python - PyTorch next(iter(training_loader)) 极慢，简单的数据，num_workers 不行吗？

Question

这里 x_dat 和 y_dat 只是很长的一维张量。

class FunctionDataset(Dataset):
    def __init__(self):
        x_dat, y_dat = data_product()

        self.length = len(x_dat)
        self.y_dat = y_dat
        self.x_dat = x_dat

    def __getitem__(self, index):
        sample = self.x_dat[index]
        label = self.y_dat[index]
        return sample, label

    def __len__(self):
        return self.length

...

data_set = FunctionDataset()

...

training_sampler = SubsetRandomSampler(train_indices)
validation_sampler = SubsetRandomSampler(validation_indices)

training_loader = DataLoader(data_set, sampler=training_sampler, batch_size=params['batch_size'], shuffle=False)
validation_loader = DataLoader(data_set, sampler=validation_sampler, batch_size=valid_size, shuffle=False)

我也试过固定两个加载程序的内存。将 num_workers 设置为 > 0 会导致进程之间出现运行时错误(例如 EOF 错误和中断错误)。我得到我的批处理:

x_val, target = next(iter(training_loader))

整个数据集将适合内存/gpu，但我想为这个实验模拟批处理。分析我的过程给我以下信息:

16276989 function calls (16254744 primitive calls) in 38.779 seconds

   Ordered by: cumulative time

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
   1745/1    0.028    0.000   38.780   38.780 {built-in method builtins.exec}
        1    0.052    0.052   38.780   38.780 simple aprox.py:3(<module>)
        1    0.000    0.000   36.900   36.900 simple aprox.py:519(exploreHeatmap)
        1    0.000    0.000   36.900   36.900 simple aprox.py:497(optFromSample)
        1    0.033    0.033   36.900   36.900 simple aprox.py:274(train)
  705/483    0.001    0.000   34.495    0.071 {built-in method builtins.next}
      222    1.525    0.007   34.493    0.155 dataloader.py:311(__next__)
      222    0.851    0.004   12.752    0.057 dataloader.py:314(<listcomp>)
  3016001   11.901    0.000   11.901    0.000 simple aprox.py:176(__getitem__)
       21    0.010    0.000   10.891    0.519 simple aprox.py:413(validationError)
      443    1.380    0.003    9.664    0.022 sampler.py:136(__iter__)
  663/221    2.209    0.003    8.652    0.039 dataloader.py:151(default_collate)
      221    0.070    0.000    6.441    0.029 dataloader.py:187(<listcomp>)
      442    6.369    0.014    6.369    0.014 {built-in method stack}
  3060221    2.799    0.000    5.890    0.000 sampler.py:68(<genexpr>)
  3060000    3.091    0.000    3.091    0.000 tensor.py:382(<lambda>)
      222    0.001    0.000    1.985    0.009 sampler.py:67(__iter__)
      222    1.982    0.009    1.982    0.009 {built-in method randperm}
  663/221    0.002    0.000    1.901    0.009 dataloader.py:192(pin_memory_batch)
      221    0.000    0.000    1.899    0.009 dataloader.py:200(<listcomp>)
....

表明与我实验的其余事件(训练模型和许多其他计算等)相比，数据加载器非常慢。出了什么问题，加快速度的最佳方法是什么？

Answer 1

检索批处理时

x, y = next(iter(training_loader))

您实际上在每次调用时都创建了一个新的数据加载器迭代器实例(!)参见this thread获取更多信息。
您应该做的是创建迭代器一次(每个纪元):

training_loader_iter = iter(training_loader)

然后为迭代器上的每个批处理调用next

for i in range(num_batches_in_epoch):
  x, y = next(training_loader_iter)

我之前遇到过类似的问题，这也让您在使用多个 worker 时遇到的 EOF 错误消失了。