各式各样的轮子们

1:同比例缩放

有时候直接进行resize会有形变，所以想到这样的方式，同比例缩放，然后补0。torchvision中是用的PIL。在推理时需要用opencv。

def ZeroPaddingResizeCV(img, size=(224, 224), interpolation=None):
    isize = img.shape
    ih, iw = isize[0], isize[1]
    h, w = size[0], size[1]
    scale = min(w / iw, h / ih)
    new_w = int(iw * scale + 0.5)
    new_h = int(ih * scale + 0.5)
 
    img = cv2.resize(img, (new_w, new_h), interpolation)
    new_img = np.zeros((h, w, 3), np.uint8)
    new_img[(h-new_h)//2:(h+new_h)//2, (w-new_w)//2:(w+new_w)//2] = img
 
    return new_img
  
new_image=ZeroPaddingResizeCV(img,(96,96))

2。log

import logging
def getLogger(log_path):
    logger = logging.getLogger()
    logger.setLevel(logging.INFO)  # Log等级总开关
    formatter = logging.Formatter(fmt="[%(asctime)s|%(filename)s|%(levelname)s] %(message)s",
                                  datefmt="%a %b %d %H:%M:%S %Y")
    # StreamHandler
    sHandler = logging.StreamHandler()
    sHandler.setFormatter(formatter)
    logger.addHandler(sHandler)
    fHandler = logging.FileHandler(log_path, mode='w')
    fHandler.setLevel(logging.DEBUG)  # 输出到file的log等级的开关
    fHandler.setFormatter(formatter)  # 定义handler的输出格式
    logger.addHandler(fHandler)  # 将logger添加到handler里面
    return logger

3，计算模型的flops与params

from models.AudioSync import AudioSync_1M_emotion,AudioSync_3M_emotion,AudioSync_21M_emotion,AudioSync_39M_emotion
import torch 
from thop import profile

use_cuda = torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
model=AudioSync_3M_emotion(26).to(device)   #26，11
input1 = torch.randn(1,1,80,16).cuda()

flops, params = profile(model, inputs=(input1, ))
# print('FLOPs = ' + str(flops/1000**3) + 'G')
# print('Params = ' + str(params/1000**2) + 'M')

print('FLOPs = ' + str(flops))
print('Params = ' + str(params))

图片到视频，无损合成

import subprocess
image_folder = '/data/dengjia/projects/digithuman/tmp/data/evalvideo/train/chuyao_white/syncnet_crop_face/%6d.jpg'
video_path = '/data/dengjia/projects/digithuman/chuyao_white.mp4'
command = ("ffmpeg -f image2 -framerate 25 -i %s  -b:v 5626k %s" % (image_folder,video_path)) 
output = subprocess.call(command, shell=True, stdout=None)

合成的视频，再逐帧打开，视频帧的分辨率与合成之前的图片一致。

在mac上可以打开使用，但是在win下格式编码不支持，需要在合成生成的视频.mp4改为.avi即可

import subprocess
image_folder = '/data/dengjia/projects/digithuman/tmp/data/evalvideo/train/chuyao_white/syncnet_crop_face/%6d.jpg'
video_path = '/data/dengjia/projects/digithuman/chuyao_white.avi'
command = ("ffmpeg -f image2 -framerate 25 -i %s  -b:v 5626k %s" % (image_folder,video_path)) 
output = subprocess.call(command, shell=True, stdout=None)

踩坑记录：

1:数据集打包

loader = data.DataLoader(
        dataset,
        num_workers=8,
        batch_size=args.batch_size // world_size,
        sampler=data.distributed.DistributedSampler(dataset, num_replicas=world_size, rank=rank, shuffle=True),
        pin_memory=True,
        drop_last=True,
    )

drop_last=True 表示最后数据不足一个batch时，丢弃
如果默认drop_last=False，有可能最后一个batch不是你设置的那个数，后面训练有可能存在数据维度对不齐的问题，eg:RuntimeError: shape ‘[4, -1, 1, 512, 4, 4]’ is invalid for input of size 73728等