tfwang's picture
Update glide_text2im/train_util.py
9d92961
import copy
import functools
import os
import blobfile as bf
import numpy as np
import torch as th
import torch.distributed as dist
from torch.nn.parallel.distributed import DistributedDataParallel as DDP
from torch.optim import AdamW
from .glide_util import sample
from . import logger
from .fp16_util import (
make_master_params,
master_params_to_model_params,
model_grads_to_master_grads,
unflatten_master_params,
zero_grad,
)
from .nn import update_ema
from .vgg import VGG
from .adv import AdversarialLoss
from .resample import LossAwareSampler, UniformSampler
import glob
import torchvision.utils as tvu
import PIL.Image as Image
# For ImageNet experiments, this was a good default value.
# We found that the lg_loss_scale quickly climbed to
# 20-21 within the first ~1K steps of training.
INITIAL_LOG_LOSS_SCALE = 20.0
class TrainLoop:
def __init__(
self,
model,
glide_options,
diffusion,
data,
val_data,
batch_size,
microbatch,
lr,
ema_rate,
log_interval,
save_interval,
resume_checkpoint,
use_fp16=False,
fp16_scale_growth=1e-3,
schedule_sampler=None,
weight_decay=0.0,
lr_anneal_steps=0,
finetune_decoder = False,
mode = '',
use_vgg = False,
use_gan = False,
uncond_p = 0,
super_res = 0,
):
self.model = model
self.glide_options=glide_options
self.diffusion = diffusion
self.data = data
self.val_data=val_data
self.batch_size = batch_size
self.microbatch = microbatch if microbatch > 0 else batch_size
self.lr = lr
self.ema_rate = (
[ema_rate]
if isinstance(ema_rate, float)
else [float(x) for x in ema_rate.split(",")]
)
self.log_interval = log_interval
self.save_interval = save_interval
self.resume_checkpoint = find_resume_checkpoint(resume_checkpoint)
self.use_fp16 = use_fp16
self.fp16_scale_growth = fp16_scale_growth
self.schedule_sampler = schedule_sampler or UniformSampler(diffusion)
self.weight_decay = weight_decay
self.lr_anneal_steps = lr_anneal_steps
self.step = 0
self.resume_step = 0
self.global_batch = self.batch_size * dist.get_world_size()
if use_vgg:
self.vgg = VGG(conv_index='22').cuda()
print('use perc')
else:
self.vgg = None
if use_gan:
self.adv = AdversarialLoss()
print('use adv')
else:
self.adv = None
self.super_res = super_res
self.uncond_p =uncond_p
self.mode = mode
self.finetune_decoder = finetune_decoder
if finetune_decoder:
self.optimize_model = self.model
else:
self.optimize_model = self.model.encoder
self.model_params = list(self.optimize_model.parameters())
self.master_params = self.model_params
self.lg_loss_scale = INITIAL_LOG_LOSS_SCALE
self.sync_cuda = th.cuda.is_available()
self._load_and_sync_parameters()
if self.use_fp16:
self._setup_fp16()
self.opt = AdamW(self.master_params, lr=self.lr, weight_decay=self.weight_decay)
if self.resume_step:
self._load_optimizer_state()
# Model was resumed, either due to a restart or a checkpoint
# being specified at the command line.
self.ema_params = [
self._load_ema_parameters(rate) for rate in self.ema_rate
]
else:
self.ema_params = [
copy.deepcopy(self.master_params) for _ in range(len(self.ema_rate))
]
if th.cuda.is_available():
self.use_ddp = True
self.ddp_model = DDP(
self.model,
device_ids=[torch.device('cuda')],
output_device=torch.device('cuda'),
broadcast_buffers=False,
bucket_cap_mb=128,
find_unused_parameters=False,
)
else:
if dist.get_world_size() > 1:
logger.warn(
"Distributed training requires CUDA. "
"Gradients will not be synchronized properly!"
)
self.use_ddp = False
self.ddp_model = self.model
def _load_and_sync_parameters(self):
resume_checkpoint = self.resume_checkpoint
if resume_checkpoint:
self.resume_step = parse_resume_step_from_filename(resume_checkpoint)
if dist.get_rank() == 0:
logger.log(f"loading model from checkpoint: {resume_checkpoint}...")
self.model.load_state_dict(th.load(resume_checkpoint, map_location="cpu"),strict=False)
#dist_util.sync_params(self.model.parameters())
def _load_ema_parameters(self, rate):
ema_params = copy.deepcopy(self.master_params)
main_checkpoint = self.resume_checkpoint
ema_checkpoint = find_ema_checkpoint(main_checkpoint, self.resume_step, rate)
if ema_checkpoint:
if dist.get_rank() == 0:
logger.log(f"loading EMA from checkpoint: {ema_checkpoint}...")
state_dict = th.load(ema_checkpoint, map_location=torch.device('cuda'))
ema_params = self._state_dict_to_master_params(state_dict)
#dist_util.sync_params(ema_params)
return ema_params
def _load_optimizer_state(self):
main_checkpoint = self.resume_checkpoint
opt_checkpoint = bf.join(
bf.dirname(main_checkpoint), f"opt{self.resume_step:06}.pt"
)
if bf.exists(opt_checkpoint):
logger.log(f"loading optimizer state from checkpoint: {opt_checkpoint}")
state_dict = th.load(opt_checkpoint, map_location="cpu")
try:
self.opt.load_state_dict(state_dict)
except:
pass
def _setup_fp16(self):
self.master_params = make_master_params(self.model_params)
self.model.convert_to_fp16()
def run_loop(self):
while (
not self.lr_anneal_steps
or self.step <= self.lr_anneal_steps
):
batch, model_kwargs = next(self.data)
# uncond_p = 0
# if self.super_res:
# uncond_p = 0
# elif self.finetune_decoder:
# uncond_p = self.uncond_p
# elif self.step > self.lr_anneal_steps - 40000:
# uncond_p = self.uncond_p
self.run_step(batch, model_kwargs)
if self.step % self.log_interval == 0:
logger.dumpkvs()
if self.step % self.save_interval == 0:
self.save()
self.val(self.step)
self.step += 1
if (self.step - 1) % self.save_interval != 0:
self.save()
def run_step(self, batch, model_kwargs):
self.forward_backward(batch, model_kwargs)
if self.use_fp16:
self.optimize_fp16()
else:
self.optimize_normal()
self.log_step()
def forward_backward(self, batch, model_kwargs):
zero_grad(self.model_params)
for i in range(0, batch.shape[0], self.microbatch):
micro = batch[i : i + self.microbatch].cuda()
micro_cond={n:model_kwargs[n][i:i+self.microbatch].cuda() for n in model_kwargs if n in ['ref', 'low_res']}
last_batch = (i + self.microbatch) >= batch.shape[0]
t, weights = self.schedule_sampler.sample(micro.shape[0], torch.device('cuda'))
if self.step <100:
vgg_loss = None
adv_loss = None
else:
vgg_loss = self.vgg
adv_loss = self.adv
compute_losses = functools.partial(
self.diffusion.training_losses,
self.ddp_model,
micro,
t,
vgg_loss,
adv_loss,
model_kwargs=micro_cond,
)
if last_batch or not self.use_ddp:
losses = compute_losses()
else:
with self.ddp_model.no_sync():
losses = compute_losses()
if isinstance(self.schedule_sampler, LossAwareSampler):
self.schedule_sampler.update_with_local_losses(
t, losses["loss"].detach()
)
loss = (losses["loss"] * weights).mean()
log_loss_dict(
self.diffusion, t, {k: v * weights for k, v in losses.items()}
)
if self.use_fp16:
loss_scale = 2 ** self.lg_loss_scale
(loss * loss_scale).backward()
else:
loss.backward()
def val(self, step):
inner_model=self.ddp_model.module
inner_model.eval()
if dist.get_rank() == 0:
print("sampling...")
s_path = os.path.join(logger.get_dir(), 'results')
os.makedirs(s_path,exist_ok=True)
img_id = 0
guidance_scale=self.glide_options['sample_c']
while (True):
if img_id >= self.glide_options['num_samples']:
break
batch, model_kwargs = next(self.val_data)
with th.no_grad():
samples=sample(
glide_model=inner_model,
glide_options=self.glide_options,
side_x=self.glide_options['image_size'],
side_y=self.glide_options['image_size'],
prompt=model_kwargs,
batch_size=self.glide_options['batch_size']//2,
guidance_scale=guidance_scale,
device=torch.device('cuda'),
prediction_respacing=self.glide_options['sample_respacing'],
upsample_enabled=self.glide_options['super_res'],
upsample_temp=0.997,
mode = self.mode,
)
samples = samples.cpu()
ref = model_kwargs['ref_ori']
# LR = model_kwargs['low_res'].cpu()
for i in range(samples.size(0)):
out_path = os.path.join(s_path, f"{dist.get_rank()}_{img_id}_step{step}_{guidance_scale}_output.png")
tvu.save_image(
(samples[i]+1)*0.5, out_path)
out_path = os.path.join(s_path, f"{dist.get_rank()}_{img_id}_step{step}_{guidance_scale}_gt.png")
tvu.save_image(
(batch[i]+1)*0.5, out_path)
out_path = os.path.join(s_path, f"{dist.get_rank()}_{img_id}_step{step}_{guidance_scale}_ref.png")
tvu.save_image(
(ref[i]+1)*0.5, out_path)
# out_path = os.path.join(s_path, f"{dist.get_rank()}_{img_id}_step{step}_{guidance_scale}_lr.png")
# tvu.save_image(
# (LR[i]+1)*0.5, out_path)
img_id += 1
inner_model.train()
def optimize_fp16(self):
if any(not th.isfinite(p.grad).all() for p in self.model_params):
self.lg_loss_scale -= 1
logger.log(f"Found NaN, decreased lg_loss_scale to {self.lg_loss_scale}")
return
model_grads_to_master_grads(self.model_params, self.master_params)
self.master_params[0].grad.mul_(1.0 / (2 ** self.lg_loss_scale))
self._log_grad_norm()
self._anneal_lr()
self.opt.step()
for rate, params in zip(self.ema_rate, self.ema_params):
update_ema(params, self.master_params, rate=rate)
master_params_to_model_params(self.model_params, self.master_params)
self.lg_loss_scale += self.fp16_scale_growth
def optimize_normal(self):
self._log_grad_norm()
self._anneal_lr()
self.opt.step()
for rate, params in zip(self.ema_rate, self.ema_params):
update_ema(params, self.master_params, rate=rate)
def _log_grad_norm(self):
sqsum = 0.0
for p in self.master_params:
sqsum += (p.grad ** 2).sum().item()
logger.logkv_mean("grad_norm", np.sqrt(sqsum))
def _anneal_lr(self):
return
def log_step(self):
logger.logkv("step", self.step + self.resume_step)
logger.logkv("samples", (self.step + self.resume_step + 1) * self.global_batch)
if self.use_fp16:
logger.logkv("lg_loss_scale", self.lg_loss_scale)
def save(self):
def save_checkpoint(rate, params):
state_dict = self._master_params_to_state_dict(params)
if dist.get_rank() == 0:
logger.log(f"saving model {rate}...")
if not rate:
filename = f"model{(self.step+self.resume_step):06d}.pt"
else:
filename = f"ema_{rate}_{(self.step+self.resume_step):06d}.pt"
with bf.BlobFile(bf.join(get_blob_logdir(), filename), "wb") as f:
th.save(state_dict, f)
save_checkpoint(0, self.master_params)
for rate, params in zip(self.ema_rate, self.ema_params):
save_checkpoint(rate, params)
if dist.get_rank() == 0:
with bf.BlobFile(
bf.join(get_blob_logdir(), f"opt{(self.step+self.resume_step):06d}.pt"),
"wb",
) as f:
th.save(self.opt.state_dict(), f)
dist.barrier()
def _master_params_to_state_dict(self, master_params):
if self.use_fp16:
master_params = unflatten_master_params(
list(self.optimize_model.parameters()), master_params
)
state_dict = self.optimize_model.state_dict()
for i, (name, _value) in enumerate(self.optimize_model.named_parameters()):
assert name in state_dict
state_dict[name] = master_params[i]
return state_dict
def _state_dict_to_master_params(self, state_dict):
params = [state_dict[name] for name, _ in self.optimize_model.named_parameters()]
if self.use_fp16:
return make_master_params(params)
else:
return params
def parse_resume_step_from_filename(filename):
"""
Parse filenames of the form path/to/modelNNNNNN.pt, where NNNNNN is the
checkpoint's number of steps.
"""
filename=filename.split('/')[-1]
assert(filename.endswith(".pt"))
filename=filename[:-3]
if filename.startswith("model"):
split = filename[5:]
elif filename.startswith("ema"):
split = filename.split("_")[-1]
else:
return 0
try:
return int(split)
except ValueError:
return 0
def get_blob_logdir():
p=os.path.join(logger.get_dir(),"checkpoints")
os.makedirs(p,exist_ok=True)
return p
def find_resume_checkpoint(resume_checkpoint):
# On your infrastructure, you may want to override this to automatically
# discover the latest checkpoint on your blob storage, etc.
if not resume_checkpoint:
return None
if "ROOT" in resume_checkpoint:
maybe_root=os.environ.get("AMLT_MAP_INPUT_DIR")
maybe_root="OUTPUT/log" if not maybe_root else maybe_root
root=os.path.join(maybe_root,"checkpoints")
resume_checkpoint=resume_checkpoint.replace("ROOT",root)
if "LATEST" in resume_checkpoint:
files=glob.glob(resume_checkpoint.replace("LATEST","*.pt"))
if not files:
return None
return max(files,key=parse_resume_step_from_filename)
return resume_checkpoint
def find_ema_checkpoint(main_checkpoint, step, rate):
if main_checkpoint is None:
return None
filename = f"ema_{rate}_{(step):06d}.pt"
path = bf.join(bf.dirname(main_checkpoint), filename)
if bf.exists(path):
return path
return None
def log_loss_dict(diffusion, ts, losses):
for key, values in losses.items():
logger.logkv_mean(key, values.mean().item())
# Log the quantiles (four quartiles, in particular).
for sub_t, sub_loss in zip(ts.cpu().numpy(), values.detach().cpu().numpy()):
quartile = int(4 * sub_t / diffusion.num_timesteps)
logger.logkv_mean(f"{key}_q{quartile}", sub_loss)