DifFace / trainer.py
Zongsheng
first upload
06f26d7
raw
history blame
38.1 kB
#!/usr/bin/env python
# -*- coding:utf-8 -*-
# Power by Zongsheng Yue 2022-05-18 13:04:06
import os
import sys
import math
import time
import lpips
import random
import datetime
import functools
import numpy as np
from pathlib import Path
from loguru import logger
from copy import deepcopy
from omegaconf import OmegaConf
from collections import OrderedDict
from einops import rearrange
from datapipe.datasets import create_dataset
from models.resample import UniformSampler
import torch
import torch.nn as nn
import torch.cuda.amp as amp
import torch.nn.functional as F
import torch.utils.data as udata
import torch.distributed as dist
import torch.multiprocessing as mp
import torchvision.utils as vutils
from torch.utils.tensorboard import SummaryWriter
from torch.nn.parallel import DistributedDataParallel as DDP
from utils import util_net
from utils import util_common
from utils import util_image
from basicsr.utils import DiffJPEG
from basicsr.utils.img_process_util import filter2D
from basicsr.data.transforms import paired_random_crop
from basicsr.data.degradations import random_add_gaussian_noise_pt, random_add_poisson_noise_pt
class TrainerBase:
def __init__(self, configs):
self.configs = configs
# setup distributed training: self.num_gpus, self.rank
self.setup_dist()
# setup seed
self.setup_seed()
# setup logger: self.logger
self.init_logger()
# logging the configurations
if self.rank == 0: self.logger.info(OmegaConf.to_yaml(self.configs))
# build model: self.model, self.loss
self.build_model()
# setup optimization: self.optimzer, self.sheduler
self.setup_optimizaton()
# resume
self.resume_from_ckpt()
def setup_dist(self):
if self.configs.gpu_id:
gpu_id = self.configs.gpu_id
num_gpus = len(gpu_id)
os.environ['CUDA_DEVICE_ORDER'] = 'PCI_BUS_ID'
os.environ['CUDA_VISIBLE_DEVICES'] = ','.join([gpu_id[ii] for ii in range(num_gpus)])
else:
num_gpus = torch.cuda.device_count()
if num_gpus > 1:
if mp.get_start_method(allow_none=True) is None:
mp.set_start_method('spawn')
rank = int(os.environ['LOCAL_RANK'])
torch.cuda.set_device(rank % num_gpus)
dist.init_process_group(
backend='nccl',
init_method='env://',
)
self.num_gpus = num_gpus
self.rank = int(os.environ['LOCAL_RANK']) if num_gpus > 1 else 0
def setup_seed(self, seed=None):
seed = self.configs.seed if seed is None else seed
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
def init_logger(self):
# only should be run on rank: 0
save_dir = Path(self.configs.save_dir)
logtxet_path = save_dir / 'training.log'
log_dir = save_dir / 'logs'
ckpt_dir = save_dir / 'ckpts'
self.ckpt_dir = ckpt_dir
if self.rank == 0:
if not save_dir.exists():
save_dir.mkdir()
else:
assert self.configs.resume, '''Please check the resume parameter. If you do not
want to resume from some checkpoint, please delete
the saving folder first.'''
# text logging
if logtxet_path.exists():
assert self.configs.resume
self.logger = logger
self.logger.remove()
self.logger.add(logtxet_path, format="{message}", mode='a')
self.logger.add(sys.stderr, format="{message}")
# tensorboard log
if not log_dir.exists():
log_dir.mkdir()
self.writer = SummaryWriter(str(log_dir))
self.log_step = {phase: 1 for phase in ['train', 'val']}
self.log_step_img = {phase: 1 for phase in ['train', 'val']}
if not ckpt_dir.exists():
ckpt_dir.mkdir()
def close_logger(self):
if self.rank == 0: self.writer.close()
def resume_from_ckpt(self):
if self.configs.resume:
if type(self.configs.resume) == bool:
ckpt_index = max([int(x.stem.split('_')[1]) for x in Path(self.ckpt_dir).glob('*.pth')])
ckpt_path = str(Path(self.ckpt_dir) / f"model_{ckpt_index}.pth")
else:
ckpt_path = self.configs.resume
assert os.path.isfile(ckpt_path)
if self.rank == 0:
self.logger.info(f"=> Loaded checkpoint {ckpt_path}")
ckpt = torch.load(ckpt_path, map_location=f"cuda:{self.rank}")
util_net.reload_model(self.model, ckpt['state_dict'])
torch.cuda.empty_cache()
# iterations
self.iters_start = ckpt['iters_start']
# learning rate scheduler
for ii in range(self.iters_start): self.adjust_lr(ii)
if self.rank == 0:
self.log_step = ckpt['log_step']
self.log_step_img = ckpt['log_step_img']
# reset the seed
self.setup_seed(self.iters_start)
else:
self.iters_start = 0
def setup_optimizaton(self):
self.optimizer = torch.optim.AdamW(self.model.parameters(),
lr=self.configs.train.lr,
weight_decay=self.configs.train.weight_decay)
def build_model(self):
params = self.configs.model.get('params', dict)
model = util_common.get_obj_from_str(self.configs.model.target)(**params)
if self.num_gpus > 1:
self.model = DDP(model.cuda(), device_ids=[self.rank,]) # wrap the network
else:
self.model = model.cuda()
# LPIPS metric
if self.rank == 0:
self.lpips_loss = lpips.LPIPS(net='vgg').cuda()
# model information
self.print_model_info()
def build_dataloader(self):
def _wrap_loader(loader):
while True: yield from loader
datasets = {}
for phase in ['train', ]:
dataset_config = self.configs.data.get(phase, dict)
datasets[phase] = create_dataset(dataset_config)
dataloaders = {}
# train dataloader
if self.rank == 0:
for phase in ['train',]:
length = len(datasets[phase])
self.logger.info('Number of images in {:s} data set: {:d}'.format(phase, length))
if self.num_gpus > 1:
shuffle = False
sampler = udata.distributed.DistributedSampler(datasets['train'],
num_replicas=self.num_gpus,
rank=self.rank)
else:
shuffle = True
sampler = None
dataloaders['train'] = _wrap_loader(udata.DataLoader(
datasets['train'],
batch_size=self.configs.train.batch[0] // self.num_gpus,
shuffle=shuffle,
drop_last=False,
num_workers=self.configs.train.num_workers // self.num_gpus,
pin_memory=True,
prefetch_factor=self.configs.train.prefetch_factor,
worker_init_fn=my_worker_init_fn,
sampler=sampler))
self.datasets = datasets
self.dataloaders = dataloaders
self.sampler = sampler
def print_model_info(self):
if self.rank == 0:
num_params = util_net.calculate_parameters(self.model) / 1000**2
self.logger.info("Detailed network architecture:")
self.logger.info(self.model.__repr__())
self.logger.info(f"Number of parameters: {num_params:.2f}M")
def prepare_data(self, phase='train'):
pass
def validation(self):
pass
def train(self):
self.build_dataloader() # prepare data: self.dataloaders, self.datasets, self.sampler
self.model.train()
num_iters_epoch = math.ceil(len(self.datasets['train']) / self.configs.train.batch[0])
for ii in range(self.iters_start, self.configs.train.iterations):
self.current_iters = ii + 1
# prepare data
data = self.prepare_data(
next(self.dataloaders['train']),
self.configs.data.train.type.lower() == 'realesrgan',
)
# training phase
self.training_step(data)
# validation phase
if (ii+1) % self.configs.train.val_freq == 0 and 'val' in self.dataloaders:
if self.rank==0:
self.validation()
#update learning rate
self.adjust_lr()
# save checkpoint
if (ii+1) % self.configs.train.save_freq == 0 and self.rank == 0:
self.save_ckpt()
if (ii+1) % num_iters_epoch == 0 and not self.sampler is None:
self.sampler.set_epoch(ii+1)
# close the tensorboard
if self.rank == 0:
self.close_logger()
def training_step(self, data):
pass
def adjust_lr(self):
if hasattr(self, 'lr_sheduler'):
self.lr_sheduler.step()
def save_ckpt(self):
ckpt_path = self.ckpt_dir / 'model_{:d}.pth'.format(self.current_iters)
torch.save({'iters_start': self.current_iters,
'log_step': {phase:self.log_step[phase] for phase in ['train', 'val']},
'log_step_img': {phase:self.log_step_img[phase] for phase in ['train', 'val']},
'state_dict': self.model.state_dict()}, ckpt_path)
class TrainerSR(TrainerBase):
def __init__(self, configs):
super().__init__(configs)
def mse_loss(self, pred, target):
return F.mse_loss(pred, target, reduction='mean')
@torch.no_grad()
def _dequeue_and_enqueue(self):
"""It is the training pair pool for increasing the diversity in a batch.
Batch processing limits the diversity of synthetic degradations in a batch. For example, samples in a
batch could not have different resize scaling factors. Therefore, we employ this training pair pool
to increase the degradation diversity in a batch.
"""
# initialize
b, c, h, w = self.lq.size()
if not hasattr(self, 'queue_size'):
self.queue_size = self.configs.data.train.params.get('queue_size', b*50)
if not hasattr(self, 'queue_lr'):
assert self.queue_size % b == 0, f'queue size {self.queue_size} should be divisible by batch size {b}'
self.queue_lr = torch.zeros(self.queue_size, c, h, w).cuda()
_, c, h, w = self.gt.size()
self.queue_gt = torch.zeros(self.queue_size, c, h, w).cuda()
self.queue_ptr = 0
if self.queue_ptr == self.queue_size: # the pool is full
# do dequeue and enqueue
# shuffle
idx = torch.randperm(self.queue_size)
self.queue_lr = self.queue_lr[idx]
self.queue_gt = self.queue_gt[idx]
# get first b samples
lq_dequeue = self.queue_lr[0:b, :, :, :].clone()
gt_dequeue = self.queue_gt[0:b, :, :, :].clone()
# update the queue
self.queue_lr[0:b, :, :, :] = self.lq.clone()
self.queue_gt[0:b, :, :, :] = self.gt.clone()
self.lq = lq_dequeue
self.gt = gt_dequeue
else:
# only do enqueue
self.queue_lr[self.queue_ptr:self.queue_ptr + b, :, :, :] = self.lq.clone()
self.queue_gt[self.queue_ptr:self.queue_ptr + b, :, :, :] = self.gt.clone()
self.queue_ptr = self.queue_ptr + b
@torch.no_grad()
def prepare_data(self, data, real_esrgan=True):
if real_esrgan:
if not hasattr(self, 'jpeger'):
self.jpeger = DiffJPEG(differentiable=False).cuda() # simulate JPEG compression artifacts
im_gt = data['gt'].cuda()
kernel1 = data['kernel1'].cuda()
kernel2 = data['kernel2'].cuda()
sinc_kernel = data['sinc_kernel'].cuda()
ori_h, ori_w = im_gt.size()[2:4]
# ----------------------- The first degradation process ----------------------- #
# blur
out = filter2D(im_gt, kernel1)
# random resize
updown_type = random.choices(
['up', 'down', 'keep'],
self.configs.degradation['resize_prob'],
)[0]
if updown_type == 'up':
scale = random.uniform(1, self.configs.degradation['resize_range'][1])
elif updown_type == 'down':
scale = random.uniform(self.configs.degradation['resize_range'][0], 1)
else:
scale = 1
mode = random.choice(['area', 'bilinear', 'bicubic'])
out = F.interpolate(out, scale_factor=scale, mode=mode)
# add noise
gray_noise_prob = self.configs.degradation['gray_noise_prob']
if random.random() < self.configs.degradation['gaussian_noise_prob']:
out = random_add_gaussian_noise_pt(
out,
sigma_range=self.configs.degradation['noise_range'],
clip=True,
rounds=False,
gray_prob=gray_noise_prob,
)
else:
out = random_add_poisson_noise_pt(
out,
scale_range=self.configs.degradation['poisson_scale_range'],
gray_prob=gray_noise_prob,
clip=True,
rounds=False)
# JPEG compression
jpeg_p = out.new_zeros(out.size(0)).uniform_(*self.configs.degradation['jpeg_range'])
out = torch.clamp(out, 0, 1) # clamp to [0, 1], otherwise JPEGer will result in unpleasant artifacts
out = self.jpeger(out, quality=jpeg_p)
# ----------------------- The second degradation process ----------------------- #
# blur
if random.random() < self.configs.degradation['second_blur_prob']:
out = filter2D(out, kernel2)
# random resize
updown_type = random.choices(
['up', 'down', 'keep'],
self.configs.degradation['resize_prob2'],
)[0]
if updown_type == 'up':
scale = random.uniform(1, self.configs.degradation['resize_range2'][1])
elif updown_type == 'down':
scale = random.uniform(self.configs.degradation['resize_range2'][0], 1)
else:
scale = 1
mode = random.choice(['area', 'bilinear', 'bicubic'])
out = F.interpolate(
out,
size=(int(ori_h / self.configs.model.params.sf * scale),
int(ori_w / self.configs.model.params.sf * scale)),
mode=mode,
)
# add noise
gray_noise_prob = self.configs.degradation['gray_noise_prob2']
if random.random() < self.configs.degradation['gaussian_noise_prob2']:
out = random_add_gaussian_noise_pt(
out,
sigma_range=self.configs.degradation['noise_range2'],
clip=True,
rounds=False,
gray_prob=gray_noise_prob,
)
else:
out = random_add_poisson_noise_pt(
out,
scale_range=self.configs.degradation['poisson_scale_range2'],
gray_prob=gray_noise_prob,
clip=True,
rounds=False,
)
# JPEG compression + the final sinc filter
# We also need to resize images to desired sizes. We group [resize back + sinc filter] together
# as one operation.
# We consider two orders:
# 1. [resize back + sinc filter] + JPEG compression
# 2. JPEG compression + [resize back + sinc filter]
# Empirically, we find other combinations (sinc + JPEG + Resize) will introduce twisted lines.
if random.random() < 0.5:
# resize back + the final sinc filter
mode = random.choice(['area', 'bilinear', 'bicubic'])
out = F.interpolate(
out,
size=(ori_h // self.configs.model.params.sf,
ori_w // self.configs.model.params.sf),
mode=mode,
)
out = filter2D(out, sinc_kernel)
# JPEG compression
jpeg_p = out.new_zeros(out.size(0)).uniform_(*self.configs.degradation['jpeg_range2'])
out = torch.clamp(out, 0, 1)
out = self.jpeger(out, quality=jpeg_p)
else:
# JPEG compression
jpeg_p = out.new_zeros(out.size(0)).uniform_(*self.configs.degradation['jpeg_range2'])
out = torch.clamp(out, 0, 1)
out = self.jpeger(out, quality=jpeg_p)
# resize back + the final sinc filter
mode = random.choice(['area', 'bilinear', 'bicubic'])
out = F.interpolate(
out,
size=(ori_h // self.configs.model.params.sf,
ori_w // self.configs.model.params.sf),
mode=mode,
)
out = filter2D(out, sinc_kernel)
# clamp and round
im_lq = torch.clamp((out * 255.0).round(), 0, 255) / 255.
# random crop
gt_size = self.configs.degradation['gt_size']
im_gt, im_lq = paired_random_crop(im_gt, im_lq, gt_size, self.configs.model.params.sf)
self.lq, self.gt = im_lq, im_gt
# training pair pool
self._dequeue_and_enqueue()
# sharpen self.gt again, as we have changed the self.gt with self._dequeue_and_enqueue
self.lq = self.lq.contiguous() # for the warning: grad and param do not obey the gradient layout contract
return {'lq':self.lq, 'gt':self.gt}
else:
return {key:value.cuda() for key, value in data.items()}
def setup_optimizaton(self):
super().setup_optimizaton() # self.optimizer
self.lr_sheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
self.optimizer,
T_max = self.configs.train.iterations,
eta_min=self.configs.train.lr_min,
)
def training_step(self, data):
current_batchsize = data['lq'].shape[0]
micro_batchsize = self.configs.train.microbatch
num_grad_accumulate = math.ceil(current_batchsize / micro_batchsize)
self.optimizer.zero_grad()
for jj in range(0, current_batchsize, micro_batchsize):
micro_data = {key:value[jj:jj+micro_batchsize,] for key, value in data.items()}
last_batch = (jj+micro_batchsize >= current_batchsize)
hq_pred = self.model(micro_data['lq'])
if last_batch or self.num_gpus <= 1:
loss = self.loss_fun(hq_pred, micro_data['gt']) / hq_pred.shape[0]
else:
with self.model.no_sync():
loss = self.loss_fun(hq_pred, micro_data['gt']) / hq_pred.shape[0]
loss /= num_grad_accumulate
loss.backward()
# make logging
self.log_step_train(hq_pred, loss, micro_data, flag=last_batch)
self.optimizer.step()
def log_step_train(self, hq_pred, loss, batch, flag=False, phase='train'):
'''
param loss: loss value
'''
if self.rank == 0:
chn = batch['lq'].shape[1]
if self.current_iters % self.configs.train.log_freq[0] == 1:
self.loss_mean = 0
self.loss_mean += loss.item()
if self.current_iters % self.configs.train.log_freq[0] == 0 and flag:
self.loss_mean /= self.configs.train.log_freq[0]
mse_pixel = self.loss_mean / batch['gt'].numel() * batch['gt'].shape[0]
log_str = 'Train:{:05d}/{:05d}, Loss:{:.2e}, MSE:{:.2e}, lr:{:.2e}'.format(
self.current_iters // 100,
self.configs.train.iterations // 100,
self.loss_mean,
mse_pixel,
self.optimizer.param_groups[0]['lr']
)
self.logger.info(log_str)
# tensorboard
self.writer.add_scalar(f'Loss-Train', self.loss_mean, self.log_step[phase])
self.log_step[phase] += 1
if self.current_iters % self.configs.train.log_freq[1] == 0 and flag:
x1 = vutils.make_grid(batch['lq'], normalize=True, scale_each=True)
self.writer.add_image("Train LQ Image", x1, self.log_step_img[phase])
x2 = vutils.make_grid(batch['gt'], normalize=True, scale_each=True)
self.writer.add_image("Train HQ Image", x2, self.log_step_img[phase])
x3 = vutils.make_grid(hq_pred.detach().data, normalize=True, scale_each=True)
self.writer.add_image("Train Recovered Image", x3, self.log_step_img[phase])
self.log_step_img[phase] += 1
if self.current_iters % self.configs.train.save_freq == 1 and flag:
self.tic = time.time()
if self.current_iters % self.configs.train.save_freq == 0 and flag:
self.toc = time.time()
elaplsed = (self.toc - self.tic)
self.logger.info(f"Elapsed time: {elaplsed:.2f}s")
self.logger.info("="*60)
def validation(self, phase='val'):
if self.rank == 0:
self.model.eval()
psnr_mean = lpips_mean = 0
total_iters = math.ceil(len(self.datasets[phase]) / self.configs.train.batch[1])
for ii, data in enumerate(self.dataloaders[phase]):
data = self.prepare_data(data)
with torch.no_grad():
hq_pred = self.model(data['lq'])
hq_pred.clamp_(0.0, 1.0)
lpips = self.lpips_loss(
util_image.normalize_th(hq_pred, reverse=False),
util_image.normalize_th(data['gt'], reverse=False),
).sum().item()
psnr = util_image.batch_PSNR(
hq_pred,
data['gt'],
ycbcr=True
)
psnr_mean += psnr
lpips_mean += lpips
if (ii+1) % self.configs.train.log_freq[2] == 0:
log_str = '{:s}:{:03d}/{:03d}, PSNR={:5.2f}, LPIPS={:6.4f}'.format(
phase,
ii+1,
total_iters,
psnr / hq_pred.shape[0],
lpips / hq_pred.shape[0]
)
self.logger.info(log_str)
x1 = vutils.make_grid(data['lq'], normalize=True, scale_each=True)
self.writer.add_image("Validation LQ Image", x1, self.log_step_img[phase])
x2 = vutils.make_grid(data['gt'], normalize=True, scale_each=True)
self.writer.add_image("Validation HQ Image", x2, self.log_step_img[phase])
x3 = vutils.make_grid(hq_pred.detach().data, normalize=True, scale_each=True)
self.writer.add_image("Validation Recovered Image", x3, self.log_step_img[phase])
self.log_step_img[phase] += 1
psnr_mean /= len(self.datasets[phase])
lpips_mean /= len(self.datasets[phase])
# tensorboard
self.writer.add_scalar('Validation PSRN', psnr_mean, self.log_step[phase])
self.writer.add_scalar('Validation LPIPS', lpips_mean, self.log_step[phase])
self.log_step[phase] += 1
# logging
self.logger.info(f'PSNR={psnr_mean:5.2f}, LPIPS={lpips_mean:6.4f}')
self.logger.info("="*60)
self.model.train()
def build_dataloader(self):
super().build_dataloader()
if self.rank == 0 and 'val' in self.configs.data:
dataset_config = self.configs.data.get('val', dict)
self.datasets['val'] = create_dataset(dataset_config)
self.dataloaders['val'] = udata.DataLoader(
self.datasets['val'],
batch_size=self.configs.train.batch[1],
shuffle=False,
drop_last=False,
num_workers=0,
pin_memory=True,
)
class TrainerDiffusionFace(TrainerBase):
def __init__(self, configs):
# ema settings
self.ema_rates = OmegaConf.to_object(configs.train.ema_rates)
super().__init__(configs)
def init_logger(self):
super().init_logger()
save_dir = Path(self.configs.save_dir)
ema_ckpt_dir = save_dir / 'ema_ckpts'
if self.rank == 0:
if not ema_ckpt_dir.exists():
util_common.mkdir(ema_ckpt_dir, delete=False, parents=False)
else:
if not self.configs.resume:
util_common.mkdir(ema_ckpt_dir, delete=True, parents=False)
self.ema_ckpt_dir = ema_ckpt_dir
def resume_from_ckpt(self):
super().resume_from_ckpt()
def _load_ema_state(ema_state, ckpt):
for key in ema_state.keys():
ema_state[key] = deepcopy(ckpt[key].detach().data)
if self.configs.resume:
# ema model
if type(self.configs.resume) == bool:
ckpt_index = max([int(x.stem.split('_')[1]) for x in Path(self.ckpt_dir).glob('*.pth')])
ckpt_path = str(Path(self.ckpt_dir) / f"model_{ckpt_index}.pth")
else:
ckpt_path = self.configs.resume
assert os.path.isfile(ckpt_path)
# EMA model
for rate in self.ema_rates:
ema_ckpt_path = self.ema_ckpt_dir / (f"ema0{int(rate*1000)}_"+Path(ckpt_path).name)
ema_ckpt = torch.load(ema_ckpt_path, map_location=f"cuda:{self.rank}")
_load_ema_state(self.ema_state[f"0{int(rate*1000)}"], ema_ckpt)
def build_model(self):
params = self.configs.model.get('params', dict)
model = util_common.get_obj_from_str(self.configs.model.target)(**params)
self.ema_model = deepcopy(model.cuda())
if self.num_gpus > 1:
self.model = DDP(model.cuda(), device_ids=[self.rank,]) # wrap the network
else:
self.model = model.cuda()
self.ema_state = {}
for rate in self.ema_rates:
self.ema_state[f"0{int(rate*1000)}"] = OrderedDict(
{key:deepcopy(value.data) for key, value in self.model.state_dict().items()}
)
# model information
self.print_model_info()
params = self.configs.diffusion.get('params', dict)
self.base_diffusion = util_common.get_obj_from_str(self.configs.diffusion.target)(**params)
self.sample_scheduler_diffusion = UniformSampler(self.base_diffusion.num_timesteps)
def prepare_data(self, data, realesrgan=False):
data = {key:value.cuda() for key, value in data.items()}
return data
def training_step(self, data):
current_batchsize = data['image'].shape[0]
micro_batchsize = self.configs.train.microbatch
num_grad_accumulate = math.ceil(current_batchsize / micro_batchsize)
if self.configs.train.use_fp16:
scaler = amp.GradScaler()
self.optimizer.zero_grad()
for jj in range(0, current_batchsize, micro_batchsize):
micro_data = {key:value[jj:jj+micro_batchsize,] for key, value in data.items()}
last_batch = (jj+micro_batchsize >= current_batchsize)
tt, weights = self.sample_scheduler_diffusion.sample(
micro_data['image'].shape[0],
device=f"cuda:{self.rank}",
use_fp16=self.configs.train.use_fp16
)
compute_losses = functools.partial(
self.base_diffusion.training_losses,
self.model,
micro_data['image'],
tt,
model_kwargs={'y':micro_data['label']} if 'label' in micro_data else None,
)
if self.configs.train.use_fp16:
with amp.autocast():
if last_batch or self.num_gpus <= 1:
losses = compute_losses()
else:
with self.model.no_sync():
losses = compute_losses()
loss = (losses["loss"] * weights).mean() / num_grad_accumulate
scaler.scale(loss).backward()
else:
if last_batch or self.num_gpus <= 1:
losses = compute_losses()
else:
with self.model.no_sync():
losses = compute_losses()
loss = (losses["loss"] * weights).mean() / num_grad_accumulate
loss.backward()
# make logging
self.log_step_train(losses, tt, micro_data, last_batch)
if self.configs.train.use_fp16:
scaler.step(self.optimizer)
scaler.update()
else:
self.optimizer.step()
self.update_ema_model()
def update_ema_model(self):
if self.num_gpus > 1:
dist.barrier()
if self.rank == 0:
for rate in self.ema_rates:
ema_state = self.ema_state[f"0{int(rate*1000)}"]
source_state = self.model.state_dict()
for key, value in ema_state.items():
ema_state[key].mul_(rate).add_(source_state[key].detach().data, alpha=1-rate)
def adjust_lr(self, ii):
base_lr = self.configs.train.lr
linear_steps = self.configs.train.milestones[0]
if ii <= linear_steps:
for params_group in self.optimizer.param_groups:
params_group['lr'] = (ii / linear_steps) * base_lr
elif ii in self.configs.train.milestones:
for params_group in self.optimizer.param_groups:
params_group['lr'] *= 0.5
def log_step_train(self, loss, tt, batch, flag=False, phase='train'):
'''
param loss: a dict recording the loss informations
param tt: 1-D tensor, time steps
'''
if self.rank == 0:
chn = batch['image'].shape[1]
num_timesteps = self.base_diffusion.num_timesteps
if self.current_iters % self.configs.train.log_freq[0] == 1:
self.loss_mean = {key:torch.zeros(size=(num_timesteps,), dtype=torch.float64)
for key in loss.keys()}
self.loss_count = torch.zeros(size=(num_timesteps,), dtype=torch.float64)
for key, value in loss.items():
self.loss_mean[key][tt, ] += value.detach().data.cpu()
self.loss_count[tt,] += 1
if self.current_iters % self.configs.train.log_freq[0] == 0 and flag:
if torch.any(self.loss_count == 0):
self.loss_count += 1e-4
for key, value in loss.items():
self.loss_mean[key] /= self.loss_count
log_str = 'Train: {:05d}/{:05d}, Loss: '.format(
self.current_iters // 100,
self.configs.train.iterations // 100)
for kk in [1, num_timesteps // 2, num_timesteps]:
if 'vb' in self.loss_mean:
log_str += 't({:d}):{:.2e}/{:.2e}/{:.2e}, '.format(
kk,
self.loss_mean['loss'][kk-1].item(),
self.loss_mean['mse'][kk-1].item(),
self.loss_mean['vb'][kk-1].item(),
)
else:
log_str += 't({:d}):{:.2e}, '.format(kk, self.loss_mean['loss'][kk-1].item())
log_str += 'lr:{:.2e}'.format(self.optimizer.param_groups[0]['lr'])
self.logger.info(log_str)
# tensorboard
for kk in [1, num_timesteps // 2, num_timesteps]:
self.writer.add_scalar(f'Loss-Step-{kk}',
self.loss_mean['loss'][kk-1].item(),
self.log_step[phase])
self.log_step[phase] += 1
if self.current_iters % self.configs.train.log_freq[1] == 0 and flag:
x1 = vutils.make_grid(batch['image'], normalize=True, scale_each=True)
self.writer.add_image("Training Image", x1, self.log_step_img[phase])
self.log_step_img[phase] += 1
if self.current_iters % self.configs.train.save_freq == 1 and flag:
self.tic = time.time()
if self.current_iters % self.configs.train.save_freq == 0 and flag:
self.toc = time.time()
elaplsed = (self.toc - self.tic) * num_timesteps / (num_timesteps - 1)
self.logger.info(f"Elapsed time: {elaplsed:.2f}s")
self.logger.info("="*130)
def validation(self, phase='val'):
self.reload_ema_model(self.ema_rates[0])
self.ema_model.eval()
indices = [int(self.base_diffusion.num_timesteps * x) for x in [0.25, 0.5, 0.75, 1]]
chn = 3
batch_size = self.configs.train.batch[1]
shape = (batch_size, chn,) + (self.configs.data.train.params.out_size,) * 2
num_iters = 0
# noise = torch.randn(shape,
# dtype=torch.float32,
# generator=torch.Generator('cpu').manual_seed(10000)).cuda()
for sample in self.base_diffusion.p_sample_loop_progressive(
model = self.ema_model,
shape = shape,
noise = None,
clip_denoised = True,
model_kwargs = None,
device = f"cuda:{self.rank}",
progress=False
):
num_iters += 1
img = util_image.normalize_th(sample['sample'], reverse=True)
if num_iters == 1:
im_recover = img
elif num_iters in indices:
im_recover_last = img
im_recover = torch.cat((im_recover, im_recover_last), dim=1)
im_recover = rearrange(im_recover, 'b (k c) h w -> (b k) c h w', c=chn)
x1 = vutils.make_grid(im_recover, nrow=len(indices)+1, normalize=False)
self.writer.add_image('Validation Sample', x1, self.log_step_img[phase])
self.log_step_img[phase] += 1
def save_ckpt(self):
if self.rank == 0:
ckpt_path = self.ckpt_dir / 'model_{:d}.pth'.format(self.current_iters)
torch.save({'iters_start': self.current_iters,
'log_step': {phase:self.log_step[phase] for phase in ['train', 'val']},
'log_step_img': {phase:self.log_step_img[phase] for phase in ['train', 'val']},
'state_dict': self.model.state_dict()}, ckpt_path)
for rate in self.ema_rates:
ema_ckpt_path = self.ema_ckpt_dir / (f"ema0{int(rate*1000)}_"+ckpt_path.name)
torch.save(self.ema_state[f"0{int(rate*1000)}"], ema_ckpt_path)
def calculate_lpips(self, inputs, targets):
inputs, targets = [(x-0.5)/0.5 for x in [inputs, targets]] # [-1, 1]
with torch.no_grad():
mean_lpips = self.lpips_loss(inputs, targets)
return mean_lpips.mean().item()
def reload_ema_model(self, rate):
model_state = {key[7:]:value for key, value in self.ema_state[f"0{int(rate*1000)}"].items()}
self.ema_model.load_state_dict(model_state)
def my_worker_init_fn(worker_id):
np.random.seed(np.random.get_state()[1][0] + worker_id)
if __name__ == '__main__':
from utils import util_image
from einops import rearrange
im1 = util_image.imread('./testdata/inpainting/val/places/Places365_val_00012685_crop000.png',
chn = 'rgb', dtype='float32')
im2 = util_image.imread('./testdata/inpainting/val/places/Places365_val_00014886_crop000.png',
chn = 'rgb', dtype='float32')
im = rearrange(np.stack((im1, im2), 3), 'h w c b -> b c h w')
im_grid = im.copy()
for alpha in [0.8, 0.4, 0.1, 0]:
im_new = im * alpha + np.random.randn(*im.shape) * (1 - alpha)
im_grid = np.concatenate((im_new, im_grid), 1)
im_grid = np.clip(im_grid, 0.0, 1.0)
im_grid = rearrange(im_grid, 'b (k c) h w -> (b k) c h w', k=5)
xx = vutils.make_grid(torch.from_numpy(im_grid), nrow=5, normalize=True, scale_each=True).numpy()
util_image.imshow(np.concatenate((im1, im2), 0))
util_image.imshow(xx.transpose((1,2,0)))