YOND / utils /utils.py
hansen97's picture
Initial clean commit
0e07d71
import os
# os.environ['OPENBLAS_WARNINGS'] = '0'
# os.environ["OMP_NUM_THREADS"] = "1"
# os.environ["MKL_NUM_THREADS"] = "1"
import cv2
cv2.setNumThreads(0)
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.distributed as dist
from torch.utils.data import Dataset, DataLoader, DistributedSampler
import torch.multiprocessing as mp
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.optim.lr_scheduler import *
import glob
import matplotlib
# matplotlib.use('AGG')
import matplotlib.pyplot as plt
import numpy as np
import gc
from PIL import Image
import time
import socket
import scipy
import scipy.io as sio
from scipy import stats
import argparse
from skimage.metrics import peak_signal_noise_ratio as compare_psnr
from skimage.metrics import structural_similarity as compare_ssim
from multiprocessing import Pool
from concurrent.futures import ThreadPoolExecutor, ProcessPoolExecutor, as_completed
import threading
from functools import wraps
from tqdm import tqdm
import exifread
import rawpy
import math
import random
import yaml
import pickle
import warnings
import h5py
import pickle
import pickle as pkl
from natsort import natsort
import scipy.io
from scipy.stats import poisson, norm
from scipy.signal import convolve
from scipy.interpolate import interp1d
import warnings
import kornia.filters as kf
from natsort import natsorted
def setup_seed(seed):
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
random.seed(seed)
torch.backends.cudnn.deterministic = True
setup_seed(1997)
fn_time = {}
def timestamp(time_points, n):
time_points[n] = time.time()
return time_points[n] - time_points[n-1]
def fn_timer(function, print_log=False):
@wraps(function)
def function_timer(*args, **kwargs):
global fn_timer
t0 = time.time()
result = function(*args, **kwargs)
t1 = time.time()
if print_log:
print ("Total time running %s: %.6f seconds" %
(function.__name__, t1-t0))
if function.__name__ in fn_time :
fn_time[function.__name__] += t1-t0
else:
fn_time[function.__name__] = t1-t0
return result
return function_timer
def log(string, log=None, str=False, end='\n', notime=False):
log_string = f'{time.strftime("%Y-%m-%d %H:%M:%S")} >> {string}' if not notime else string
print(log_string)
if log is not None:
with open(log,'a+') as f:
f.write(log_string+'\n')
else:
pass
# os.makedirs('worklog', exist_ok=True)
# log = f'worklog/worklog-{time.strftime("%Y-%m-%d")}.txt'
# with open(log,'a+') as f:
# f.write(log_string+'\n')
if str:
return string+end
class AverageMeter(object):
"""Computes and stores the average and current value"""
def __init__(self, name, fmt=':f', log=True, last_epoch=0):
self.name = name
self.fmt = fmt
self.log = log
self.history = []
self.last_epoch = last_epoch
self.history_init_flag = False
self.reset()
def reset(self):
if self.log:
try:
if self.avg>0: self.history.append(self.avg)
except:
pass#print(f'Start log {self.name}!')
self.val = 0
self.avg = 0
self.sum = 0
self.count = 0
def update(self, val, n=1):
self.val = val
self.sum += val * n
self.count += n
self.avg = self.sum / self.count
def plot_history(self, savefile='log.jpg', logfile='log.pkl'):
# 读取老log
if os.path.exists(logfile) and not self.history_init_flag:
self.history_init_flag = True
with open(logfile, 'rb') as f:
history_old = pickle.load(f)
if self.last_epoch: # 为0则重置
self.history = history_old + self.history[:self.last_epoch]
# 记录log
with open(logfile, 'wb') as f:
pickle.dump(self.history, f)
# 画图
plt.figure(figsize=(12,9))
plt.title(f'{self.name} log')
x = list(range(len(self.history)))
plt.plot(x, self.history)
plt.xlabel('Epoch')
plt.ylabel(self.name)
plt.savefig(savefile, bbox_inches='tight')
plt.close()
def __str__(self):
fmtstr = '{name}:{val' + self.fmt + '}({avg' + self.fmt + '})'
return fmtstr.format(**self.__dict__)
def pkl_convert(param):
return {
k.replace("module.", ""): v
for k, v in param.items()
if "module." in k
}
def load_weights(model, pretrained_dict, multi_gpu=False, by_name=False):
model_dict = model.module.state_dict() if multi_gpu else model.state_dict()
# 1. filter out unnecessary keys
tsm_replace = []
for k in pretrained_dict:
if 'tsm_shift' in k:
k_new = k.replace('tsm_shift', 'tsm_buffer')
tsm_replace.append((k, k_new))
for k, k_new in tsm_replace:
pretrained_dict[k_new] = pretrained_dict[k]
if by_name:
del_list = []
for k, v in pretrained_dict.items():
if k in model_dict:
if model_dict[k].shape != pretrained_dict[k].shape:
# 1. Delete values not in key
del_list.append(k)
# 2. Cat it to the end
# diff = model_dict[k].size()[1] - pretrained_dict[k].size()[1]
# v = torch.cat((v, v[:,:diff]), dim=1)
# 3. Repeat it to same
# nframe = model_dict[k].shape[1] // pretrained_dict[k].shape[1]
# v = torch.repeat_interleave(v, nframe, dim=1)
# 4. Clip it to same
# c_model = model_dict[k].shape[1]
# c_save = pretrained_dict[k].shape[1]
# c_diff = c_model - c_save
# if c_model > c_save:
# v = torch.cat((v, torch.empty(b_model, c_diff, h_model, w_model).cuda()), dim=1)
# else:
# v = v[:,:c_diff]
log(f'Warning: "{k}":{pretrained_dict[k].shape}->{model_dict[k].shape}')
pretrained_dict[k] = v
else:
del_list.append(k)
log(f'Warning: "{k}" is not exist and has been deleted!!')
for k in del_list:
del pretrained_dict[k]
# 2. overwrite entries in the existing state dict
model_dict.update(pretrained_dict)
if multi_gpu:
model.module.load_state_dict(model_dict)
else:
model.load_state_dict(model_dict)
# for name, parameter in model.named_parameters():
# if name not in ["input_blocks.0.0.weight", "out.2.weight", "out.2.bias"]:
# # if name not in del_list:
# parameter.requires_grad = False
# log(f'Warning: layer except "{name}" is freezed...')
return model
def tensor_dimxto4(tensor):
c, h, w = tensor.shape[-3:]
tensor = tensor.reshape(-1, c, h, w)
return tensor
def tensor_dimxto5(tensor):
t, c, h, w = tensor.shape[-4:]
tensor = tensor.reshape(-1, t, c, h, w)
return tensor
def tensor_dim5to4(tensor):
batchsize, crops, c, h, w = tensor.shape
tensor = tensor.reshape(batchsize*crops, c, h, w)
return tensor
def tensor_dim6to5(tensor):
batchsize, crops, t, c, h, w = tensor.shape
tensor = tensor.reshape(batchsize*crops, t, c, h, w)
return tensor
def get_host_with_dir(dataset_name=''):
multi_gpu = False
hostname = socket.gethostname()
log(f"User's hostname is '{hostname}'")
if hostname == 'ubun':
host = '/data/fenghansen/datasets'
elif hostname == 'ubuntu':
host = '/data4/fenghansen/datasets'
elif hostname == 'DESKTOP-FCAMIOQ':
host = 'F:/datasets'
elif hostname == 'DESKTOP-LGD8S6F': # BIT-816
host = 'E:/datasets'
elif hostname[:6] == 'isp-gn':
host = '/home/fenghansen/datasets'
else:
host = '/data'
multi_gpu = True if torch.cuda.device_count() > 1 else False
return hostname, host + dataset_name, multi_gpu
def get_p2d(shape, base=16):
xb, xc, xh, xw = shape
yh, yw = ((xh-1)//base+1)*base, ((xw-1)//base+1)*base
diffY = yh - xh
diffX = yw - xw
p2d = (diffX // 2, diffX - diffX//2, diffY // 2, diffY - diffY//2)
return p2d
# def big_image_split(data, n=2, pad=64):
# # 把大图分割成小图
# p2d = get_p2d(data.shape, base=pad)
# data = F.pad(data, p2d, mode='reflect')
# data = torch.cat(torch.chunk(data, n, dim=2), dim=0)
# data = torch.cat(torch.chunk(data, n, dim=3), dim=0)
# return data, p2d
# def big_image_merge(data, n=2, p2d=[0,0,0,0]):
# # 把小图合并成大图
# data = torch.cat(torch.chunk(data, n, dim=0), dim=3)
# data = torch.cat(torch.chunk(data, n, dim=0), dim=2)
# H, W = data.shape[-2:]
# data = data[..., p2d[-2]:H-p2d[-1], p2d[0]:W-p2d[1]]
# return data
def calculate_padding(shape, target_size, overlap_ratio=0.25):
"""计算需要的padding以确保图片能被均匀切块,只在右下角pad"""
h, w = shape[-2:]
# 如果图片尺寸小于目标尺寸,不进行padding和裁剪
if h <= target_size and w <= target_size:
return [0, 0, 0, 0], (1, 1), True # 不进行padding,只生成一个块
stride_h = int(target_size * (1 - overlap_ratio))
stride_w = int(target_size * (1 - overlap_ratio))
# 计算需要的块数
n_h = max(1, (h - target_size + stride_h - 1) // stride_h + 1)
n_w = max(1, (w - target_size + stride_w - 1) // stride_w + 1)
# 计算最终需要的尺寸
final_h = (n_h - 1) * stride_h + target_size
final_w = (n_w - 1) * stride_w + target_size
# 只在右下角添加padding
pad_bottom = final_h - h
pad_right = final_w - w
return [0, pad_right, 0, pad_bottom], (n_h, n_w), False # [左, 右, 上, 下]
def big_image_split(data, target_size=512, overlap_ratio=0.25, pad_mode='reflect'):
"""
将大图分割成多个有重叠区域的小图,只在右下角进行padding
参数:
data: 输入的图像张量 [B,C,H,W]
target_size: 每个切块的目标尺寸
overlap_ratio: 重叠区域比例(0-1)
pad_mode: padding模式
返回:
patches: 切块后的图像张量 [B*num_patches,C,target_size,target_size]
metadata: 包含padding信息和切块数量的元数据
"""
# 确保输入是4D张量
if data.dim() == 3:
data = data.unsqueeze(0)
# 计算padding(只在右下角)
p2d, (n_h, n_w), is_original = calculate_padding(data.shape, target_size, overlap_ratio)
# 保存原始尺寸
original_shape = data.shape
# 如果图片尺寸小于目标尺寸,不进行padding和裁剪
if is_original:
# 确保输出格式一致,但实际上只有一个块
patches = data
metadata = {
'p2d': p2d,
'n_h': n_h,
'n_w': n_w,
'stride_h': target_size, # 整个图片作为一个块
'stride_w': target_size,
'target_size': target_size,
'original_shape': original_shape,
'is_original': True
}
return patches, metadata
data = F.pad(data, p2d, mode=pad_mode)
# 获取padding后的尺寸
B, C, H, W = data.shape
# 计算步长
stride_h = int(target_size * (1 - overlap_ratio))
stride_w = int(target_size * (1 - overlap_ratio))
# 创建滑动窗口
patches = data.unfold(2, target_size, stride_h).unfold(3, target_size, stride_w)
patches = patches.contiguous().view(B, C, n_h, n_w, target_size, target_size)
# 重新排列维度
patches = patches.permute(0, 2, 3, 1, 4, 5).contiguous()
patches = patches.view(B * n_h * n_w, C, target_size, target_size)
# 保存元数据
metadata = {
'p2d': p2d,
'n_h': n_h,
'n_w': n_w,
'stride_h': stride_h,
'stride_w': stride_w,
'target_size': target_size,
'original_shape': original_shape,
'is_original': False
}
return patches, metadata
def big_image_merge(patches, metadata, blend_mode='triangle'):
"""
将多个小图重新合并成大图
参数:
patches: 切块后的图像张量 [B*num_patches,C,target_size,target_size]
metadata: 包含padding信息和切块数量的元数据
blend_mode: 融合模式,支持 'avg'(平均)、'max'(最大值)或'triangle'(三角模糊过渡)
返回:
merged: 合并后的图像张量 [B,C,H,W]
"""
# 如果是原图,直接返回
if metadata.get('is_original', False):
return patches
# 提取元数据
p2d = metadata['p2d']
n_h = metadata['n_h']
n_w = metadata['n_w']
stride_h = metadata['stride_h']
stride_w = metadata['stride_w']
target_size = metadata['target_size']
original_shape = metadata['original_shape']
B_merged = patches.shape[0] // (n_h * n_w)
# 重塑patches
patches = patches.view(B_merged, n_h, n_w, patches.shape[1], target_size, target_size)
patches = patches.permute(0, 3, 1, 4, 2, 5).contiguous() # [B,C,n_h,target_size,n_w,target_size]
# 创建用于合并的张量和权重图
merged = torch.zeros((B_merged, patches.shape[1],
(n_h-1)*stride_h+target_size,
(n_w-1)*stride_w+target_size), device=patches.device)
if blend_mode in ['avg', 'triangle']:
weight_map = torch.zeros_like(merged)
# 创建三角模糊权重图(如果使用该模式)
if blend_mode == 'triangle':
# 创建水平方向的权重渐变
overlap_w = target_size - stride_w
if overlap_w > 0:
h_weights = torch.ones(target_size, target_size, device=patches.device)
# 左侧权重从0线性增加到1
left_weights = torch.linspace(0, 1, overlap_w, device=patches.device).view(1, overlap_w)
h_weights[:, :overlap_w] = left_weights
# 右侧权重从1线性减少到0
right_weights = torch.linspace(1, 0, overlap_w, device=patches.device).view(1, overlap_w)
h_weights[:, -overlap_w:] = right_weights
else:
h_weights = torch.ones(target_size, target_size, device=patches.device)
# 创建垂直方向的权重渐变
overlap_h = target_size - stride_h
if overlap_h > 0:
v_weights = torch.ones(target_size, target_size, device=patches.device)
# 顶部权重从0线性增加到1
top_weights = torch.linspace(0, 1, overlap_h, device=patches.device).view(overlap_h, 1)
v_weights[:overlap_h, :] = top_weights
# 底部权重从1线性减少到0
bottom_weights = torch.linspace(1, 0, overlap_h, device=patches.device).view(overlap_h, 1)
v_weights[-overlap_h:, :] = bottom_weights
else:
v_weights = torch.ones(target_size, target_size, device=patches.device)
# 合并水平和垂直权重
triangle_weights = h_weights * v_weights
# 填充合并后的张量
for i in range(n_h):
for j in range(n_w):
h_start = i * stride_h
w_start = j * stride_w
if blend_mode == 'avg':
# 创建权重图以处理重叠区域
weight = torch.ones_like(patches[:, :, i, :, j, :])
merged[:, :, h_start:h_start+target_size, w_start:w_start+target_size] += patches[:, :, i, :, j, :]
weight_map[:, :, h_start:h_start+target_size, w_start:w_start+target_size] += weight
elif blend_mode == 'triangle':
# 使用三角模糊权重
current_patch = patches[:, :, i, :, j, :]
current_weight = triangle_weights.expand_as(current_patch)
merged[:, :, h_start:h_start+target_size, w_start:w_start+target_size] += current_patch * current_weight
weight_map[:, :, h_start:h_start+target_size, w_start:w_start+target_size] += current_weight
elif blend_mode == 'max':
# 使用最大值融合
current_patch = patches[:, :, i, :, j, :]
current_region = merged[:, :, h_start:h_start+target_size, w_start:w_start+target_size]
merged[:, :, h_start:h_start+target_size, w_start:w_start+target_size] = torch.max(current_region, current_patch)
# 处理平均融合或三角模糊融合
if blend_mode in ['avg', 'triangle']:
# 避免除零
weight_map = torch.clamp(weight_map, min=1e-8)
merged = merged / weight_map
# 去除padding(只去除右下角的padding)
pad_left, pad_right, pad_top, pad_bottom = p2d
H, W = merged.shape[-2:]
merged = merged[..., :H-pad_bottom, :W-pad_right] # 只移除右下角的padding
return merged
def tensor2numpy(data, eval=True, transpose=True):
if eval: data = data[0]
data = data.detach().cpu().numpy()
if transpose:
length = len(data.shape)
if length == 3:
data = data.transpose(1,2,0)
elif length == 4:
data = data.transpose(0,2,3,1)
return data
def numpy2tensor(data, device='cpu', eval=True, transpose=True, clone=False):
if clone: data = data.copy()
data = torch.from_numpy(np.ascontiguousarray(data)).float().to(device)
length = len(data.shape)
if transpose:
if length == 3:
data = data.permute(2,0,1)
elif length == 2:
data = data[None,:]
elif length == 4:
return data.permute(0,3,1,2)
if eval:
data = data[None,:]
return data
def read_paired_fns(filename):
with open(filename) as f:
fns = f.readlines()
fns = [tuple(fn.strip().split(' ')) for fn in fns]
return fns
def metrics_recorder(file, names, psnrs, ssims):
if os.path.exists(file):
with open(file, 'rb') as f:
metrics = pkl.load(f)
else:
metrics = {}
for name, psnr, ssim in zip(names, psnrs, ssims):
metrics[name] = [psnr, ssim]
with open(file, 'wb') as f:
pkl.dump(metrics, f)
return metrics
def mpop(func, idx, *args, **kwargs):
data = func(*args, **kwargs)
log(f'Finish task No.{idx}...')
return idx, func(*args, **kwargs)
def dataload(path):
suffix = path[-4:].lower()
if suffix in ['.arw','.dng','.nef','.cr2']:
data = rawpy.imread(path).raw_image_visible
elif suffix in ['.raw']:
data = np.fromfile(path, np.uint16).reshape(1440, 2560)
elif suffix in ['.npy']:
data = np.load(path)
elif suffix in ['.mat']:
if 'metadata' in path.lower():
data = scipy.io.loadmat(path)
else:
with h5py.File(path, 'r') as f:
data = np.array(f['x'])
elif suffix in ['.jpg', '.png', '.bmp', 'tiff']:
data = cv2.imread(path)[:,:,::-1]
return data
# 把ELD模型中的Unet权重单独提取出来
def pth_transfer(src_path='/data/ELD/checkpoints/sid-ours-inc4/model_200_00257600.pt',
dst_path='checkpoints/SonyA7S2_Official.pth',
reverse=False):
model_src = torch.load(src_path, map_location='cpu')
if reverse:
model_dst = torch.load(dst_path, map_location='cpu')
model_src['netG'] = model_dst
save_dir = os.path.join('pth_transfer', os.path.basename(dst_path)[9:-15])
os.makedirs(save_dir, exist_ok=True)
save_path = os.path.join(save_dir, os.path.basename(src_path))
torch.save(model_src, save_path)
else:
model_src = model_src['netG']
torch.save(model_src, dst_path)