File size: 9,565 Bytes
d6682b6 |
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 |
from typing import Dict, List, Tuple
import numpy as np
import torch
from matplotlib.style import context
from transformers import AutoModelForCausalLM, AutoTokenizer
from ..rome import repr_tools
from ...util import nethook
from .rome_hparams import ROMEHyperParams
def compute_v(
model: AutoModelForCausalLM,
tok: AutoTokenizer,
request: Dict,
hparams: ROMEHyperParams,
layer: int,
left_vector: torch.Tensor,
context_templates: List[str],
) -> torch.Tensor:
"""
Computes the value (right) vector for the rank-1 update.
Runs a simple optimization procedure.
"""
print("Computing right vector (v)")
# Tokenize target into list of int token IDs
target_ids = tok.encode(request["target_new"], return_tensors="pt", add_special_tokens=False).to('cpu')[0]
# if target_ids[0] == tok.bos_token_id or target_ids[0] == tok.unk_token_id:
# target_ids = target_ids[1:]
# Compile list of rewriting and KL x/y pairs
rewriting_prompts, kl_prompts = [
context.format(request["prompt"]) + tok.decode(target_ids[:-1])
for context in context_templates
], ["{} is a"]
all_prompts = rewriting_prompts + kl_prompts
input_tok = tok(
[prompt.format(request["subject"]) for prompt in all_prompts],
return_tensors="pt",
padding=True,
).to("cpu")
# Compute rewriting targets
rewriting_targets = torch.tensor(-100, device='cpu').repeat(
len(rewriting_prompts), *input_tok["input_ids"].shape[1:]
)
for i in range(len(rewriting_prompts)):
ex_len = input_tok["attention_mask"][i].sum()
rewriting_targets[i, ex_len - len(target_ids) : ex_len] = target_ids
# Compute indices of the tokens where the fact is looked up
vanilla_input_prompts = [
context.format(request["prompt"]).format(request['subject'])
for context in context_templates
] + [f"{request['subject']} is a"]
lookup_idxs = [
find_fact_lookup_idx(
prompt, request["subject"], tok, hparams.fact_token, verbose=(i == 0), input_prompt=vanilla_input_prompts[i]
)
for i, prompt in enumerate(all_prompts)
]
# Finalize rewrite and loss layers
loss_layer = max(hparams.v_loss_layer, layer)
print(f"Rewrite layer is {layer}")
print(f"Tying optimization objective to {loss_layer}")
# Set up an optimization over a latent vector that, when output at the
# rewrite layer, i.e. hypothesized fact lookup location, will induce the
# target token to be predicted at the final layer.
if hasattr(model.config, 'n_embd'):
delta = torch.zeros((model.config.n_embd,), requires_grad=True, device=f"cpu")
else:
delta = torch.zeros((model.config.hidden_size,), requires_grad=True, device=f"cpu")
target_init, kl_distr_init = None, None
# Inserts new "delta" variable at the appropriate part of the computation
def edit_output_fn(cur_out, cur_layer):
nonlocal target_init
if cur_layer == hparams.mlp_module_tmp.format(layer):
# Store initial value of the vector of interest
if target_init is None:
print("Recording initial value of v*")
# Initial value is recorded for the clean sentence
target_init = cur_out[0, lookup_idxs[0]].detach().clone()
for i, idx in enumerate(lookup_idxs):
if len(lookup_idxs)!=len(cur_out):
cur_out[idx, i, :] += delta
else:
cur_out[i, idx, :] += delta
return cur_out
# Optimizer
opt = torch.optim.Adam([delta], lr=hparams.v_lr)
nethook.set_requires_grad(False, model)
# Execute optimization
for it in range(hparams.v_num_grad_steps):
opt.zero_grad()
# Forward propagation
with nethook.TraceDict(
module=model,
layers=[
hparams.layer_module_tmp.format(loss_layer),
hparams.mlp_module_tmp.format(layer),
],
retain_input=False,
retain_output=True,
edit_output=edit_output_fn,
) as tr:
logits = model(**input_tok).logits
# Compute distribution for KL divergence
kl_logits = torch.stack(
[
logits[i - len(kl_prompts), idx, :]
for i, idx in enumerate(lookup_idxs[-len(kl_prompts) :])
],
dim=0,
)
kl_log_probs = torch.nn.functional.log_softmax(kl_logits, dim=1)
if kl_distr_init is None:
kl_distr_init = kl_log_probs.detach().clone()
# Compute loss on rewriting targets
log_probs = torch.log_softmax(logits, dim=2)
loss = torch.gather(
log_probs,
2,
torch.where(rewriting_targets != -100, rewriting_targets, 0).unsqueeze(2),
).squeeze(2)
mask = (rewriting_targets != -100).float()
# Aggregate total losses
nll_loss_each = -(loss * mask).sum(1) / target_ids.size(0)
nll_loss = nll_loss_each.mean()
kl_loss = hparams.kl_factor * torch.nn.functional.kl_div(
kl_distr_init, kl_log_probs, log_target=True, reduction="batchmean"
)
weight_decay = hparams.v_weight_decay * (
torch.norm(delta) / torch.norm(target_init) ** 2
)
# weight_decay = hparams.v_weight_decay * torch.norm(delta) ** 2
loss = nll_loss + kl_loss + weight_decay
print(
f"loss {np.round(loss.item(), 3)} = {np.round(nll_loss.item(), 3)} + {np.round(kl_loss.item(), 3)} + {np.round(weight_decay.item(), 3)} "
f"avg prob of [{request['target_new']}] "
f"{torch.exp(-nll_loss_each).mean().item()}"
)
if loss < 5e-2:
break
if it == hparams.v_num_grad_steps - 1:
break
# Backpropagate
loss.backward()
opt.step()
# Project within L2 ball
max_norm = hparams.clamp_norm_factor * target_init.norm()
if delta.norm() > max_norm:
with torch.no_grad():
delta[...] = delta * max_norm / delta.norm()
target = target_init + delta.to(target_init.dtype)
# Retrieve cur_input, the current input to the 2nd MLP layer, and
# cur_output, the original output of the 2nd MLP layer.
cur_input, cur_output = get_module_input_output_at_word(
model,
tok,
layer,
context_template=request["prompt"],
word=request["subject"],
module_template=hparams.rewrite_module_tmp,
fact_token_strategy=hparams.fact_token,
)
# Solving the linear system to compute the right vector
right_vector = (target - cur_output) / torch.dot(cur_input, left_vector)
print(f"Delta norm: {(target - cur_output).norm().item()}")
print(
f"Change in target norm: {target_init.norm().item()} to {target.norm().item()} => {(target.norm() - target_init.norm()).item()}"
)
print(f"Division Factor: {torch.dot(cur_input, left_vector).item()}")
print(f"Right vector norm: {right_vector.norm()}")
return right_vector
def get_module_input_output_at_word(
model: AutoModelForCausalLM,
tok: AutoTokenizer,
layer: int,
context_template: str,
word: str,
module_template: str,
fact_token_strategy: str,
) -> Tuple[torch.Tensor]:
"""
Retrieves detached representations for a word at the input and
output of a particular layer module.
"""
word_repr_args = dict(
model=model,
tok=tok,
layer=layer,
module_template=module_template,
)
if "subject_" in fact_token_strategy and fact_token_strategy.index("subject_") == 0:
subtoken = fact_token_strategy[len("subject_") :]
l_input, l_output = repr_tools.get_reprs_at_word_tokens(
track="both",
subtoken=subtoken,
context_templates=[context_template],
words=[word],
**word_repr_args,
)
elif fact_token_strategy == "last":
l_input, l_output = repr_tools.get_reprs_at_idxs(
track="both",
contexts=[context_template.format(word)],
idxs=[[-1]],
**word_repr_args,
)
else:
raise ValueError(f"fact_token={fact_token_strategy} not recognized")
l_input, l_output = l_input[0], l_output[0]
return l_input.detach(), l_output.detach()
def find_fact_lookup_idx(
prompt: str,
subject: str,
tok: AutoTokenizer,
fact_token_strategy: str,
verbose=True,
input_prompt=None
) -> int:
"""
Computes hypothesized fact lookup index given a sentence and subject.
"""
ret = None
if fact_token_strategy == "last":
ret = len(tok.encode(input_prompt)) - 1
elif (
"subject_" in fact_token_strategy and fact_token_strategy.index("subject_") == 0
):
ret = repr_tools.get_words_idxs_in_templates(
tok=tok,
context_templates=[prompt],
words=[subject],
subtoken=fact_token_strategy[len("subject_") :],
)[0][0]
else:
raise ValueError(f"fact_token={fact_token_strategy} not recognized")
sentence = prompt.format(subject)
if verbose:
print(
f"Lookup index found: {ret} | Sentence: {sentence} | Token:",
tok.decode(tok(sentence)["input_ids"][ret]),
)
return ret
|