File size: 10,423 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 | import datetime
import os
import os.path
import numpy as np
import torch
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel as DDP
import torch.nn.functional as F
import pickle
from functools import partial
import math
import data
import losses
import sampling
import utils.utils as utils
import utils.seq_utils as sutils
from model.ema import ExponentialMovingAverage
from hypersphere import Hypersphere
from hydra.utils import instantiate
torch.backends.cudnn.benchmark = True
# torch.autograd.set_detect_anomaly(True)
import wandb
from evaluation import Eval
def setup(rank, world_size, port):
os.environ["MASTER_ADDR"] = "localhost"
os.environ["MASTER_PORT"] = str(port)
# 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")
checkpoint_dir = os.path.join(work_dir, "checkpoints")
checkpoint_meta_dir = os.path.join(work_dir, "checkpoints-meta", "checkpoint.pth")
if rank == 0:
utils.makedirs(sample_dir)
utils.makedirs(checkpoint_dir)
utils.makedirs(os.path.dirname(checkpoint_meta_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)
if rank == 0 and cfg.use_wandb:
wandb.init(project=f'discrete-{cfg.data.train}', name=cfg.wandb_run_name)
wandb.config = 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())
# build sde
token_size = cfg.tokens
vocab_size = cfg.tokens + (1 if 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 = cfg.model.length * base_max_length
_base_n = partial(sutils.base_n, base=token_size, max_length=base_max_length)
mprint(f"Data {cfg.data.train} with base-{token_size}, base_length={base_max_length}")
scheduler = instantiate(cfg.scheduler, device=device)
prior_dist = instantiate(cfg.sde.prior_dist, device=device, batch_dims=(cfg.training.batch_size * 2 // (cfg.ngpus * cfg.training.accum), seq_length, vocab_size))
try:
with open(os.path.join(cfg.work_dir, "sde.pkl"), "rb") as f:
alphas, rhos = pickle.load(f)
sde = instantiate(
cfg.sde, manifold=manifold, scheduler=scheduler, prior_dist=prior_dist,
preprocessed=(alphas.to(device), rhos.to(device))
)
mprint(f"Loaded preprocessed sde.")
except:
mprint(f"No preprocessed sde found.")
sde = instantiate(
cfg.sde, manifold=manifold, scheduler=scheduler, prior_dist=prior_dist, device=device,
)
if rank==0:
with open(os.path.join(work_dir, "sde.pkl"), "wb") as f:
pickle.dump((sde.alphas.detach().cpu(), sde.rhos.detach().cpu()), f, protocol=pickle.HIGHEST_PROTOCOL)
if rank == 0 and len(sde.alphas.shape)==1:
import matplotlib.pyplot as plt
ts = np.linspace(0, 1, sde.preprocess_steps)
fig, axs = plt.subplots(1, 2, figsize=(18,6))
axs[0].plot(ts, sde.alphas.detach().cpu().numpy(), color='b')
axs[0].title.set_text("alphas")
axs[1].plot(ts, sde.rhos.detach().cpu().numpy(), color='b')
axs[1].title.set_text("rhos")
fig.suptitle(f"beta_0={sde.scheduler.beta_0:.1e}, beta_f={sde.scheduler.beta_f:.1e}")
fig.savefig(os.path.join(cfg.work_dir, 'moments.png'), dpi=300, bbox_inches="tight")
# build model
drift_model = instantiate(cfg.model).to(device)
drift_model = torch.compile(drift_model)
drift_model = DDP(drift_model, device_ids=[rank], static_graph=True, find_unused_parameters=True)
num_parameters = sum(p.numel() for p in drift_model.parameters())
mprint(f"Number of parameters in the model: {num_parameters}")
ema = ExponentialMovingAverage(
drift_model.parameters(), decay=cfg.training.ema
)
# build optimization state
optimizer = losses.get_optimizer(cfg, drift_model.parameters())
scaler = torch.cuda.amp.GradScaler()
state = dict(
optimizer=optimizer, scaler=scaler, model=drift_model, sde=sde, ema=ema, step=0, config=cfg
)
# load in state
state = utils.restore_checkpoint(checkpoint_meta_dir, state, device)
initial_step = int(state['step'])
# Build data iterators
train_ds, eval_ds = data.get_dataloaders(cfg)
train_iter = iter(train_ds)
eval_iter = iter(eval_ds)
# load in tokenizer
tokenizer = data.get_tokenizer(cfg.data.train)
# Build one-step training and evaluation functions
optimize_fn = losses.optimization_manager(cfg)
loss_type = cfg.training.loss_type
eval_loss_type = "elbo"
mprint(f"Training loss type: {loss_type} | Eval loss type: {eval_loss_type} | weight type: {scheduler.weight_type}")
train_step_fn = losses.get_step_fn(
loss_type=loss_type, sde=sde, train=True, optimize_fn=optimize_fn, accum=cfg.training.accum
)
eval_step_fn = losses.get_step_fn(
loss_type=eval_loss_type, sde=sde, train=False, optimize_fn=optimize_fn, accum=cfg.training.accum
)
sampling_eps = 1e-5
if cfg.training.snapshot_sampling:
sampling_batch_size = cfg.training.batch_size * 2 // (cfg.ngpus * cfg.training.accum)
sampling_shape = (sampling_batch_size, seq_length, vocab_size)
sampling_fn = sampling.get_sampling_fn(cfg, sde, sampling_shape, sampling_eps, device)
evaluator = Eval(cfg, sde, distributed=True)
shift_and_decode = sutils.find_bos_and_shift_fn(token_size, base_max_length, tokenizer)
num_train_steps = cfg.training.n_iters
mprint(f"Starting training loop at step {initial_step}.")
while state['step'] < num_train_steps + 1:
step = state['step']
# Convert indices into onehot vectors
batch = next(train_iter)['input_ids'].to(device)
batch = _base_n(batch)
batch = F.one_hot(batch, num_classes=vocab_size).to(torch.float32)
loss = train_step_fn(state, batch)
if step != state['step']:
if step % cfg.training.log_freq == 0 and step > 0:
dist.all_reduce(loss)
loss /= world_size
mprint("step: %d, train_loss: %.3e" % (step, loss.item()))
if rank == 0 and cfg.use_wandb:
wandb.log({'train_loss' : loss.item()}, step=step)
if step % cfg.training.snapshot_freq_for_preemption == 0 and rank == 0:
utils.save_checkpoint(checkpoint_meta_dir, state)
if step % cfg.training.eval_freq == 0 and step > 0:
# Convert indices into onehot vectors
eval_batch = next(eval_iter)['input_ids'].to(device)
eval_batch = _base_n(eval_batch)
eval_batch = F.one_hot(eval_batch, num_classes=vocab_size).to(torch.float32)
eval_loss = eval_step_fn(state, eval_batch)
dist.all_reduce(eval_loss)
eval_loss /= world_size
mprint("step: %d, val_loss: %.3e" % (step, eval_loss.item()))
if rank == 0 and cfg.use_wandb:
wandb.log({'val_loss' : eval_loss.item()}, step=step)
del eval_batch
if step > 0 and step % cfg.training.snapshot_freq == 0 or step == num_train_steps:
# Save the checkpoint.
save_step = step // cfg.training.snapshot_freq
if rank == 0:
utils.save_checkpoint(os.path.join(
checkpoint_dir, f'checkpoint_{save_step}.pth'), state
)
# Generate and save samples
if cfg.training.snapshot_sampling:
mprint(f"Generating text at step: {step}")
this_sample_dir = os.path.join(sample_dir, "iter_{}".format(step))
utils.makedirs(this_sample_dir)
ema.store(drift_model.parameters())
ema.copy_to(drift_model.parameters())
sample = sampling_fn(drift_model) # indices in base-n
sentences, sample = shift_and_decode(sample.detach().cpu())
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")
result_dict = evaluator.eval(
sample=sample, drift_model=drift_model
)
for k, v in result_dict.items():
mprint(f"Step {step}. {k:8s}: {v:.3f}")
if rank == 0 and cfg.use_wandb:
wandb.log({k : v}, step=step)
ema.restore(drift_model.parameters())
dist.barrier() |