File size: 9,839 Bytes
a2ffd07 | 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 | import os
import random
from dataclasses import dataclass
from math import exp, log
from pathlib import Path
from typing import Callable, Literal
import numpy as np
import pandas as pd
import torch
from torch import nn, optim
from jaxtyping import Int, Float
from torch.utils.data import DataLoader, TensorDataset
# Autocheck if the instance is a notebook or not (fixes weird bugs in colab)
from tqdm.autonotebook import tqdm
from .util import DEFAULT_DEVICE
class LinearProbe(nn.Module):
"""
Based on by https://github.com/jbloomAus/alphabetical_probe/blob/main/src/probes.py
"""
def __init__(self, input_dim, num_outputs: int = 1):
super().__init__()
self.fc = nn.Linear(input_dim, num_outputs)
def forward(self, x):
return self.fc(x)
@property
def weights(self):
return self.fc.weight
@property
def biases(self):
return self.fc.bias
def _calc_pos_weights(y: torch.Tensor) -> torch.Tensor:
num_pos_samples = y.sum(dim=0)
num_neg_samples = len(y) - num_pos_samples
return num_neg_samples / num_pos_samples
def train_multi_probe(
x_train: torch.Tensor, # tensor of shape (num_samples, input_dim)
y_train: torch.Tensor, # tensor of shape (num_samples, num_probes), with values in [0, 1]
num_probes: int | None = None, # inferred from y_train if None
batch_size: int = 4096,
num_epochs: int = 100,
lr: float = 0.01,
end_lr: float = 1e-5,
weight_decay: float = 1e-6,
show_progress: bool = True,
optimizer: Literal["Adam", "SGD", "AdamW"] = "Adam",
extra_loss_fn: (
Callable[[LinearProbe, torch.Tensor, torch.Tensor], torch.Tensor] | None
) = None,
verbose: bool = False,
device: torch.device = DEFAULT_DEVICE,
map_acts: Callable[[torch.Tensor], torch.Tensor] | None = None,
probe_dim: int | None = None,
) -> LinearProbe:
"""
Train a multi-class one-vs-rest logistic regression probe on the given data.
This is equivalent to training num_probes separate binary logistic regression probes.
Args:
x_train: tensor of shape (num_samples, input_dim)
y_train: one_hot (or multi-hot) tensor of shape (num_samples, num_probes), with values in [0, 1]
num_probes: number of probes to train simultaneously
batch_size: batch size for training
num_epochs: number of epochs to train for
lr: learning rate
weight_decay: weight decay
show_progress: whether to show a progress bar
device: device to train on
"""
dtype = x_train.dtype
num_probes = num_probes or y_train.shape[-1]
dataset = TensorDataset(x_train, y_train.to(dtype=dtype))
loader = DataLoader(dataset, batch_size=batch_size, shuffle=True)
if probe_dim is None:
probe_dim = x_train.shape[-1]
probe = LinearProbe(probe_dim, num_outputs=num_probes).to(device, dtype=dtype)
_run_probe_training(
probe,
loader,
loss_fn=nn.BCEWithLogitsLoss(pos_weight=_calc_pos_weights(y_train).to(device)),
num_epochs=num_epochs,
lr=lr,
end_lr=end_lr,
weight_decay=weight_decay,
show_progress=show_progress,
optimizer_name=optimizer,
extra_loss_fn=extra_loss_fn,
verbose=verbose,
device=device,
map_acts=map_acts,
)
return probe
def train_binary_probe(
x_train: torch.Tensor, # tensor of shape (num_samples, input_dim)
y_train: torch.Tensor, # tensor of shape (num_samples,), with values in [0, 1]
batch_size: int = 256,
num_epochs: int = 100,
lr: float = 0.01,
end_lr: float = 1e-5,
weight_decay: float = 1e-6,
show_progress: bool = True,
optimizer: Literal["Adam", "SGD", "AdamW"] = "Adam",
extra_loss_fn: (
Callable[[LinearProbe, torch.Tensor, torch.Tensor], torch.Tensor] | None
) = None,
verbose: bool = False,
device: torch.device = DEFAULT_DEVICE,
) -> LinearProbe:
"""
Train a logistic regression probe on the given data. This is a thin wrapped around train_multi_probe.
Args:
x_train: tensor of shape (num_samples, input_dim)
y_train: tensor of shape (num_samples,), with values in [0, 1]
batch_size: batch size for training
num_epochs: number of epochs to train for
lr: learning rate
weight_decay: weight decay
show_progress: whether to show a progress bar
device: device to train on
"""
return train_multi_probe(
x_train,
y_train.unsqueeze(1),
num_probes=1,
batch_size=batch_size,
num_epochs=num_epochs,
lr=lr,
end_lr=end_lr,
weight_decay=weight_decay,
show_progress=show_progress,
optimizer=optimizer,
extra_loss_fn=extra_loss_fn,
verbose=verbose,
device=device,
)
def _get_exponential_decay_scheduler(
optimizer: optim.Optimizer, # type: ignore
start_lr: float,
end_lr: float,
num_steps: int,
) -> optim.lr_scheduler.ExponentialLR:
gamma = exp(log(end_lr / start_lr) / num_steps)
return optim.lr_scheduler.ExponentialLR(optimizer, gamma=gamma)
def _run_probe_training(
probe: LinearProbe,
loader: DataLoader,
loss_fn: Callable[[torch.Tensor, torch.Tensor], torch.Tensor],
num_epochs: int,
lr: float,
end_lr: float,
weight_decay: float,
show_progress: bool,
optimizer_name: Literal["Adam", "SGD", "AdamW"],
extra_loss_fn: (
Callable[[LinearProbe, torch.Tensor, torch.Tensor], torch.Tensor] | None
),
verbose: bool,
device: torch.device,
map_acts: Callable[[torch.Tensor], torch.Tensor] | None = None,
) -> None:
probe.train()
if optimizer_name == "Adam":
optimizer = optim.Adam(probe.parameters(), lr=lr, weight_decay=weight_decay) # type: ignore
elif optimizer_name == "SGD":
optimizer = optim.SGD(probe.parameters(), lr=lr, weight_decay=weight_decay) # type: ignore
elif optimizer_name == "AdamW":
optimizer = optim.AdamW(probe.parameters(), lr=lr, weight_decay=weight_decay) # type: ignore
else:
raise ValueError(f"Unknown optimizer: {optimizer_name}")
scheduler = _get_exponential_decay_scheduler(
optimizer, start_lr=lr, end_lr=end_lr, num_steps=num_epochs
)
epoch_pbar = tqdm(range(num_epochs), disable=not show_progress, desc="Epochs")
for epoch in epoch_pbar:
epoch_sum_loss = 0
batch_pbar = tqdm(
loader,
disable=not show_progress,
leave=False,
desc=f"Epoch {epoch + 1}/{num_epochs}",
)
for batch_embeddings, batch_labels in batch_pbar:
batch_embeddings = batch_embeddings.to(device)
if map_acts is not None:
batch_embeddings = map_acts(batch_embeddings)
batch_labels = batch_labels.to(device)
optimizer.zero_grad()
logits = probe(batch_embeddings)
loss = loss_fn(logits, batch_labels)
if extra_loss_fn is not None:
loss += extra_loss_fn(probe, batch_embeddings, batch_labels)
loss.backward()
optimizer.step()
batch_loss = loss.item()
epoch_sum_loss += batch_loss
batch_pbar.set_postfix({"Loss": f"{batch_loss:.8f}"})
epoch_mean_loss = epoch_sum_loss / len(loader)
current_lr = scheduler.get_last_lr()[0]
epoch_pbar.set_postfix(
{"Mean Loss": f"{epoch_mean_loss:.8f}", "LR": f"{current_lr:.2e}"}
)
if verbose:
print(
f"Epoch {epoch + 1}: Mean Loss: {epoch_mean_loss:.8f}, LR: {current_lr:.2e}"
)
scheduler.step()
probe.eval()
def select_k_features(
l1_probe: LinearProbe,
k: int,
label: int,
acts: torch.Tensor | None = None, # (n_sample, n_feats)
) -> Int[torch.Tensor, "k"]:
if acts is None:
return l1_probe.weights[label].topk(k).indices
else:
return (torch.sum(acts, dim=0) * l1_probe.weights[label].cpu()).topk(k).indices
def load_probe(
save_name: str,
d_model: int,
num_outputs: int,
device: torch.device | str = DEFAULT_DEVICE,
) -> LinearProbe:
probe = LinearProbe(d_model, num_outputs)
probe.load_state_dict(torch.load(save_name, map_location=device))
probe.to(device)
return probe
def train_or_load_lr(
x_train: torch.Tensor, # tensor of shape (num_samples, input_dim)
y_train: torch.Tensor, # tensor of shape (num_samples, num_probes), with values in [0, 1]
save_name: str,
batch_size_lr: int = 4096,
num_epochs: int = 100,
lr: float = 0.01,
end_lr: float = 1e-5,
weight_decay: float = 1e-6,
show_progress: bool = True,
optimizer: Literal["Adam", "SGD", "AdamW"] = "Adam",
verbose: bool = False,
device: torch.device | str = DEFAULT_DEVICE,
map_acts: Callable[[torch.Tensor], torch.Tensor] | None = None,
probe_dim: int | None = None,
) -> LinearProbe:
if not os.path.exists(save_name):
probe = train_multi_probe(
x_train,
y_train,
batch_size=batch_size_lr,
num_epochs=num_epochs,
lr=lr,
end_lr=end_lr,
weight_decay=weight_decay,
show_progress=show_progress,
optimizer=optimizer,
verbose=verbose,
device=device,
map_acts=map_acts,
probe_dim=probe_dim
).to("cpu")
if verbose:
print(f"Probe trained with {probe.num_probes} probes")
torch.save(probe.state_dict(), save_name)
else:
probe = load_probe(
save_name, x_train.shape[-1], y_train.shape[-1], "cpu",
)
return probe |