File size: 11,578 Bytes
ae8ade0 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 | import gc
import logging
from utils.dataset import cycle
from utils.dataset import LatentLMDBDataset
from utils.distributed import EMA_FSDP, fsdp_wrap, fsdp_state_dict, launch_distributed_job
from utils.misc import (
set_seed,
merge_dict_list
)
import torch.distributed as dist
from omegaconf import OmegaConf
import torch
import wandb
import time
import os
from model import NaiveConsistency
class Trainer:
def __init__(self, config):
self.config = config
self.step = 0
# Step 1: Initialize the distributed training environment (rank, seed, dtype, logging etc.)
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
launch_distributed_job()
global_rank = dist.get_rank()
self.world_size = dist.get_world_size()
self.dtype = torch.bfloat16 if config.mixed_precision else torch.float32
self.device = torch.cuda.current_device()
self.is_main_process = global_rank == 0
self.causal = config.causal
self.disable_wandb = config.disable_wandb
# use a random seed for the training
if config.seed == 0:
random_seed = torch.randint(0, 10000000, (1,), device=self.device)
dist.broadcast(random_seed, src=0)
config.seed = random_seed.item()
set_seed(config.seed + global_rank)
if self.is_main_process and not self.disable_wandb:
wandb.login(host=config.wandb_host, key=config.wandb_key)
wandb.init(
config=OmegaConf.to_container(config, resolve=True),
name=config.config_name,
mode="online",
entity=config.wandb_entity,
project=config.wandb_project,
dir=config.wandb_save_dir
)
self.output_path = config.logdir
# Step 2: Initialize the model and optimizer
self.model = NaiveConsistency(config, device=self.device)
self.model.generator = fsdp_wrap(
self.model.generator,
sharding_strategy=config.sharding_strategy,
mixed_precision=config.mixed_precision,
wrap_strategy=config.generator_fsdp_wrap_strategy,
cpu_offload=True
)
self.model.generator_ema = fsdp_wrap(
self.model.generator_ema,
sharding_strategy=config.sharding_strategy,
mixed_precision=config.mixed_precision,
wrap_strategy=config.generator_fsdp_wrap_strategy,
cpu_offload=True
)
self.model.teacher = fsdp_wrap(
self.model.teacher,
sharding_strategy=config.sharding_strategy,
mixed_precision=config.mixed_precision,
wrap_strategy=config.real_score_fsdp_wrap_strategy,
cpu_offload=True
)
self.model.text_encoder = fsdp_wrap(
self.model.text_encoder,
sharding_strategy=config.sharding_strategy,
mixed_precision=config.mixed_precision,
wrap_strategy=config.text_encoder_fsdp_wrap_strategy,
cpu_offload=True
)
self.generator_optimizer = torch.optim.AdamW(
[param for param in self.model.generator.parameters()
if param.requires_grad],
lr=config.lr,
betas=(config.beta1, config.beta2),
weight_decay=config.weight_decay
)
self.generator_ema = EMA_FSDP(self.model.generator, decay=self.config.ema_weight)
# Step 3: Initialize the dataloader
dataset = LatentLMDBDataset(
config.data_path, max_pair=int(1e8))
sampler = torch.utils.data.distributed.DistributedSampler(
dataset, shuffle=True, drop_last=True)
dataloader = torch.utils.data.DataLoader(
dataset,
batch_size=config.batch_size,
sampler=sampler,
num_workers=8)
if dist.get_rank() == 0:
print("DATASET SIZE %d" % len(dataset))
self.dataloader = cycle(dataloader)
##############################################################################################################
# 6. Set up EMA parameter containers
rename_param = (
lambda name: name.replace("_fsdp_wrapped_module.", "")
.replace("_checkpoint_wrapped_module.", "")
.replace("_orig_mod.", "")
)
self.name_to_trainable_params = {}
for n, p in self.model.generator.named_parameters():
if not p.requires_grad:
continue
renamed_n = rename_param(n)
self.name_to_trainable_params[renamed_n] = p
ema_weight = config.ema_weight
self.generator_ema = None
if (ema_weight is not None) and (ema_weight > 0.0):
print(f"Setting up EMA with weight {ema_weight}")
self.generator_ema = EMA_FSDP(self.model.generator, decay=ema_weight)
##############################################################################################################
# 7. Load the causal diffusion model as the teacher model
if getattr(config, "generator_ckpt", False):
print(f"Loading pretrained generator from {config.generator_ckpt}")
state_dict = torch.load(config.generator_ckpt, map_location="cpu")
if "generator" in state_dict:
state_dict = state_dict["generator"]
fixed = {}
for k, v in state_dict.items():
if k.startswith("model._fsdp_wrapped_module."):
k = k.replace("model._fsdp_wrapped_module.", "model.", 1)
fixed[k] = v
state_dict = fixed
elif "model" in state_dict:
state_dict = state_dict["model"]
elif "generator_ema" in state_dict:
gen_sd = state_dict["generator_ema"]
fixed = {}
for k, v in gen_sd.items():
if k.startswith("model._fsdp_wrapped_module."):
k = k.replace("model._fsdp_wrapped_module.", "model.", 1)
fixed[k] = v
state_dict = fixed
self.model.generator.load_state_dict(
state_dict, strict=True
)
self.model.teacher.load_state_dict(
state_dict, strict=True
)
#############################################################################################################
self.max_grad_norm_generator = getattr(config, "max_grad_norm_generator", 10.0)
self.max_grad_norm_critic = getattr(config, "max_grad_norm_critic", 10.0)
self.previous_time = None
def save(self):
print("Start gathering distributed model states...")
generator_state_dict = fsdp_state_dict(
self.model.generator)
if self.config.ema_start_step < self.step:
state_dict = {
"generator_ema": self.generator_ema.full_state_dict(self.model.generator),
}
else:
state_dict = {
"generator": generator_state_dict,
}
if self.is_main_process:
os.makedirs(os.path.join(self.output_path,
f"checkpoint_model_{self.step:06d}"), exist_ok=True)
torch.save(state_dict, os.path.join(self.output_path,
f"checkpoint_model_{self.step:06d}", "model.pt"))
print("Model saved to", os.path.join(self.output_path,
f"checkpoint_model_{self.step:06d}", "model.pt"))
def fwdbwd_one_step(self, batch, clean_latent=None):
self.model.eval()
if self.step % 20 == 0:
torch.cuda.empty_cache()
# Step 1: Get the next batch of text prompts
text_prompts = batch["prompts"]
batch_size = len(text_prompts)
image_or_video_shape = list(self.config.image_or_video_shape)
image_or_video_shape[0] = batch_size
# Step 2: Extract the conditional infos
with torch.no_grad():
conditional_dict = self.model.text_encoder(
text_prompts=text_prompts)
if not getattr(self, "unconditional_dict", None):
unconditional_dict = self.model.text_encoder(
text_prompts=[self.config.negative_prompt] * batch_size)
unconditional_dict = {k: v.detach()
for k, v in unconditional_dict.items()}
self.unconditional_dict = unconditional_dict # cache the unconditional_dict
else:
unconditional_dict = self.unconditional_dict
# Step 3: Store gradients for the generator (if training the generator)
generator_loss, generator_log_dict = self.model.generator_loss(
conditional_dict=conditional_dict,
unconditional_dict=unconditional_dict,
clean_latent=clean_latent,
ema_model = self.generator_ema
)
generator_loss.backward()
generator_grad_norm = self.model.generator.clip_grad_norm_(
self.max_grad_norm_generator)
generator_log_dict.update({"generator_loss": generator_loss,
"generator_grad_norm": generator_grad_norm})
return generator_log_dict
def train(self):
start_step = self.step
while True:
self.generator_optimizer.zero_grad(set_to_none=True)
batch = next(self.dataloader)
generator_log_dict = self.fwdbwd_one_step(batch, clean_latent=batch["clean_latent"])
self.generator_optimizer.step()
if self.generator_ema is not None:
self.generator_ema.update(self.model.generator)
# Increment the step since we finished gradient update
self.step += 1
# Save the model
if (not self.config.no_save) and (self.step - start_step) > 0 and self.step % self.config.log_iters == 0:
torch.cuda.empty_cache()
self.save()
torch.cuda.empty_cache()
# Logging
if self.is_main_process:
wandb_loss_dict = {}
wandb_loss_dict.update(
{
"generator_loss": generator_log_dict["generator_loss"].mean().item(),
"generator_grad_norm": generator_log_dict["generator_grad_norm"].mean().item()
}
)
if not self.disable_wandb:
wandb.log(wandb_loss_dict, step=self.step)
if self.step % self.config.gc_interval == 0:
if dist.get_rank() == 0:
logging.info("DistGarbageCollector: Running GC.")
gc.collect()
torch.cuda.empty_cache()
if self.is_main_process:
current_time = time.time()
if self.previous_time is None:
self.previous_time = current_time
else:
if not self.disable_wandb:
wandb.log({"per iteration time": current_time - self.previous_time}, step=self.step)
self.previous_time = current_time
|