File size: 7,429 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 | import datetime
import os
import os.path
import torch
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel as DDP
import math
import pickle
from tqdm import tqdm
import gc
from hydra.utils import instantiate
from omegaconf import OmegaConf
import data
import sampling
import utils.utils as utils
from model.ema import ExponentialMovingAverage
from hypersphere import Hypersphere
from evaluation import Eval
import utils.seq_utils as sutils
torch.backends.cudnn.benchmark = True
# torch.autograd.set_detect_anomaly(True)
def setup(rank, world_size, port):
os.environ["MASTER_ADDR"] = "localhost"
os.environ["MASTER_PORT"] = str(port)
os.environ["TOKENIZERS_PARALLELISM"] = "false"
# initialize the process group
dist.init_process_group(
"nccl", rank=rank, world_size=world_size, timeout=datetime.timedelta(minutes=30)
)
# for clean terminal print
import sys
f = open(os.devnull, "w")
if rank != 0:
sys.stdout = f
sys.stderr = f
def cleanup():
dist.destroy_process_group()
def run_multiprocess(rank, world_size, cfg, port):
try:
setup(rank, world_size, port)
_run(rank, world_size, cfg)
finally:
cleanup()
def _run(rank, world_size, cfg):
torch.cuda.set_device(rank)
work_dir = cfg.work_dir
# Create directories for experimental logs
sample_dir = os.path.join(work_dir, "samples")
if rank == 0:
utils.makedirs(sample_dir)
# logging
if rank == 0:
logger = utils.get_logger(os.path.join(work_dir, "run.log"))
def mprint(msg):
if rank == 0:
logger.info(msg)
mprint(work_dir)
mprint(cfg)
device = torch.device(f"cuda:{rank}" if torch.cuda.is_available() else "cpu")
if device.type == "cuda":
mprint("Found {} CUDA devices.".format(torch.cuda.device_count()))
for i in range(torch.cuda.device_count()):
props = torch.cuda.get_device_properties(i)
mprint(
"{} \t Memory: {:.2f}GB".format(
props.name, props.total_memory / (1024 ** 3)
)
)
else:
mprint("WARNING: Using device {}".format(device))
mprint(f"Found {os.cpu_count()} total number of CPUs.")
mprint(f"SEED: {cfg.seed}")
utils.set_seed(cfg.seed+dist.get_rank())
# load saved state from checkpoint
loaded_state = torch.load(cfg.model_path, map_location=device)
train_cfg = loaded_state['config']
# Update sampling config from training config
OmegaConf.set_struct(cfg, False)
OmegaConf.update(cfg, "model", train_cfg.model)
OmegaConf.update(cfg, "training", train_cfg.training)
OmegaConf.update(cfg, "data", train_cfg.data)
OmegaConf.update(cfg, "tokens", train_cfg.tokens)
OmegaConf.update(cfg, "original_tokens", train_cfg.original_tokens)
if cfg.sampling.batch_per_gpu == 0:
batch_per_gpu = train_cfg.training.batch_size // (train_cfg.ngpus * train_cfg.training.accum)
OmegaConf.update(cfg, "sampling.batch_per_gpu", batch_per_gpu)
if cfg.eval.batch_size == 0:
OmegaConf.update(cfg, "eval.batch_size", cfg.sampling.batch_per_gpu * cfg.ngpus * cfg.training.accum)
# build sde
token_size = train_cfg.tokens
vocab_size = train_cfg.tokens + (1 if train_cfg.sde.prior_dist.add_mask_token else 0)
manifold = Hypersphere(vocab_size-1)
base_max_length = math.ceil(math.log(cfg.original_tokens) / math.log(token_size))
seq_length = train_cfg.model.length * base_max_length
mprint(f"Data {train_cfg.data.train} with base-{token_size}, base_length={base_max_length}")
scheduler = instantiate(train_cfg.scheduler, device=device)
prior_dist = instantiate(train_cfg.sde.prior_dist, device=device, batch_dims=(cfg.sampling.batch_per_gpu, seq_length, vocab_size))
try:
base_dir = os.path.dirname(os.path.dirname(cfg.model_path))
with open(os.path.join(base_dir, "sde.pkl"), "rb") as f:
alphas, rhos = pickle.load(f)
sde = instantiate(
train_cfg.sde, manifold=manifold, scheduler=scheduler, prior_dist=prior_dist,
preprocessed=(alphas.to(device), rhos.to(device))
)
print("Loaded preprocessed sde.")
except:
print("No preprocessed sde found.")
sde = instantiate(
cfg.sde, manifold=manifold, scheduler=scheduler, prior_dist=prior_dist,
device=device
)
# build model
drift_model = instantiate(train_cfg.model).to(device)
# checkpoint for torch compiled model
if "_orig_mod" in [*loaded_state['model']][0]:
drift_model = torch.compile(drift_model)
drift_model.load_state_dict(loaded_state["model"])
else:
drift_model.load_state_dict(loaded_state["model"])
drift_model = torch.compile(drift_model)
drift_model = DDP(drift_model, device_ids=[rank]).eval()
ema = ExponentialMovingAverage(
drift_model.parameters(), decay=train_cfg.training.ema
)
ema.load_state_dict(loaded_state["ema"])
step = int(loaded_state['step'])
# load ema parameters
ema.store(drift_model.parameters())
ema.copy_to(drift_model.parameters())
# load in tokenizer
tokenizer = data.get_tokenizer(cfg.data.train)
# define sampling function
sampling_eps = 1e-5
sampling_shape = (cfg.sampling.batch_per_gpu, seq_length, vocab_size)
sampling_fn = sampling.get_sampling_fn(
cfg, sde, sampling_shape, sampling_eps, device, proj_fn=(lambda x: x)
)
mprint(f"Generating text at step: {step}")
this_sample_dir = os.path.join(sample_dir, f"iter_{step}")
utils.makedirs(this_sample_dir)
num_samples = 256 if cfg.data.train=='text8' else 8
# num_samples = 4000 if cfg.data.train=='text8' else 512
samples = []
for i in tqdm(range(math.ceil(num_samples / (sampling_shape[0] * world_size))), leave=False):
sample = sampling_fn(drift_model).detach().cpu() # Shape: B x (mxL) x V
samples.append(sample)
gc.collect()
torch.cuda.empty_cache()
samples = torch.cat(samples, dim=0)
shift_and_decode = sutils.find_bos_and_shift_fn(token_size, base_max_length, tokenizer)
sentences, samples = shift_and_decode(samples)
file_name = os.path.join(this_sample_dir, f"sample_{rank}.txt")
with open(file_name, 'w', encoding='utf8') as file:
for sentence in sentences:
file.write(sentence + "\n")
file.write("="*100+"\n")
with open(os.path.join(this_sample_dir, f"samples_{rank}.pkl"), 'wb') as file:
pickle.dump(samples, file, protocol=pickle.HIGHEST_PROTOCOL)
dist.barrier()
if rank == 0:
# Evaluation performed on single GPU due to measuring entropy.
# May be modified to perform evaluation on multiple GPUs.
evaluator = Eval(cfg, sde, distributed=False)
samples = []
for i in range(world_size):
with open(os.path.join(this_sample_dir, f"samples_{i}.pkl"), 'rb') as file:
sample = pickle.load(file)
samples.extend(sample)
print(f"{len(samples)} samples loaded.")
result_dict = evaluator.eval(sample=samples, drift_model=drift_model)
for k, v in result_dict.items():
mprint(f"Step {step}. {k:8s}: {v:.3f}")
dist.barrier() |