yolov5 test. Py broken pipe error: [errno 32] broken pipe problem solving

• dataset.py

    batch_size = min(batch_size, len(dataset))
    nw = min([os.cpu_count() // world_size, batch_size if batch_size > 1 else 0, workers])  # number of workers
    sampler = torch.utils.data.distributed.DistributedSampler(dataset) if rank != -1 else None
    loader = torch.utils.data.DataLoader if image_weights else InfiniteDataLoader
    # Use torch.utils.data.DataLoader() if dataset.properties will update during training else InfiniteDataLoader()
    dataloader = loader(dataset,
                        batch_size=batch_size,
                        num_workers= nw,
                        sampler=sampler,
                        pin_memory=True,
                        collate_fn=LoadImagesAndLabels.collate_fn4 if quad else LoadImagesAndLabels.collate_fn)
    return dataloader, dataset

• terms of settlement
  take loader Function num_workers Parameter set to 0

• num_worker The role of
  1、 Every turn dataloader When loading data ：dataloader One-time creation num_worker individual worker,（ It can also be said that dataloader One-time creation num_worker Working process ,worker It's also an ordinary work process ）, And use batch_sampler Will specify batch Assigned to the designated worker,worker Put it in charge of batch Load in RAM. then ,dataloader from RAM Find out what to use in this iteration batch, If you find it , Just use . If not , Will be num_worker individual worker Continue loading batch To the memory , until dataloader stay RAM Find the target in batch. In general, it can be found , because batch_sampler Appoint batch Of course, priority should be given to the... To be used in this round batch.
  2、num_worker Set it big , The advantage is to find batch Fast , Because the next iteration is batch Probably in the last round / Last round … The iteration is already loaded . The disadvantage is that the memory overhead is high , It's aggravating CPU burden （worker Load data to RAM The process is CPU Copy it ）.num_workers The experience setting value is your own computer / Server's CPU The core number , If CPU strong 、RAM And it's abundant , You can make it bigger .
  3、 If num_worker Set to 0, It means every iteration ,dataloader No more self loading data to RAM This step （ Because no worker 了 ）, But in RAM In looking for batch, When it cannot be found, load the corresponding batch. The disadvantage, of course, is that it's slower .