torch.utils.data.dataloader中的collate_fn数据拼接方式那种最快

1、cat

		'
		'
		'
        if index == 0:
            noisy_wola_sec2 = noisy_wola_sec
            clean_wola_sec2 = clean_wola_sec
            noisy_wola_main2 = noisy_wola_main
            clean_wola_main2 = clean_wola_main
        else:
            noisy_wola_sec2 = torch.cat((noisy_wola_sec2, noisy_wola_sec), 0)
            clean_wola_sec2 = torch.cat((clean_wola_sec2, clean_wola_sec), 0)     
            noisy_wola_main2 = torch.cat((noisy_wola_main2, noisy_wola_main), 0)
            clean_wola_main2 = torch.cat((clean_wola_main2, clean_wola_main), 0)  

平均一次的时间 0.32737993749600012 s
torch.Size([24, 2, 200, 321])

2、stack

    noisy_wola_sec2 = []
    clean_wola_sec2 = []
    noisy_wola_main2 = []
    clean_wola_main2 = []
		
		'
		'
		'
        noisy_wola_sec2.append(noisy_wola_sec)
        clean_wola_sec2.append(clean_wola_sec)
        noisy_wola_main2.append(noisy_wola_main)
        clean_wola_main2.append(clean_wola_main)  
        index += 1
    noisy_wola_sec2 = torch.stack(noisy_wola_sec2,dim=0)
    clean_wola_sec2 = torch.stack(clean_wola_sec2,dim=0)
    noisy_wola_main2 = torch.stack(noisy_wola_main2,dim=0)
    clean_wola_main2 = torch.stack(clean_wola_main2,dim=0)

平均一次的时间 0.23807710859812956 s
torch.Size([24, 2, 200, 321])
3、矩阵赋值

   noisy_wola_sec2 = torch.zeros((24, 2, 200, 321))
   clean_wola_sec2 = torch.zeros((24, 2, 200, 321))
   noisy_wola_main2 = torch.zeros((24, 2, 200, 321))
   clean_wola_main2 = torch.zeros((24, 2, 200, 321))
		'
		'
		'
       noisy_wola_sec2[index] = noisy_wola_sec
       clean_wola_sec2[index] = clean_wola_sec
       noisy_wola_main2[index] = noisy_wola_main
       clean_wola_main2[index] = clean_wola_main
 
       index += 1

平均一次的时间 0.20699568727310752 s
torch.Size([24, 2, 200, 321])

总结
虽然3是最快的但是不灵活推荐2,因为最后一个batch不是固定的大小可能每次循环到最后会报错。