F.silu(x1) * x2)
class FeedForwardLayer(nn.Module):
def __init__(
self, d_in: int, d_out: int, dropout: float = 0.1, layer_norm_eps: Optional[float] = 1e-12
):
super().__init__()
self.ff = FeedForward(d_in, d_out)
self.skip = nn.Linear(d_in, d_out) if d_in != d_out else nn.Identity()
self.dropout = nn.Dropout(dropout)
self.norm = (
nn.LayerNorm(d_out, eps=layer_norm_eps)
if layer_norm_eps is not None else nn.Identity()
)
def forward(self, x):
y = self.ff(self.dropout(x)) + self.skip(x)
return self.norm(y)
# ─── pure PyTorch compressor ────────────────────────────────────
class CompressionModel(nn.Module):
"""
Encoder → (optional) Decoder.
"""
def __init__(
self,
d_in: int,
d_comp: int,
encoder_layers: int,
decoder_layers: int,
dropout: float,
layer_norm_eps: Optional[float],
):
super().__init__()
enc_layers: List[nn.Module] = []
for i in range(encoder_layers):
last = i == encoder_layers - 1
enc_layers.append(
FeedForwardLayer(
d_in,
d_comp if last else d_in,
dropout if not last else 0.0,
None if last else layer_norm_eps,
)
)
self.encoder = nn.Sequential(*enc_layers)
# optional decoder
dec_layers: List[nn.Module] = []
for i in range(decoder_layers):
last = i == decoder_layers - 1
d_prev = d_comp if i==0 else d_in
dec_layers.append(
FeedForwardLayer(
d_prev,
d_in,
dropout if not last else 0.0,
None if last else layer_norm_eps,
)
)
self.decoder = nn.Sequential(*dec_layers) if dec_layers else None
def forward(self, x) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
z = self.encoder(x)
x_recon = self.decoder(z) if self.decoder is not None else None
return z, x_recon
# ─── HF wrapper ─────────────────────────────────────────────────
@dataclass
class RZCompressionOutput(ModelOutput):
loss: torch.FloatTensor
loss_terms: Dict[str, torch.Tensor] | None = None
compressed: Dict[int, torch.FloatTensor] | None = None
reconstructed: torch.FloatTensor | None = None
class RZCompressionModel(PreTrainedModel):
config_class = RZCompressionConfig
def __init__(self, config: RZCompressionConfig):
super().__init__(config)
self.compressors = nn.ModuleDict(
{
str(size): CompressionModel(
d_in=config.input_size,
d_comp=size,
encoder_layers=config.encoder_layers,
decoder_layers=config.decoder_layers,
dropout=config.dropout,
layer_norm_eps=config.layer_norm_eps,
)
for size in config.compression_sizes
}
)
self.post_init()
def get_encoders(self, unpack_single=False):
encoders = {}
for k,v in self.compressors.items():
v = v.encoder
if len(v)==1 and unpack_single:
# unpack from nn.Sequential if only a single layer
v = v[0]
encoders[k] = v
encoders = nn.ModuleDict(encoders)
return encoders
def save_encoders(self, path, unpack_single=False):
encoders = self.get_encoders(unpack_single)
torch.save(encoders.state_dict(), path)
def compress(self,
inputs: torch.Tensor,
compression_sizes: List[int]):
compressed = {d: self.compressors[str(d)].encoder(inputs) for d in compression_sizes}
return compressed
def forward(self, embedding, return_dict=True, compute_loss=True):
# ---------- forward passes ------------------------------------------------
compressed, recons = {}, []
for size, module in self.compressors.items():
z, rec = module(embedding)
compressed[int(size)] = z
if rec is not None:
recons.append(rec)
recon_stack = torch.stack(recons, dim=1) if recons else None
# ---------- losses --------------------------------------------------------
if compute_loss:
loss_total, terms = compute_losses(embedding, compressed, recon_stack, self.config)
else:
loss_total, terms = torch.zeros((), device=embedding.device), {}
if not return_dict:
return compressed, recon_stack, loss_total, terms
return RZCompressionOutput(
loss=loss_total,
loss_terms=terms,
compressed=compressed,
reconstructed=recon_stack,
) |