pytorch.utils.data
可兼容迭代数据训练处理,在dataloader中使用提高训练效率:借助迭代器避免内存溢出不足的现象、借助链式处理使得数据读取利用更高效(可类比操作系统的资源调控)
书接上文,使用迭代器链式处理数据,在Process
类的__iter__
方法中执行挂载的预处理方法,可以嵌套包裹多层处理方法,类似KoaJs洋葱模型,在for循环时,自动执行预处理方法返回处理后的数据
分析下述示例中输入数据依次执行顺序:travel -> deep -> shuffle -> sort -> batch
,实际由于嵌套循环或设置缓存的存在,数据流式会有变化,具体如后图分析
from torch.utils.data import IterableDataset
# ...
import random
class Process(IterableDataset):
def __init__(self, data, f):
self.data = data
# 绑定处理函数
self.f = f
def __iter__(self):
# for循环遍历时,返回一个当前环节处理的迭代器对象
return self.f(iter(self.data))
a = ['a0', 'a1', 'a2', 'a3', 'a4', 'a5', 'a6', 'a7', 'a8', 'a9']
b = ['b0', 'b1', 'b2', 'b3', 'b4', 'b5', 'b6', 'b7', 'b8', 'b9']
c = ['c0', 'c1', 'c2', 'c3', 'c4', 'c5', 'c6', 'c7', 'c8', 'c9']
# data = [[j + str(i) for i in range(10)] for j in ['a','b', 'c'] ]
data = [a, b, c]
def travel(d):
for i in d:
# print('travel ', i)
yield i
def deep(d):
for arr in d:
for item in arr:
yield item
def shuffle(d, sf_size=5):
buf = []
for i in d:
buf.append(i)
if len(buf) >= sf_size:
random.shuffle(buf)
for j in buf:
# print('shuffle', j)
yield j
buf = []
for k in buf:
yield k
def sort(d):
buf = []
for i in d:
buf.append(i)
if len(buf) >= 3:
for i in buf:
# print('sort', i)
yield i
buf = []
for k in buf:
yield k
def batch(d):
buf = []
for i in d:
buf.append(i)
if len(buf) >= 16:
for i in buf:
# print('batch', i)
yield i
buf = []
# 对训练数据进行的多个预处理步骤
dataset = Process(data, travel)
dataset = Process(dataset , deep)
dataset = Process(dataset , shuffle)
dataset = Process(dataset , sort)
train_dataset = Process(p, batch)
# 可在此处断点测试
for i in p:
print(i, 'train')
# train_data_loader = DataLoader(train_dataset,num_workers=args.num_workers,prefetch_factor=args.prefetch)
# train(model , train_data_loader)
由上可以构造数据流式方向 :batch(iter(sort(iter(shuffle(iter(deep(iter(travel(iter( d ))))))))))
根据数据流式抽取部分过程画出时序图如下: