PyReader

class paddle.fluid.io.PyReader(feed_list=None, capacity=None, use_double_buffer=True, iterable=True, return_list=False)[源代码]

在python中为数据输入创建一个reader对象。将使用python线程预取数据,并将其异步插入队列。当调用Executor.run(…)时,将自动提取队列中的数据。

参数

  • feed_list (list(Variable)|tuple(Variable)) - feed变量列表,由 fluid.layers.data() 创建。
  • capacity (int) - PyReader对象内部维护队列的容量大小。单位是batch数量。若reader读取速度较快,建议设置较大的capacity值。
  • use_double_buffer (bool) - 是否使用 double_buffer_reader 。若use_double_buffer=True,PyReader会异步地预读取下一个batch的数据,可加速数据读取过程,但同时会占用少量的CPU/GPU存储,即一个batch输入数据的存储空间。
  • iterable (bool) - 所创建的DataLoader对象是否可迭代。
  • return_list (bool) - 每个设备上的数据是否以list形式返回。仅在iterable = True模式下有效。若return_list = False,每个设备上的返回数据均是str -> LoDTensor的映射表,其中映射表的key是每个输入变量的名称。若return_list = True,则每个设备上的返回数据均是list(LoDTensor)。推荐在静态图模式下使用return_list = False,在动态图模式下使用return_list = True。

返回

被创建的reader对象

返回类型

reader (Reader)

代码示例

1.如果iterable=False,则创建的PyReader对象几乎与 fluid.layers.py_reader() 相同。算子将被插入program中。用户应该在每个epoch之前调用 start() ,并在epoch结束时捕获 Executor.run() 抛出的 fluid.core.EOFException 。一旦捕获到异常,用户应该调用 reset() 手动重置reader。

  1. import paddle
  2. import paddle.fluid as fluid
  3. import numpy as np
  4. EPOCH_NUM = 3
  5. ITER_NUM = 5
  6. BATCH_SIZE = 3
  7. def network(image, label):
  8. # 用户定义网络,此处以softmax回归为例
  9. predict = fluid.layers.fc(input=image, size=10, act='softmax')
  10. return fluid.layers.cross_entropy(input=predict, label=label)
  11. def reader_creator_random_image_and_label(height, width):
  12. def reader():
  13. for i in range(ITER_NUM):
  14. fake_image = np.random.uniform(low=0,
  15. high=255,
  16. size=[height, width])
  17. fake_label = np.ones([1])
  18. yield fake_image, fake_label
  19. return reader
  20. image = fluid.layers.data(name='image', shape=[784, 784], dtype='float32')
  21. label = fluid.layers.data(name='label', shape=[1], dtype='int64')
  22. reader = fluid.io.PyReader(feed_list=[image, label],
  23. capacity=4,
  24. iterable=False)
  25. user_defined_reader = reader_creator_random_image_and_label(784, 784)
  26. reader.decorate_sample_list_generator(
  27. paddle.batch(user_defined_reader, batch_size=BATCH_SIZE))
  28. loss = network(image, label)
  29. executor = fluid.Executor(fluid.CPUPlace())
  30. executor.run(fluid.default_startup_program())
  31. for i in range(EPOCH_NUM):
  32. reader.start()
  33. while True:
  34. try:
  35. executor.run(feed=None)
  36. except fluid.core.EOFException:
  37. reader.reset()
  38. break

2.如果iterable=True,则创建的PyReader对象与程序分离。程序中不会插入任何算子。在本例中,创建的reader是一个python生成器,它是可迭代的。用户应将从PyReader对象生成的数据输入 Executor.run(feed=...)

  1. import paddle
  2. import paddle.fluid as fluid
  3. import numpy as np
  4. EPOCH_NUM = 3
  5. ITER_NUM = 5
  6. BATCH_SIZE = 10
  7. def network(image, label):
  8. # 用户定义网络,此处以softmax回归为例
  9. predict = fluid.layers.fc(input=image, size=10, act='softmax')
  10. return fluid.layers.cross_entropy(input=predict, label=label)
  11. def reader_creator_random_image(height, width):
  12. def reader():
  13. for i in range(ITER_NUM):
  14. fake_image = np.random.uniform(low=0, high=255, size=[height, width]),
  15. fake_label = np.ones([1])
  16. yield fake_image, fake_label
  17. return reader
  18. image = fluid.layers.data(name='image', shape=[784, 784], dtype='float32')
  19. label = fluid.layers.data(name='label', shape=[1], dtype='int64')
  20. reader = fluid.io.PyReader(feed_list=[image, label], capacity=4, iterable=True, return_list=False)
  21. user_defined_reader = reader_creator_random_image(784, 784)
  22. reader.decorate_sample_list_generator(
  23. paddle.batch(user_defined_reader, batch_size=BATCH_SIZE),
  24. fluid.core.CPUPlace())
  25. loss = network(image, label)
  26. executor = fluid.Executor(fluid.CPUPlace())
  27. executor.run(fluid.default_startup_program())
  28. for _ in range(EPOCH_NUM):
  29. for data in reader():
  30. executor.run(feed=data, fetch_list=[loss])
  1. return_list=True,返回值将用list表示而非dict,通常用于动态图模式中。
  1. import paddle
  2. import paddle.fluid as fluid
  3. import numpy as np
  4. EPOCH_NUM = 3
  5. ITER_NUM = 5
  6. BATCH_SIZE = 10
  7. def reader_creator_random_image(height, width):
  8. def reader():
  9. for i in range(ITER_NUM):
  10. yield np.random.uniform(low=0, high=255, size=[height, width]), \
  11. np.random.random_integers(low=0, high=9, size=[1])
  12. return reader
  13. place = fluid.CPUPlace()
  14. with fluid.dygraph.guard(place):
  15. py_reader = fluid.io.PyReader(capacity=2, return_list=True)
  16. user_defined_reader = reader_creator_random_image(784, 784)
  17. py_reader.decorate_sample_list_generator(
  18. paddle.batch(user_defined_reader, batch_size=BATCH_SIZE),
  19. place)
  20. for image, label in py_reader():
  21. relu = fluid.layers.relu(image)

启动数据输入线程。只能在reader对象不可迭代时调用。

代码示例

  1. import paddle
  2. import paddle.fluid as fluid
  3. import numpy as np
  4. BATCH_SIZE = 10
  5. def generator():
  6. for i in range(5):
  7. yield np.random.uniform(low=0, high=255, size=[784, 784]),
  8. image = fluid.layers.data(name='image', shape=[784, 784], dtype='float32')
  9. reader = fluid.io.PyReader(feed_list=[image], capacity=4, iterable=False)
  10. reader.decorate_sample_list_generator(
  11. paddle.batch(generator, batch_size=BATCH_SIZE))
  12. executor = fluid.Executor(fluid.CPUPlace())
  13. executor.run(fluid.default_startup_program())
  14. for i in range(3):
  15. reader.start()
  16. while True:
  17. try:
  18. executor.run(feed=None)
  19. except fluid.core.EOFException:
  20. reader.reset()
  21. break

reset()

fluid.core.EOFException 抛出时重置reader对象。只能在reader对象不可迭代时调用。

代码示例

  1. import paddle
  2. import paddle.fluid as fluid
  3. import numpy as np
  4. BATCH_SIZE = 10
  5. def generator():
  6. for i in range(5):
  7. yield np.random.uniform(low=0, high=255, size=[784, 784]),
  8. image = fluid.layers.data(name='image', shape=[784, 784], dtype='float32')
  9. reader = fluid.io.PyReader(feed_list=[image], capacity=4, iterable=False)
  10. reader.decorate_sample_list_generator(
  11. paddle.batch(generator, batch_size=BATCH_SIZE))
  12. executor = fluid.Executor(fluid.CPUPlace())
  13. executor.run(fluid.default_startup_program())
  14. for i in range(3):
  15. reader.start()
  16. while True:
  17. try:
  18. executor.run(feed=None)
  19. except fluid.core.EOFException:
  20. reader.reset()
  21. break

decorate_sample_generator(sample_generator, batch_size, drop_last=True, places=None)

设置PyReader对象的数据源。

提供的 sample_generator 应该是一个python生成器,它生成的数据类型应为list(numpy.ndarray)。

当PyReader对象可迭代时,必须设置 places

如果所有的输入都没有LOD,这个方法比 decorate_sample_list_generator(paddle.batch(sample_generator, ...)) 更快。

参数

  • sample_generator (generator) – Python生成器,yield 类型为list(numpy.ndarray)
  • batch_size (int) – batch size,必须大于0
  • drop_last (bool) – 当样本数小于batch数量时,是否删除最后一个batch
  • places (None|list(CUDAPlace)|list(CPUPlace)) – 位置列表。当PyReader可迭代时必须被提供

代码示例

  1. import paddle.fluid as fluid
  2. import numpy as np
  3. EPOCH_NUM = 3
  4. ITER_NUM = 15
  5. BATCH_SIZE = 3
  6. def network(image, label):
  7. # 用户定义网络,此处以softmax回归为例
  8. predict = fluid.layers.fc(input=image, size=10, act='softmax')
  9. return fluid.layers.cross_entropy(input=predict, label=label)
  10. def random_image_and_label_generator(height, width):
  11. def generator():
  12. for i in range(ITER_NUM):
  13. fake_image = np.random.uniform(low=0,
  14. high=255,
  15. size=[height, width])
  16. fake_label = np.array([1])
  17. yield fake_image, fake_label
  18. return generator
  19. image = fluid.layers.data(name='image', shape=[784, 784], dtype='float32')
  20. label = fluid.layers.data(name='label', shape=[1], dtype='int64')
  21. reader = fluid.io.PyReader(feed_list=[image, label], capacity=4, iterable=True)
  22. user_defined_generator = random_image_and_label_generator(784, 784)
  23. reader.decorate_sample_generator(user_defined_generator,
  24. batch_size=BATCH_SIZE,
  25. places=[fluid.CPUPlace()])
  26. loss = network(image, label)
  27. executor = fluid.Executor(fluid.CPUPlace())
  28. executor.run(fluid.default_startup_program())
  29. for _ in range(EPOCH_NUM):
  30. for data in reader():
  31. executor.run(feed=data, fetch_list=[loss])

decorate_sample_list_generator(reader, places=None)

设置PyReader对象的数据源。

提供的 reader 应该是一个python生成器,它生成列表(numpy.ndarray)类型的批处理数据。

当PyReader对象不可迭代时,必须设置 places

参数

  • reader (generator) – 返回列表(numpy.ndarray)类型的批处理数据的Python生成器
  • places (None|list(CUDAPlace)|list(CPUPlace)) – 位置列表。当PyReader可迭代时必须被提供

代码示例

  1. import paddle
  2. import paddle.fluid as fluid
  3. import numpy as np
  4. EPOCH_NUM = 3
  5. ITER_NUM = 15
  6. BATCH_SIZE = 3
  7. def network(image, label):
  8. # 用户定义网络,此处以softmax回归为例
  9. predict = fluid.layers.fc(input=image, size=10, act='softmax')
  10. return fluid.layers.cross_entropy(input=predict, label=label)
  11. def random_image_and_label_generator(height, width):
  12. def generator():
  13. for i in range(ITER_NUM):
  14. fake_image = np.random.uniform(low=0,
  15. high=255,
  16. size=[height, width])
  17. fake_label = np.ones([1])
  18. yield fake_image, fake_label
  19. return generator
  20. image = fluid.layers.data(name='image', shape=[784, 784], dtype='float32')
  21. label = fluid.layers.data(name='label', shape=[1], dtype='int64')
  22. reader = fluid.io.PyReader(feed_list=[image, label], capacity=4, iterable=True)
  23. user_defined_generator = random_image_and_label_generator(784, 784)
  24. reader.decorate_sample_list_generator(
  25. paddle.batch(user_defined_generator, batch_size=BATCH_SIZE),
  26. fluid.core.CPUPlace())
  27. loss = network(image, label)
  28. executor = fluid.Executor(fluid.core.CPUPlace())
  29. executor.run(fluid.default_startup_program())
  30. for _ in range(EPOCH_NUM):
  31. for data in reader():
  32. executor.run(feed=data, fetch_list=[loss])

decorate_batch_generator(reader, places=None)

设置PyReader对象的数据源。

提供的 reader 应该是一个python生成器,它生成列表(numpy.ndarray)类型或LoDTensor类型的批处理数据。

当PyReader对象不可迭代时,必须设置 places

参数

  • reader (generator) – 返回LoDTensor类型的批处理数据的Python生成器
  • places (None|list(CUDAPlace)|list(CPUPlace)) – 位置列表。当PyReader可迭代时必须被提供

代码示例

  1. import paddle.fluid as fluid
  2. import numpy as np
  3. EPOCH_NUM = 3
  4. ITER_NUM = 15
  5. BATCH_SIZE = 3
  6. def network(image, label):
  7. # 用户定义网络,此处以softmax回归为例
  8. predict = fluid.layers.fc(input=image, size=10, act='softmax')
  9. return fluid.layers.cross_entropy(input=predict, label=label)
  10. def random_image_and_label_generator(height, width):
  11. def generator():
  12. for i in range(ITER_NUM):
  13. batch_image = np.random.uniform(low=0,
  14. high=255,
  15. size=[BATCH_SIZE, height, width])
  16. batch_label = np.ones([BATCH_SIZE, 1])
  17. batch_image = batch_image.astype('float32')
  18. batch_label = batch_label.astype('int64')
  19. yield batch_image, batch_label
  20. return generator
  21. image = fluid.layers.data(name='image', shape=[784, 784], dtype='float32')
  22. label = fluid.layers.data(name='label', shape=[1], dtype='int64')
  23. reader = fluid.io.PyReader(feed_list=[image, label], capacity=4, iterable=True)
  24. user_defined_generator = random_image_and_label_generator(784, 784)
  25. reader.decorate_batch_generator(user_defined_generator, fluid.CPUPlace())
  26. loss = network(image, label)
  27. executor = fluid.Executor(fluid.CPUPlace())
  28. executor.run(fluid.default_startup_program())
  29. for _ in range(EPOCH_NUM):
  30. for data in reader():
  31. executor.run(feed=data, fetch_list=[loss])

next()

获取下一个数据。用户不应直接调用此方法。此方法用于PaddlePaddle框架内部实现Python 2.x的迭代器协议。