File size: 9,484 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 | import torch
import torch.distributed as dist
import torch.nn.functional as F
import math
import numpy as np
from tqdm import tqdm
from transformers import logging, GPT2LMHeadModel, AutoModelForCausalLM, AutoTokenizer
logging.set_verbosity_error() # Set logging to show only errors (this suppresses warnings)
import gc
from functools import partial
from data import get_test_dataloaders, get_tokenizer
import losses
import utils.seq_utils as sutils
class Eval:
def __init__(self, cfg, sde, **kwargs):
self.cfg = cfg
self.sde = sde
self.distributed = kwargs.get('distributed', True)
self.tokenizer = get_tokenizer(cfg.data.train)
# Load gpt2 tokenizer if the current tokenizer is not gpt2 tokenizer
# E.g., text8 and lm1b
if cfg.data.train in ["text8", "lm1b"]:
self.gpt_tokenizer = AutoTokenizer.from_pretrained("gpt2")
self.gpt_tokenizer.pad_token = self.gpt_tokenizer.eos_token
token_size = cfg.tokens
base_max_length = math.ceil(math.log(cfg.original_tokens) / math.log(token_size))
self._base_n = partial(sutils.base_n, base=token_size, max_length=base_max_length)
self.gen_ppl_batch_size = 16 # Adjust this for your GPU memory
def eval(self, sample=None, drift_model=None):
result_dict = {}
self.device = drift_model.device
# Measure token entropy
if self.cfg.eval.entropy:
result_dict["entropy"] = self.entropy(sample)
if self.cfg.data.train == "text8":
try:
if self.cfg.eval.nll:
result_dict['bpc'] = self.bpc(drift_model)
if self.cfg.eval.eval_model=="gptj":
result_dict['gptj_nll'] = self.gptj_nll(sample)
except Exception as e:
print(e)
return result_dict
# Measure generative perplexity
if self.cfg.eval.gen_ppl:
try:
result_dict['gpt2_ppl'] = self.gpt2_ppl(sample)
except Exception as e:
print(e)
if self.cfg.eval.nll:
try:
result_dict["nll"] = self.nll(drift_model)
except Exception as e:
print(e)
return result_dict
# Token Entropy
def entropy(self, sample):
if self.cfg.data.train in ["text8"]:
sample = self.gpt_tokenizer(self.tokenizer.batch_decode(sample)).input_ids
sample_stats = {}
for tokens in sample:
for token in tokens:
if isinstance(token, torch.Tensor):
token = token.item()
if token not in sample_stats:
sample_stats[token] = 0
sample_stats[token] += 1
sample_probs = {}
total_sample_stats = sum(sample_stats.values())
sample_entropy = 0
for k, v in sample_stats.items():
p = v / total_sample_stats
sample_probs[k] = p
if p > 0:
sample_entropy += -p * np.log(p)
entropy = torch.tensor(sample_entropy, device=self.device)
if self.distributed:
dist.all_reduce(entropy)
entropy /= dist.get_world_size()
return entropy
def mc_estimate(self, drift_model, batch):
loss_fn = losses.elbo_loss_fn(
self.sde, False, simul_steps=self.cfg.eval.simul_steps
)
return loss_fn(drift_model, batch)
# Upper bound of negative log-likelihood (nll)
def nll(self, drift_model):
test_ds = get_test_dataloaders(self.cfg)
test_iter = iter(test_ds)
with torch.no_grad():
nelbo, N = 0., 0
for _ in tqdm(range(self.cfg.eval.n_iters), desc='[NLL]', position=0, leave=False):
batch = next(test_iter)['input_ids'].to(self.device)
batch = self._base_n(batch)
batch = F.one_hot(batch, num_classes=self.sde.manifold.dim+1).to(torch.float32)
nelbo_batch = self.mc_estimate(drift_model, batch).sum()
nelbo += nelbo_batch
N += batch.shape[0]
nelbo = nelbo / N
if self.distributed:
dist.all_reduce(nelbo)
nelbo /= dist.get_world_size()
del batch
return nelbo
# Bits-per-character (bpc)
def bpc(self, drift_model):
nelbo = self.nll(drift_model)
return nelbo / self.cfg.model.length / np.log(2)
# GPT-J generative ppl
def gptj_nll(self, sample):
batch_size = self.gen_ppl_batch_size
with torch.no_grad():
eval_model = AutoModelForCausalLM.from_pretrained(
"EleutherAI/gpt-j-6B", torch_dtype=torch.bfloat16, device_map="auto"
)
gen_nll, N = 0., 0
for k in tqdm(range(len(sample) // batch_size), leave=False, desc='[GPTJ NLL]'):
batch = sample[batch_size*k:batch_size*(k+1)]
output = self.gpt_tokenizer(
self.tokenizer.batch_decode(batch), padding='longest', return_tensors='pt'
).to(self.device)
input_ids = output.input_ids
ori_len = output.attention_mask.sum(dim=1)
logits = eval_model(
input_ids=input_ids,
attention_mask=output.attention_mask,
labels=input_ids,
return_dict=True
).logits
logits = logits.transpose(-1, -2)
for j in range(len(logits)):
s = input_ids[j][:ori_len[j]].unsqueeze(0)
l = logits[j][...,:ori_len[j]].unsqueeze(0)
gen_nll += F.cross_entropy(l[..., :-1], s[..., 1:], reduction="none").mean()
N += 1
gen_nll /= N
if self.distributed:
dist.all_reduce(gen_nll)
gen_nll /= dist.get_world_size()
del eval_model, logits, output, sample, l
gc.collect()
torch.cuda.empty_cache()
return gen_nll
# GPT2 generative ppl
def gpt2_ppl(self, sample):
eval_model = GPT2LMHeadModel.from_pretrained("gpt2-large").to(self.device).eval()
batch_size = self.gen_ppl_batch_size
gen_ppl, N = 0., 0
with torch.no_grad():
if self.cfg.data.train in ["text8", "lm1b"]:
for k in tqdm(range(len(sample) // batch_size), leave=False, desc='[Gen. PPL]'):
batch = sample[batch_size*k:batch_size*(k+1)]
output = self.gpt_tokenizer(
self.tokenizer.batch_decode(batch), padding='longest', return_tensors='pt'
).to(self.device)
input_ids = output.input_ids
ori_len = output.attention_mask.sum(dim=1)
logits = eval_model(
input_ids=input_ids,
attention_mask=output.attention_mask,
labels=input_ids,
return_dict=True
).logits
logits = logits.transpose(-1, -2)
for j in range(len(logits)):
s = input_ids[j][:ori_len[j]].unsqueeze(0)
l = logits[j][...,:ori_len[j]].unsqueeze(0)
gen_ppl += F.cross_entropy(
l[..., :-1], s[..., 1:], reduction="none"
).mean().exp()
N += 1
gen_ppl /= N
else:
_sample = torch.ones(
(len(sample), self.cfg.model.length), device=self.device, dtype=torch.long
) * self.tokenizer.pad_token_id
attention_mask = torch.zeros_like(_sample)
for i, s in enumerate(sample):
_sample[i, :len(s)] = s
attention_mask[i, :len(s)] = 1
sample = _sample.clip(max=self.tokenizer.vocab_size-1)
for k in tqdm(range(len(sample) // batch_size), leave=False, desc='[Gen. PPL]'):
input_ids = sample[batch_size*k:batch_size*(k+1)]
ori_len = attention_mask[batch_size*k:batch_size*(k+1)].sum(dim=1)
logits = eval_model(
input_ids=input_ids,
attention_mask=attention_mask[batch_size*k:batch_size*(k+1)],
labels=input_ids,
return_dict=True
).logits
logits = logits.transpose(-1, -2)
for j in range(len(logits)):
s = input_ids[j][:ori_len[j]].unsqueeze(0)
l = logits[j][...,:ori_len[j]].unsqueeze(0)
gen_ppl += F.cross_entropy(
l[..., :-1], s[..., 1:], reduction="none"
).mean().exp()
N += 1
gen_ppl /= N
if self.distributed:
dist.all_reduce(gen_ppl)
gen_ppl /= dist.get_world_size()
del eval_model, logits
gc.collect()
torch.cuda.empty_cache()
return gen_ppl |