File size: 9,311 Bytes
93cd35a | 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 | import math
import statistics
from typing import Dict, List, Union
import numpy as np
import torch
import torch.nn.functional as F
import sacrebleu
import transformers
from tqdm import tqdm
from utils.logging_utils import log_for_0
# ============================================
# Text-similarity metrics (BLEU / ROUGE)
# ============================================
def _mean_std_sem(values):
n = len(values)
mean = sum(values) / n
std = statistics.pstdev(values) if n > 1 else 0.0
sem = std / math.sqrt(n) if n > 1 else 0.0
return mean, std, sem
def compute_bleu(hypotheses, references):
return sacrebleu.corpus_bleu(hypotheses, [references], lowercase=True, use_effective_order=True).score
def compute_rouge(hypotheses, references, return_std=False):
from rouge_score import rouge_scorer
scorer = rouge_scorer.RougeScorer(["rouge1", "rouge2", "rougeL"], use_stemmer=True)
r1, r2, rL = [], [], []
for hyp, ref in zip(hypotheses, references):
s = scorer.score(ref, hyp)
r1.append(s["rouge1"].fmeasure * 100)
r2.append(s["rouge2"].fmeasure * 100)
rL.append(s["rougeL"].fmeasure * 100)
m1, s1, e1 = _mean_std_sem(r1)
m2, s2, e2 = _mean_std_sem(r2)
mL, sL, eL = _mean_std_sem(rL)
means = {"rouge1": m1, "rouge2": m2, "rougeL": mL}
if not return_std:
return means
stds = {
"rouge1_std": s1, "rouge2_std": s2, "rougeL_std": sL,
"rouge1_sem": e1, "rouge2_sem": e2, "rougeL_sem": eL,
}
return means, stds
# ============================================
# Perplexity / entropy metrics (PyTorch)
# ============================================
class NLL:
"""PyTorch implementation of NLL metric."""
def __init__(self):
self.reset()
def reset(self):
self.mean_value = torch.tensor(0.0, dtype=torch.float64)
self.weight = torch.tensor(0.0, dtype=torch.float64)
def update(self, value: Union[float, torch.Tensor], weight: Union[float, torch.Tensor] = 1.0):
if not isinstance(value, torch.Tensor):
value = torch.tensor(value, dtype=torch.float64)
if not isinstance(weight, torch.Tensor):
weight = torch.tensor(weight, dtype=torch.float64)
weight = torch.broadcast_to(weight, value.shape)
if value.numel() == 0:
return
self.mean_value = self.mean_value + value.sum()
self.weight = self.weight + weight.sum()
class Perplexity(NLL):
def compute(self) -> torch.Tensor:
return torch.exp(self.mean_value / self.weight)
class MeanMetric:
def __init__(self):
self.reset()
def reset(self):
self.sum_value = torch.tensor(0.0, dtype=torch.float64)
self.count = torch.tensor(0.0, dtype=torch.float64)
def update(self, value: Union[float, torch.Tensor]):
if not isinstance(value, torch.Tensor):
value = torch.tensor(value, dtype=torch.float64)
self.sum_value = self.sum_value + value.sum()
self.count = self.count + value.numel()
def compute(self) -> torch.Tensor:
return self.sum_value / self.count
class Metrics:
def __init__(
self,
gen_ppl_eval_model_name_or_path=None,
eval_ppl_batch_size=None,
eval_context_size=1024,
) -> None:
self.gen_ppl = Perplexity()
self.sample_entropy = MeanMetric()
self.eval_ppl_batch_size = eval_ppl_batch_size
self.gen_ppl_eval_model_name_or_path = gen_ppl_eval_model_name_or_path
self.eval_context_size = eval_context_size
self._eval_model = None
self._eval_device = None
# mT5 needs use_fast=False to avoid Tiktoken/SentencePiece conversion issues.
use_fast = "mt5" not in gen_ppl_eval_model_name_or_path.lower()
self.tokenizer = transformers.AutoTokenizer.from_pretrained(
gen_ppl_eval_model_name_or_path, use_fast=use_fast,
)
if self.tokenizer.pad_token is None:
self.tokenizer.pad_token = self.tokenizer.eos_token
self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
def reset(self):
self.gen_ppl.reset()
self.sample_entropy.reset()
def _eval_retokenize(self, text_samples, max_length):
out = self.tokenizer(
text_samples,
return_tensors="np",
return_token_type_ids=False,
return_attention_mask=True,
truncation=True,
padding=True,
max_length=max_length,
)
return out["input_ids"], out["attention_mask"], self.eval_context_size
@torch.no_grad()
def _compute_batch_nlls(self, input_ids: torch.Tensor, attention_mask: torch.Tensor):
model = self._eval_model
eos_token_id = self.tokenizer.eos_token_id
logits = model(input_ids=input_ids, attention_mask=attention_mask).logits
targets = input_ids[:, 1:]
logits_pred = logits[:, :-1, :]
log_normalizers = torch.logsumexp(logits_pred.to(torch.float32), dim=-1)
target_logits = logits_pred.gather(-1, targets.unsqueeze(-1)).squeeze(-1).to(torch.float32)
nlls = log_normalizers - target_logits
is_eos = (input_ids == eos_token_id)
first_eos = (is_eos.to(torch.int32).cumsum(dim=-1) == 1)
token_mask = (input_ids != eos_token_id)
valid_tokens = first_eos[:, 1:].to(torch.int32) + token_mask[:, 1:].to(torch.int32)
return nlls, valid_tokens
def record_generative_perplexity(
self,
text_samples: List[str],
max_length: int,
retokenize: bool = True,
) -> Dict:
import os
os.environ["TOKENIZERS_PARALLELISM"] = "false"
if self._eval_model is None:
from transformers import AutoModelForCausalLM
log_for_0(f"Loading PyTorch model: {self.gen_ppl_eval_model_name_or_path}")
self._eval_device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self._eval_model = AutoModelForCausalLM.from_pretrained(
self.gen_ppl_eval_model_name_or_path,
torch_dtype=torch.bfloat16,
).to(self._eval_device).eval()
log_for_0("PPL model cached for reuse")
device = self._eval_device
if retokenize:
samples, attn_mask, eval_context_size = self._eval_retokenize(text_samples, max_length=max_length)
else:
samples = text_samples
attn_mask = np.ones(samples.shape)
eval_context_size = samples.shape[-1]
batch_size = self.eval_ppl_batch_size or samples.shape[0]
batch_size = min(batch_size, samples.shape[0]) or 1
num_batches = (samples.shape[0] + batch_size - 1) // batch_size
log_for_0(f"PPL: batch_size={batch_size}, {num_batches} batches")
per_sample_nll_sum = np.zeros(samples.shape[0], dtype=np.float64)
per_sample_token_count = np.zeros(samples.shape[0], dtype=np.float64)
for i in tqdm(range(num_batches), desc="Evaluating perplexity"):
batch_start = i * batch_size
batch_end = min((i + 1) * batch_size, samples.shape[0])
batch_samples = samples[batch_start:batch_end]
batch_attn_mask = attn_mask[batch_start:batch_end]
for chunk_start in range(0, batch_samples.shape[1], eval_context_size):
chunk_end = min(chunk_start + eval_context_size, batch_samples.shape[1])
sample_chunk = batch_samples[:, chunk_start:chunk_end]
attn_mask_chunk = batch_attn_mask[:, chunk_start:chunk_end]
input_ids = torch.from_numpy(sample_chunk).to(device).long()
attn = torch.from_numpy(attn_mask_chunk).to(device).long()
nlls, valid_tokens = self._compute_batch_nlls(input_ids, attn)
nlls_np = nlls.detach().cpu().numpy().astype(np.float64)
valid_tokens_np = valid_tokens.detach().cpu().numpy().astype(np.float64)
weighted_nlls = nlls_np * valid_tokens_np
self.gen_ppl.update(torch.from_numpy(weighted_nlls),
torch.from_numpy(valid_tokens_np))
per_sample_nll_sum[batch_start:batch_end] += weighted_nlls.sum(axis=-1)
per_sample_token_count[batch_start:batch_end] += valid_tokens_np.sum(axis=-1)
with np.errstate(divide="ignore", invalid="ignore"):
per_sample_ppl = np.exp(per_sample_nll_sum / per_sample_token_count)
per_sample_ppl = np.where(per_sample_token_count > 0, per_sample_ppl, np.nan).tolist()
per_sample_entropy = []
for i in range(samples.shape[0]):
valid_len = int(attn_mask[i].sum())
valid_tokens = samples[i, :valid_len]
_, counts = np.unique(valid_tokens, return_counts=True)
probs = counts.astype(np.float32) / counts.sum()
entropy = float(-np.sum(probs * np.log(probs + 1e-10)))
per_sample_entropy.append(entropy)
self.sample_entropy.update(entropy)
return {
"ppl": float(self.gen_ppl.compute()),
"per_sample_ppl": per_sample_ppl,
"mean_entropy": sum(per_sample_entropy) / len(per_sample_entropy),
}
|