hallucination / probe /util.py
ToiTenBao's picture
Upload hallucination folder
a2ffd07 verified
Raw
History Blame Contribute Delete
13.7 kB
from collections.abc import Generator, Sequence
from typing import TypeVar, overload
import torch
from tqdm.autonotebook import tqdm
from transformer_lens.hook_points import HookedRootModule
from torch.utils.data import TensorDataset, DataLoader
from sae_lens import SAE
from typing import Dict, List, Tuple
from sae.SAE_Trainer import DataConfig
from sae.Load_Data import load_lvlm_data
from sae.SAE_Tools import *
from IPython.display import HTML, display
T = TypeVar("T")
K = TypeVar("K")
DEFAULT_DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
@overload
def batchify(
data: Sequence[T], batch_size: int, show_progress: bool = False
) -> Generator[Sequence[T], None, None]: ...
@overload
def batchify(
data: torch.Tensor, batch_size: int, show_progress: bool = False
) -> Generator[torch.Tensor, None, None]: ...
def batchify(
data: Sequence[T] | torch.Tensor, batch_size: int, show_progress: bool = False
) -> Generator[Sequence[T] | torch.Tensor, None, None]:
"""Generate batches from data. If show_progress is True, display a progress bar."""
for i in tqdm(
range(0, len(data), batch_size),
total=(len(data) // batch_size + (len(data) % batch_size != 0)),
disable=not show_progress,
):
yield data[i : i + batch_size]
def flip_dict(d: dict[T, T]) -> dict[T, T]:
"""Flip a dictionary, i.e. {a: b} -> {b: a}"""
return {v: k for k, v in d.items()}
def listify(item: T | list[T]) -> list[T]:
"""Convert an item or list of items to a list."""
if isinstance(item, list):
return item
return [item]
def dict_zip(*dicts: dict[T, K]) -> Generator[tuple[T, tuple[K, ...]], None, None]:
"""Zip together multiple dictionaries, iterating their common keys and a tuple of values."""
if not dicts:
return
keys = set(dicts[0]).intersection(*dicts[1:])
for key in keys:
yield key, tuple(d[key] for d in dicts)
def from_top_k_to_tensor(
d_sae: int,
indices: torch.Tensor,
values: torch.Tensor,
device: torch.device | str = "cpu",
):
if len(indices.shape) == 2:
b, _ = indices.shape
latents = torch.zeros((b, d_sae), device=device, dtype=values.dtype)
elif len(indices.shape) == 3:
b, s, _ = indices.shape
latents = torch.zeros((b, s, d_sae), device=device, dtype=values.dtype)
latents = latents.scatter_(
-1, indices.to(device), values.to(device)
)
return latents
def get_sae_acts(
input_activations: torch.Tensor,
sae: HookedRootModule,
batch_size: int = 4096,
device: torch.device | str = "cpu",
convert_to_cpu: bool = False,
verbose: bool = True,
) -> torch.Tensor | Tuple[torch.Tensor, torch.Tensor]:
indices, values = get_sae_activations(
sae,
DataLoader(TensorDataset(input_activations), batch_size=batch_size, shuffle=False),
device,
no_tqdm=not verbose,
)
return from_top_k_to_tensor(
sae.cfg.d_sae,
indices,
values,
device=device if not convert_to_cpu else "cpu",
)
def load_sae(
model: HookedRootModule, list_release: List[str], list_sae_id: List[str], device: torch.device | str
) -> List[HookedRootModule]:
list_saes = []
for release, sae_id in zip(list_release, list_sae_id):
if release == "local":
list_saes.append(
load_sae_model(
sae_id,
model,
device=str(device),
)
)
else:
sae, _, _ = SAE.from_pretrained(
release=release,
sae_id=sae_id,
device=str(device),
)
list_saes.append(sae)
return list_saes
def load_data_toks(data_length: int, tok_name: str) -> Tensor:
num_workers=4
hf_dataset="yerevann/coco-karpathy"
local_train_path="./COCO-Dataset/train_rest"
local_val_path="./COCO-Dataset/val"
tok_name="Salesforce/blip-image-captioning-base"
batch_size=16
max_length=512
data_config = DataConfig(
batch_size=batch_size,
hf_dataset=hf_dataset,
local_train_path=local_train_path,
local_val_path=local_val_path,
num_workers=num_workers,
max_length=max_length, # the processor of blip only allow max tokens (fixed)
processor = tok_name,
)
_, data_loader = load_lvlm_data(data_config)
data_toks = extract_data(data_loader, num_batches=data_length) # data batch size is 10
return data_toks
def cache_activation_model(
hook_name: str,
model: HookedTransformer,
x: Tensor,
batch_size_model: int,
verbose: bool = True
) -> Tensor:
target_cache = []
def hook_fn(tens: Tensor, hook: HookPoint):
batch, seq = tens.shape[0], tens.shape[1]
target_cache.append(tens.reshape(batch * seq, -1).cpu().detach())
with t.no_grad():
with model.hooks(
fwd_hooks=[
(hook_name, hook_fn)
]
):
for toks in batchify(x, batch_size_model, show_progress=verbose):
model(toks)
acts = t.cat(target_cache)
return acts
@t.inference_mode()
def subsample_tensor(tensor: torch.Tensor, max_samples: int) -> torch.Tensor:
"""
Subsample a 2D tensor if the number of samples exceeds the specified maximum.
Args:
tensor (torch.Tensor): Input 2D tensor of shape (n_sample, f).
max_samples (int): Maximum number of samples to retain.
Returns:
torch.Tensor: Subsampled tensor of shape (min(n_sample, max_samples), f).
"""
n_sample, f = tensor.shape
if n_sample > max_samples:
indices = torch.randperm(n_sample)[:max_samples] # Randomly select max_samples indices
return tensor[indices]
return tensor
@t.no_grad()
def select_feature_from_probe(
probe_weight: Tensor, # (1, d_model)
W_dec: Tensor, # (d_sae, d_model)
sae_acts: Tensor, # (n_sample, d_sae)
labels: Tensor, # (n_sample, 1) --> the binary label
):
mask = t.where(labels, t.ones_like(labels).float(), -t.ones_like(labels).float()).unsqueeze(-1)
positive_label_acts = (sae_acts * mask).mean(0).clamp(min=0) # (d_sae)
positive_label_directions = positive_label_acts.unsqueeze(-1) * W_dec # (d_sae, d_model)
def normalize(tens: Tensor):
return tens / tens.norm(2, dim=1).max()
scores = normalize(positive_label_directions) @ normalize(probe_weight).T # (d_sae, d_model) @ (d_model, 1) -> (d_sae, 1)
return scores
@t.inference_mode()
def compute_f1(
masks: t.Tensor,
indices: t.Tensor,
target_idx: int,
device: t.device,
pad_value: int = -1,
feature_batch_size: int = 64,
target_batch_size: int = 16,
compute_dtype: t.dtype = t.float32,
other_feat_idx: t.Tensor | None = None,
):
"""
Compute maximum F1 scores for a batch of target feature activations vs. other features,
and return both the F1 scores and the indices of the features that achieved them.
Args:
masks: Bool or 0/1 tensor of shape (num_targets, n_sample). masks[i, n] == 1 if target i is active on sample n.
indices: Int tensor of shape (n_sample, k). Each row holds up to k activated feature indices for that sample.
target_idx: The feature index to exclude from the candidates (e.g., the "self" feature).
device: Torch device to run on.
pad_value: Padding value in `indices` rows.
feature_batch_size: Number of candidate features per batch.
target_batch_size: Number of target rows per batch.
compute_dtype: Accumulation dtype (float32 by default).
other_feat_idx: (n_index) If we only compute F1 among a certain feature indices.
Returns:
Tuple of (f1_scores, feature_indices):
- f1_scores: 1D tensor of shape (num_targets,) with the max F1 over all other features for each target.
- feature_indices: 1D tensor of shape (num_targets,) with the index of the feature that achieved the max F1.
"""
masks = masks.to(device=device).bool()
indices = indices.to(device=device)
n_sample = indices.shape[0]
num_targets = masks.shape[0]
# Unique candidate feature ids present in indices (ignoring padding).
if other_feat_idx is None:
valid_mask = indices.ne(pad_value)
if valid_mask.any():
flat_indices = indices[valid_mask]
# Avoid unnecessary sort during unique
unique_indices = t.unique(flat_indices, sorted=False)
# Exclude the single target_idx from candidate pool
other_feat_idx = unique_indices[unique_indices.ne(t.as_tensor(target_idx, device=unique_indices.device))]
else:
other_feat_idx = t.empty(0, dtype=indices.dtype, device=indices.device)
if other_feat_idx.numel() == 0:
zeros = t.zeros(num_targets, dtype=compute_dtype, device=device)
neg_ones = t.full((num_targets,), -1, dtype=indices.dtype, device=device)
return zeros, neg_ones
# Batch over targets
num_target_batches = (num_targets + target_batch_size - 1) // target_batch_size
all_max_f1_scores = []
all_best_indices = []
for target_batch_idx in range(num_target_batches):
st = target_batch_idx * target_batch_size
en = min(st + target_batch_size, num_targets)
# T: (Tb, N) boolean activation for this batch of targets
T = masks[st:en] # bool
Tb = T.shape[0]
# Precompute |target| per row (A): (Tb,)
A = T.sum(dim=1, dtype=compute_dtype) # float
# Track best F1 and corresponding feature index across all feature batches for each target in this batch
best_f1 = t.zeros(Tb, dtype=compute_dtype, device=device)
best_indices = t.full((Tb,), -1, dtype=other_feat_idx.dtype, device=device)
# Batch over candidate features
num_feature_batches = (other_feat_idx.numel() + feature_batch_size - 1) // feature_batch_size
for j in range(num_feature_batches):
fs = j * feature_batch_size
fe = min(fs + feature_batch_size, other_feat_idx.numel())
feature_batch = other_feat_idx[fs:fe] # (Fb,)
# Build F: (N, Fb) boolean activation of these features across samples
# indices: (N, k), feature_batch: (Fb,)
# Equality broadcasting => (N, k, Fb) then any over k -> (N, Fb)
F = (indices.unsqueeze(-1) == feature_batch.view(1, 1, -1)).any(dim=1)
# |feature| per column (B): (Fb,)
B = F.sum(dim=0, dtype=compute_dtype)
# TP = T @ F (both 0/1) => (Tb, Fb)
TP = t.matmul(T.to(dtype=compute_dtype), F.to(dtype=compute_dtype))
# F1 = 2*TP / (|target| + |feature|)
denom = A.unsqueeze(1) + B.unsqueeze(0) # (Tb, Fb)
f1 = t.where(denom > 0, (2.0 * TP) / denom, t.zeros((), dtype=compute_dtype, device=device))
# For each target, find the best feature in this batch
batch_best_f1, batch_best_idx = f1.max(dim=1)
batch_best_idx = batch_best_idx + fs # Convert to global index in other_feat_idx
# Update if this batch has better F1 scores
update_mask = batch_best_f1 > best_f1
best_f1 = t.where(update_mask, batch_best_f1, best_f1)
best_indices = t.where(
update_mask,
other_feat_idx[batch_best_idx], # Get the actual feature index
best_indices
)
all_max_f1_scores.append(best_f1)
all_best_indices.append(best_indices)
return t.cat(all_max_f1_scores, dim=0), t.cat(all_best_indices, dim=0)
def extract_context(data_tensor: Tensor, index_tensor: Tensor, context_size=15):
"""
Extract context windows of ±context_size around each index.
Args:
data_tensor: 2D tensor of shape (batch, seq_len)
index_tensor: 1D tensor of shape (batch,) containing indices
context_size: Size of context window on each side (default: 15)
Returns:
2D tensor of shape (batch, 2 * context_size + 1) with context windows
"""
batch_size, seq_len = data_tensor.shape
window_size = 2 * context_size + 1
# Create relative indices for the context window
indices = t.arange(-context_size, context_size + 1, device=data_tensor.device)
indices = indices.view(1, -1).expand(index_tensor.shape[0], window_size)
# Add the center indices
indices = indices + index_tensor
# Handle boundary conditions by clamping
indices = t.clamp(indices, 0, data_tensor.flatten().shape[0]-1)
context_windows = data_tensor.flatten()[indices]
return context_windows
def highlight_html(strings: List[str], highlight_index: int):
"""
For Jupyter notebooks - uses HTML formatting
"""
html_str = ""
for i, s in enumerate(strings):
s = s.replace("�", "").replace("\n", "↵")
if i == highlight_index:
html_str += f'<span style="color: red; font-weight: bold">{s}</span>'
else:
html_str += f'{s}'
display(HTML(html_str))
def show_activation(
model: HookedTransformer,
data_toks: Tensor, # (b, s)
index_tensor: Tensor, # (n_index)
num_examples: int = 50,
context_size: int = 15,
):
contexts = extract_context(data_toks, index_tensor, context_size=context_size)
for context in contexts[:num_examples]:
highlight_html(model.to_str_tokens(context), context_size)