resshift / trainer.py
yuhj95's picture
Upload folder using huggingface_hub
4730cdc verified
#!/usr/bin/env python
# -*- coding:utf-8 -*-
# Power by Zongsheng Yue 2022-05-18 13:04:06
import os, sys, math, time, random, datetime, functools
import lpips
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 contextlib import nullcontext
from datapipe.datasets import create_dataset
from utils import util_net
from utils import util_common
from utils import util_image
from basicsr.utils import DiffJPEG, USMSharp
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
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
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()
def setup_dist(self):
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(
timeout=datetime.timedelta(seconds=3600),
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, global_seeding=None):
if seed is None:
seed = self.configs.train.get('seed', 12345)
if global_seeding is None:
global_seeding = self.configs.train.global_seeding
assert isinstance(global_seeding, bool)
if not global_seeding:
seed += self.rank
torch.cuda.manual_seed(seed)
else:
torch.cuda.manual_seed_all(seed)
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
def init_logger(self):
if self.configs.resume:
assert self.configs.resume.endswith(".pth")
save_dir = Path(self.configs.resume).parents[1]
project_id = save_dir.name
else:
project_id = datetime.datetime.now().strftime("%Y-%m-%d-%H-%M")
save_dir = Path(self.configs.save_dir) / project_id
if not save_dir.exists() and self.rank == 0:
save_dir.mkdir(parents=True)
# setting log counter
if self.rank == 0:
self.log_step = {phase: 1 for phase in ['train', 'val']}
self.log_step_img = {phase: 1 for phase in ['train', 'val']}
# text logging
logtxet_path = save_dir / 'training.log'
if self.rank == 0:
if logtxet_path.exists():
assert self.configs.resume
self.logger = logger
self.logger.remove()
self.logger.add(logtxet_path, format="{message}", mode='a', level='INFO')
self.logger.add(sys.stdout, format="{message}")
# tensorboard logging
log_dir = save_dir / 'tf_logs'
self.tf_logging = self.configs.train.tf_logging
if self.rank == 0 and self.tf_logging:
if not log_dir.exists():
log_dir.mkdir()
self.writer = SummaryWriter(str(log_dir))
# checkpoint saving
ckpt_dir = save_dir / 'ckpts'
self.ckpt_dir = ckpt_dir
if self.rank == 0 and (not ckpt_dir.exists()):
ckpt_dir.mkdir()
if 'ema_rate' in self.configs.train:
self.ema_rate = self.configs.train.ema_rate
assert isinstance(self.ema_rate, float), "Ema rate must be a float number"
ema_ckpt_dir = save_dir / 'ema_ckpts'
self.ema_ckpt_dir = ema_ckpt_dir
if self.rank == 0 and (not ema_ckpt_dir.exists()):
ema_ckpt_dir.mkdir()
# save images into local disk
self.local_logging = self.configs.train.local_logging
if self.rank == 0 and self.local_logging:
image_dir = save_dir / 'images'
if not image_dir.exists():
(image_dir / 'train').mkdir(parents=True)
(image_dir / 'val').mkdir(parents=True)
self.image_dir = image_dir
# logging the configurations
if self.rank == 0:
self.logger.info(OmegaConf.to_yaml(self.configs))
def close_logger(self):
if self.rank == 0 and self.tf_logging:
self.writer.close()
def resume_from_ckpt(self):
def _load_ema_state(ema_state, ckpt):
for key in ema_state.keys():
if key not in ckpt and key.startswith('module'):
ema_state[key] = deepcopy(ckpt[7:].detach().data)
elif key not in ckpt and (not key.startswith('module')):
ema_state[key] = deepcopy(ckpt['module.'+key].detach().data)
else:
ema_state[key] = deepcopy(ckpt[key].detach().data)
if self.configs.resume:
assert self.configs.resume.endswith(".pth") and os.path.isfile(self.configs.resume)
if self.rank == 0:
self.logger.info(f"=> Loaded checkpoint from {self.configs.resume}")
ckpt = torch.load(self.configs.resume, map_location=f"cuda:{self.rank}")
util_net.reload_model(self.model, ckpt['state_dict'])
torch.cuda.empty_cache()
# learning rate scheduler
self.iters_start = ckpt['iters_start']
for ii in range(1, self.iters_start+1):
self.adjust_lr(ii)
# logging
if self.rank == 0:
self.log_step = ckpt['log_step']
self.log_step_img = ckpt['log_step_img']
# EMA model
if self.rank == 0 and hasattr(self, 'ema_rate'):
ema_ckpt_path = self.ema_ckpt_dir / ("ema_"+Path(self.configs.resume).name)
self.logger.info(f"=> Loaded EMA checkpoint from {str(ema_ckpt_path)}")
ema_ckpt = torch.load(ema_ckpt_path, map_location=f"cuda:{self.rank}")
_load_ema_state(self.ema_state, ema_ckpt)
torch.cuda.empty_cache()
# AMP scaler
if self.amp_scaler is not None:
if "amp_scaler" in ckpt:
self.amp_scaler.load_state_dict(ckpt["amp_scaler"])
if self.rank == 0:
self.logger.info("Loading scaler from resumed state...")
# reset the seed
self.setup_seed(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)
# amp settings
self.amp_scaler = amp.GradScaler() if self.configs.train.use_amp else None
def build_model(self):
params = self.configs.model.get('params', dict)
model = util_common.get_obj_from_str(self.configs.model.target)(**params)
model.cuda()
if self.configs.model.ckpt_path is not None:
ckpt_path = self.configs.model.ckpt_path
if self.rank == 0:
self.logger.info(f"Initializing model from {ckpt_path}")
ckpt = torch.load(ckpt_path, map_location=f"cuda:{self.rank}")
if 'state_dict' in ckpt:
ckpt = ckpt['state_dict']
util_net.reload_model(model, ckpt)
if self.configs.train.compile.flag:
if self.rank == 0:
self.logger.info("Begin compiling model...")
model = torch.compile(model, mode=self.configs.train.compile.mode)
if self.rank == 0:
self.logger.info("Compiling Done")
if self.num_gpus > 1:
self.model = DDP(model, device_ids=[self.rank,], static_graph=False) # wrap the network
else:
self.model = model
# EMA
if self.rank == 0 and hasattr(self.configs.train, 'ema_rate'):
self.ema_model = deepcopy(model).cuda()
self.ema_state = OrderedDict(
{key:deepcopy(value.data) for key, value in self.model.state_dict().items()}
)
self.ema_ignore_keys = [x for x in self.ema_state.keys() if ('running_' in x or 'num_batches_tracked' in x)]
# model information
self.print_model_info()
def build_dataloader(self):
def _wrap_loader(loader):
while True: yield from loader
# make datasets
datasets = {'train': create_dataset(self.configs.data.get('train', dict)), }
if hasattr(self.configs.data, 'val') and self.rank == 0:
datasets['val'] = create_dataset(self.configs.data.get('val', dict))
if self.rank == 0:
for phase in datasets.keys():
length = len(datasets[phase])
self.logger.info('Number of images in {:s} data set: {:d}'.format(phase, length))
# make dataloaders
if self.num_gpus > 1:
sampler = udata.distributed.DistributedSampler(
datasets['train'],
num_replicas=self.num_gpus,
rank=self.rank,
)
else:
sampler = None
dataloaders = {'train': _wrap_loader(udata.DataLoader(
datasets['train'],
batch_size=self.configs.train.batch[0] // self.num_gpus,
shuffle=False if self.num_gpus > 1 else True,
drop_last=True,
num_workers=min(self.configs.train.num_workers, 4),
pin_memory=True,
prefetch_factor=self.configs.train.get('prefetch_factor', 2),
worker_init_fn=my_worker_init_fn,
sampler=sampler,
))}
if hasattr(self.configs.data, 'val') and self.rank == 0:
dataloaders['val'] = udata.DataLoader(datasets['val'],
batch_size=self.configs.train.batch[1],
shuffle=False,
drop_last=False,
num_workers=0,
pin_memory=True,
)
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, data, dtype=torch.float32, phase='train'):
data = {key:value.cuda().to(dtype=dtype) for key, value in data.items()}
return data
def validation(self):
pass
def train(self):
self.init_logger() # setup logger: self.logger
self.build_model() # build model: self.model, self.loss
self.setup_optimizaton() # setup optimization: self.optimzer, self.sheduler
self.resume_from_ckpt() # resume if necessary
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']))
# training phase
self.training_step(data)
# validation phase
if 'val' in self.dataloaders and (ii+1) % self.configs.train.get('val_freq', 10000) == 0:
self.validation()
#update learning rate
self.adjust_lr()
# save checkpoint
if (ii+1) % self.configs.train.save_freq == 0:
self.save_ckpt()
if (ii+1) % num_iters_epoch == 0 and self.sampler is not None:
self.sampler.set_epoch(ii+1)
# close the tensorboard
self.close_logger()
def training_step(self, data):
pass
def adjust_lr(self, current_iters=None):
assert hasattr(self, 'lr_scheduler')
self.lr_scheduler.step()
def save_ckpt(self):
if self.rank == 0:
ckpt_path = self.ckpt_dir / 'model_{:d}.pth'.format(self.current_iters)
ckpt = {
'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(),
}
if self.amp_scaler is not None:
ckpt['amp_scaler'] = self.amp_scaler.state_dict()
torch.save(ckpt, ckpt_path)
if hasattr(self, 'ema_rate'):
ema_ckpt_path = self.ema_ckpt_dir / 'ema_model_{:d}.pth'.format(self.current_iters)
torch.save(self.ema_state, ema_ckpt_path)
def reload_ema_model(self):
if self.rank == 0:
if self.num_gpus > 1:
model_state = {key[7:]:value for key, value in self.ema_state.items()}
else:
model_state = self.ema_state
self.ema_model.load_state_dict(model_state)
@torch.no_grad()
def update_ema_model(self):
if self.num_gpus > 1:
dist.barrier()
if self.rank == 0:
source_state = self.model.state_dict()
rate = self.ema_rate
for key, value in self.ema_state.items():
if key in self.ema_ignore_keys:
self.ema_state[key] = source_state[key]
else:
self.ema_state[key].mul_(rate).add_(source_state[key].detach().data, alpha=1-rate)
def logging_image(self, im_tensor, tag, phase, add_global_step=False, nrow=8):
"""
Args:
im_tensor: b x c x h x w tensor
im_tag: str
phase: 'train' or 'val'
nrow: number of displays in each row
"""
assert self.tf_logging or self.local_logging
im_tensor = vutils.make_grid(im_tensor, nrow=nrow, normalize=True, scale_each=True) # c x H x W
if self.local_logging:
im_path = str(self.image_dir / phase / f"{tag}-{self.log_step_img[phase]}.png")
im_np = im_tensor.cpu().permute(1,2,0).numpy()
util_image.imwrite(im_np, im_path)
if self.tf_logging:
self.writer.add_image(
f"{phase}-{tag}-{self.log_step_img[phase]}",
im_tensor,
self.log_step_img[phase],
)
if add_global_step:
self.log_step_img[phase] += 1
def logging_metric(self, metrics, tag, phase, add_global_step=False):
"""
Args:
metrics: dict
tag: str
phase: 'train' or 'val'
"""
if self.tf_logging:
tag = f"{phase}-{tag}"
if isinstance(metrics, dict):
self.writer.add_scalars(tag, metrics, self.log_step[phase])
else:
self.writer.add_scalar(tag, metrics, self.log_step[phase])
if add_global_step:
self.log_step[phase] += 1
else:
pass
def load_model(self, model, ckpt_path=None):
if self.rank == 0:
self.logger.info(f'Loading from {ckpt_path}...')
ckpt = torch.load(ckpt_path, map_location=f"cuda:{self.rank}")
if 'state_dict' in ckpt:
ckpt = ckpt['state_dict']
util_net.reload_model(model, ckpt)
if self.rank == 0:
self.logger.info('Loaded Done')
def freeze_model(self, net):
for params in net.parameters():
params.requires_grad = False
class TrainerDifIR(TrainerBase):
def setup_optimizaton(self):
super().setup_optimizaton()
if self.configs.train.lr_schedule == 'cosin':
self.lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer=self.optimizer,
T_max=self.configs.train.iterations - self.configs.train.warmup_iterations,
eta_min=self.configs.train.lr_min,
)
def build_model(self):
super().build_model()
if self.rank == 0 and hasattr(self.configs.train, 'ema_rate'):
self.ema_ignore_keys.extend([x for x in self.ema_state.keys() if 'relative_position_index' in x])
# autoencoder
if self.configs.autoencoder is not None:
ckpt = torch.load(self.configs.autoencoder.ckpt_path, map_location=f"cuda:{self.rank}")
if self.rank == 0:
self.logger.info(f"Restoring autoencoder from {self.configs.autoencoder.ckpt_path}")
params = self.configs.autoencoder.get('params', dict)
autoencoder = util_common.get_obj_from_str(self.configs.autoencoder.target)(**params)
autoencoder.cuda()
autoencoder.load_state_dict(ckpt, True)
for params in autoencoder.parameters():
params.requires_grad_(False)
autoencoder.eval()
if self.configs.train.compile.flag:
if self.rank == 0:
self.logger.info("Begin compiling autoencoder model...")
autoencoder = torch.compile(autoencoder, mode=self.configs.train.compile.mode)
if self.rank == 0:
self.logger.info("Compiling Done")
self.autoencoder = autoencoder
else:
self.autoencoder = None
# LPIPS metric
lpips_loss = lpips.LPIPS(net='vgg').to(f"cuda:{self.rank}")
for params in lpips_loss.parameters():
params.requires_grad_(False)
lpips_loss.eval()
if self.configs.train.compile.flag:
if self.rank == 0:
self.logger.info("Begin compiling LPIPS Metric...")
lpips_loss = torch.compile(lpips_loss, mode=self.configs.train.compile.mode)
if self.rank == 0:
self.logger.info("Compiling Done")
self.lpips_loss = lpips_loss
params = self.configs.diffusion.get('params', dict)
self.base_diffusion = util_common.get_obj_from_str(self.configs.diffusion.target)(**params)
@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.degradation.get('queue_size', b*10)
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, dtype=torch.float32, realesrgan=None, phase='train'):
if realesrgan is None:
realesrgan = self.configs.data.get(phase, dict).type == 'realesrgan'
if realesrgan and phase == 'train':
if not hasattr(self, 'jpeger'):
self.jpeger = DiffJPEG(differentiable=False).cuda() # simulate JPEG compression artifacts
if not hasattr(self, 'use_sharpener'):
self.use_sharpener = USMSharp().cuda()
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]
if isinstance(self.configs.degradation.sf, int):
sf = self.configs.degradation.sf
else:
assert len(self.configs.degradation.sf) == 2
sf = random.uniform(*self.configs.degradation.sf)
if self.configs.degradation.use_sharp:
im_gt = self.use_sharpener(im_gt)
# ----------------------- 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 ----------------------- #
if random.random() < self.configs.degradation['second_order_prob']:
# 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 / sf * scale), int(ori_w / 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 // sf, ori_w // 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 // sf, ori_w // sf),
mode=mode,
)
out = filter2D(out, sinc_kernel)
# resize back
if self.configs.degradation.resize_back:
out = F.interpolate(out, size=(ori_h, ori_w), mode='bicubic')
temp_sf = self.configs.degradation['sf']
else:
temp_sf = self.configs.degradation['sf']
# 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, temp_sf)
im_lq = (im_lq - 0.5) / 0.5 # [0, 1] to [-1, 1]
im_gt = (im_gt - 0.5) / 0.5 # [0, 1] to [-1, 1]
self.lq, self.gt, flag_nan = replace_nan_in_batch(im_lq, im_gt)
if flag_nan:
with open(f"records_nan_rank{self.rank}.log", 'a') as f:
f.write(f'Find Nan value in rank{self.rank}\n')
# training pair pool
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}
elif phase == 'val':
offset = self.configs.train.get('val_resolution', 256)
for key, value in data.items():
h, w = value.shape[2:]
if h > offset and w > offset:
h_end = int((h // offset) * offset)
w_end = int((w // offset) * offset)
data[key] = value[:, :, :h_end, :w_end]
else:
h_pad = math.ceil(h / offset) * offset - h
w_pad = math.ceil(w / offset) * offset - w
padding_mode = self.configs.train.get('val_padding_mode', 'reflect')
data[key] = F.pad(value, pad=(0, w_pad, 0, h_pad), mode=padding_mode)
return {key:value.cuda().to(dtype=dtype) for key, value in data.items()}
else:
return {key:value.cuda().to(dtype=dtype) for key, value in data.items()}
def backward_step(self, dif_loss_wrapper, micro_data, num_grad_accumulate, tt):
context = torch.cuda.amp.autocast if self.configs.train.use_amp else nullcontext
with context():
losses, z_t, z0_pred = dif_loss_wrapper()
losses['loss'] = losses['mse']
loss = losses['loss'].mean() / num_grad_accumulate
if self.amp_scaler is None:
loss.backward()
else:
self.amp_scaler.scale(loss).backward()
return losses, z0_pred, z_t
def training_step(self, data):
current_batchsize = data['gt'].shape[0]
micro_batchsize = self.configs.train.microbatch
num_grad_accumulate = math.ceil(current_batchsize / micro_batchsize)
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 = torch.randint(
0, self.base_diffusion.num_timesteps,
size=(micro_data['gt'].shape[0],),
device=f"cuda:{self.rank}",
)
latent_downsamping_sf = 2**(len(self.configs.autoencoder.params.ddconfig.ch_mult) - 1)
latent_resolution = micro_data['gt'].shape[-1] // latent_downsamping_sf
if 'autoencoder' in self.configs:
noise_chn = self.configs.autoencoder.params.embed_dim
else:
noise_chn = micro_data['gt'].shape[1]
noise = torch.randn(
size= (micro_data['gt'].shape[0], noise_chn,) + (latent_resolution, ) * 2,
device=micro_data['gt'].device,
)
if self.configs.model.params.cond_lq:
model_kwargs = {'lq':micro_data['lq'],}
if 'mask' in micro_data:
model_kwargs['mask'] = micro_data['mask']
else:
model_kwargs = None
compute_losses = functools.partial(
self.base_diffusion.training_losses,
self.model,
micro_data['gt'],
micro_data['lq'],
tt,
first_stage_model=self.autoencoder,
model_kwargs=model_kwargs,
noise=noise,
)
if last_batch or self.num_gpus <= 1:
losses, z0_pred, z_t = self.backward_step(compute_losses, micro_data, num_grad_accumulate, tt)
else:
with self.model.no_sync():
losses, z0_pred, z_t = self.backward_step(compute_losses, micro_data, num_grad_accumulate, tt)
# make logging
if last_batch:
self.log_step_train(losses, tt, micro_data, z_t, z0_pred.detach())
if self.configs.train.use_amp:
self.amp_scaler.step(self.optimizer)
self.amp_scaler.update()
else:
self.optimizer.step()
# grad zero
self.model.zero_grad()
if hasattr(self.configs.train, 'ema_rate'):
self.update_ema_model()
def adjust_lr(self, current_iters=None):
base_lr = self.configs.train.lr
warmup_steps = self.configs.train.warmup_iterations
current_iters = self.current_iters if current_iters is None else current_iters
if current_iters <= warmup_steps:
for params_group in self.optimizer.param_groups:
params_group['lr'] = (current_iters / warmup_steps) * base_lr
else:
if hasattr(self, 'lr_scheduler'):
self.lr_scheduler.step()
def log_step_train(self, loss, tt, batch, z_t, z0_pred, phase='train'):
'''
param loss: a dict recording the loss informations
param tt: 1-D tensor, time steps
'''
if self.rank == 0:
chn = batch['gt'].shape[1]
num_timesteps = self.base_diffusion.num_timesteps
record_steps = [1, (num_timesteps // 2) + 1, num_timesteps]
if self.current_iters % self.configs.train.log_freq[0] == 1:
self.loss_mean = {key:torch.zeros(size=(len(record_steps),), dtype=torch.float64)
for key in loss.keys()}
self.loss_count = torch.zeros(size=(len(record_steps),), dtype=torch.float64)
for jj in range(len(record_steps)):
for key, value in loss.items():
index = record_steps[jj] - 1
mask = torch.where(tt == index, torch.ones_like(tt), torch.zeros_like(tt))
current_loss = torch.sum(value.detach() * mask)
self.loss_mean[key][jj] += current_loss.item()
self.loss_count[jj] += mask.sum().item()
if self.current_iters % self.configs.train.log_freq[0] == 0:
if torch.any(self.loss_count == 0):
self.loss_count += 1e-4
for key in loss.keys():
self.loss_mean[key] /= self.loss_count
log_str = 'Train: {:06d}/{:06d}, Loss/MSE: '.format(
self.current_iters,
self.configs.train.iterations)
for jj, current_record in enumerate(record_steps):
log_str += 't({:d}):{:.1e}/{:.1e}, '.format(
current_record,
self.loss_mean['loss'][jj].item(),
self.loss_mean['mse'][jj].item(),
)
log_str += 'lr:{:.2e}'.format(self.optimizer.param_groups[0]['lr'])
self.logger.info(log_str)
self.logging_metric(self.loss_mean, tag='Loss', phase=phase, add_global_step=True)
if self.current_iters % self.configs.train.log_freq[1] == 0:
self.logging_image(batch['lq'], tag='lq', phase=phase, add_global_step=False)
self.logging_image(batch['gt'], tag='gt', phase=phase, add_global_step=False)
x_t = self.base_diffusion.decode_first_stage(
self.base_diffusion._scale_input(z_t, tt),
self.autoencoder,
)
self.logging_image(x_t, tag='diffused', phase=phase, add_global_step=False)
x0_pred = self.base_diffusion.decode_first_stage(
z0_pred,
self.autoencoder,
)
self.logging_image(x0_pred, tag='x0-pred', phase=phase, add_global_step=True)
if self.current_iters % self.configs.train.save_freq == 1:
self.tic = time.time()
if self.current_iters % self.configs.train.save_freq == 0:
self.toc = time.time()
elaplsed = (self.toc - self.tic)
self.logger.info(f"Elapsed time: {elaplsed:.2f}s")
self.logger.info("="*100)
def validation(self, phase='val'):
if self.rank == 0:
if self.configs.train.use_ema_val:
self.reload_ema_model()
self.ema_model.eval()
else:
self.model.eval()
indices = np.linspace(
0,
self.base_diffusion.num_timesteps,
self.base_diffusion.num_timesteps if self.base_diffusion.num_timesteps < 5 else 4,
endpoint=False,
dtype=np.int64,
).tolist()
if not (self.base_diffusion.num_timesteps-1) in indices:
indices.append(self.base_diffusion.num_timesteps-1)
batch_size = self.configs.train.batch[1]
num_iters_epoch = math.ceil(len(self.datasets[phase]) / batch_size)
mean_psnr = mean_lpips = 0
for ii, data in enumerate(self.dataloaders[phase]):
data = self.prepare_data(data, phase='val')
if 'gt' in data:
im_lq, im_gt = data['lq'], data['gt']
else:
im_lq = data['lq']
num_iters = 0
if self.configs.model.params.cond_lq:
model_kwargs = {'lq':data['lq'],}
if 'mask' in data:
model_kwargs['mask'] = data['mask']
else:
model_kwargs = None
tt = torch.tensor(
[self.base_diffusion.num_timesteps, ]*im_lq.shape[0],
dtype=torch.int64,
).cuda()
for sample in self.base_diffusion.p_sample_loop_progressive(
y=im_lq,
model=self.ema_model if self.configs.train.use_ema_val else self.model,
first_stage_model=self.autoencoder,
noise=None,
clip_denoised=True if self.autoencoder is None else False,
model_kwargs=model_kwargs,
device=f"cuda:{self.rank}",
progress=False,
):
sample_decode = {}
if num_iters in indices:
for key, value in sample.items():
if key in ['sample', ]:
sample_decode[key] = self.base_diffusion.decode_first_stage(
value,
self.autoencoder,
).clamp(-1.0, 1.0)
im_sr_progress = sample_decode['sample']
if num_iters + 1 == 1:
im_sr_all = im_sr_progress
else:
im_sr_all = torch.cat((im_sr_all, im_sr_progress), dim=1)
num_iters += 1
tt -= 1
if 'gt' in data:
mean_psnr += util_image.batch_PSNR(
sample_decode['sample'] * 0.5 + 0.5,
im_gt * 0.5 + 0.5,
ycbcr=self.configs.train.val_y_channel,
)
mean_lpips += self.lpips_loss(
sample_decode['sample'],
im_gt,
).sum().item()
if (ii + 1) % self.configs.train.log_freq[2] == 0:
self.logger.info(f'Validation: {ii+1:02d}/{num_iters_epoch:02d}...')
im_sr_all = rearrange(im_sr_all, 'b (k c) h w -> (b k) c h w', c=im_lq.shape[1])
self.logging_image(
im_sr_all,
tag='progress',
phase=phase,
add_global_step=False,
nrow=len(indices),
)
if 'gt' in data:
self.logging_image(im_gt, tag='gt', phase=phase, add_global_step=False)
self.logging_image(im_lq, tag='lq', phase=phase, add_global_step=True)
if 'gt' in data:
mean_psnr /= len(self.datasets[phase])
mean_lpips /= len(self.datasets[phase])
self.logger.info(f'Validation Metric: PSNR={mean_psnr:5.2f}, LPIPS={mean_lpips:6.4f}...')
self.logging_metric(mean_psnr, tag='PSNR', phase=phase, add_global_step=False)
self.logging_metric(mean_lpips, tag='LPIPS', phase=phase, add_global_step=True)
self.logger.info("="*100)
if not (self.configs.train.use_ema_val and hasattr(self.configs.train, 'ema_rate')):
self.model.train()
class TrainerDifIRLPIPS(TrainerDifIR):
def backward_step(self, dif_loss_wrapper, micro_data, num_grad_accumulate, tt):
loss_coef = self.configs.train.get('loss_coef')
context = torch.cuda.amp.autocast if self.configs.train.use_amp else nullcontext
# diffusion loss
with context():
losses, z_t, z0_pred = dif_loss_wrapper()
x0_pred = self.base_diffusion.decode_first_stage(
z0_pred,
self.autoencoder,
) # f16
self.current_x0_pred = x0_pred.detach()
# classification loss
losses["lpips"] = self.lpips_loss(
x0_pred.clamp(-1.0, 1.0),
micro_data['gt'],
).to(z0_pred.dtype).view(-1)
flag_nan = torch.any(torch.isnan(losses["lpips"]))
if flag_nan:
losses["lpips"] = torch.nan_to_num(losses["lpips"], nan=0.0)
losses["mse"] *= loss_coef[0]
losses["lpips"] *= loss_coef[1]
assert losses["mse"].shape == losses["lpips"].shape
if flag_nan:
losses["loss"] = losses["mse"]
else:
losses["loss"] = losses["mse"] + losses["lpips"]
loss = losses['loss'].mean() / num_grad_accumulate
if self.amp_scaler is None:
loss.backward()
else:
self.amp_scaler.scale(loss).backward()
return losses, z0_pred, z_t
def log_step_train(self, loss, tt, batch, z_t, z0_pred, phase='train'):
'''
param loss: a dict recording the loss informations
param tt: 1-D tensor, time steps
'''
if self.rank == 0:
chn = batch['gt'].shape[1]
num_timesteps = self.base_diffusion.num_timesteps
record_steps = [1, (num_timesteps // 2) + 1, num_timesteps]
if self.current_iters % self.configs.train.log_freq[0] == 1:
self.loss_mean = {key:torch.zeros(size=(len(record_steps),), dtype=torch.float64)
for key in loss.keys()}
self.loss_count = torch.zeros(size=(len(record_steps),), dtype=torch.float64)
for jj in range(len(record_steps)):
for key, value in loss.items():
index = record_steps[jj] - 1
mask = torch.where(tt == index, torch.ones_like(tt), torch.zeros_like(tt))
assert value.shape == mask.shape
current_loss = torch.sum(value.detach() * mask)
self.loss_mean[key][jj] += current_loss.item()
self.loss_count[jj] += mask.sum().item()
if self.current_iters % self.configs.train.log_freq[0] == 0:
if torch.any(self.loss_count == 0):
self.loss_count += 1e-4
for key in loss.keys():
self.loss_mean[key] /= self.loss_count
log_str = 'Train: {:06d}/{:06d}, MSE/LPIPS: '.format(
self.current_iters,
self.configs.train.iterations)
for jj, current_record in enumerate(record_steps):
log_str += 't({:d}):{:.1e}/{:.1e}, '.format(
current_record,
self.loss_mean['mse'][jj].item(),
self.loss_mean['lpips'][jj].item(),
)
log_str += 'lr:{:.2e}'.format(self.optimizer.param_groups[0]['lr'])
self.logger.info(log_str)
self.logging_metric(self.loss_mean, tag='Loss', phase=phase, add_global_step=True)
if self.current_iters % self.configs.train.log_freq[1] == 0:
self.logging_image(batch['lq'], tag='lq', phase=phase, add_global_step=False)
self.logging_image(batch['gt'], tag='gt', phase=phase, add_global_step=False)
x_t = self.base_diffusion.decode_first_stage(
self.base_diffusion._scale_input(z_t, tt),
self.autoencoder,
)
self.logging_image(x_t, tag='diffused', phase=phase, add_global_step=False)
self.logging_image(self.current_x0_pred, tag='x0-pred', phase=phase, add_global_step=True)
if self.current_iters % self.configs.train.save_freq == 1:
self.tic = time.time()
if self.current_iters % self.configs.train.save_freq == 0:
self.toc = time.time()
elaplsed = (self.toc - self.tic)
self.logger.info(f"Elapsed time: {elaplsed:.2f}s")
self.logger.info("="*100)
def replace_nan_in_batch(im_lq, im_gt):
'''
Input:
im_lq, im_gt: b x c x h x w
'''
if torch.isnan(im_lq).sum() > 0:
valid_index = []
im_lq = im_lq.contiguous()
for ii in range(im_lq.shape[0]):
if torch.isnan(im_lq[ii,]).sum() == 0:
valid_index.append(ii)
assert len(valid_index) > 0
im_lq, im_gt = im_lq[valid_index,], im_gt[valid_index,]
flag = True
else:
flag = False
return im_lq, im_gt, flag
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)))