Paddle的 Dataloader 遇到纯文本数据时报错 StopIteration

[cocoshe]

bug描述 Describe the Bug

Paddle的 Dataloader 遇到纯文本数据时报错 StopIteration

看上去Paddle的Dataloader并不支持纯str类型的数据，而Pytorch则不然

下面是Dataloader文档中的样例代码，后面代码都是基于这份代码来做复现。

• 样例代码

import numpy as np

import paddle
import paddle.nn as nn
import paddle.nn.functional as F
from paddle.io import Dataset, BatchSampler, DataLoader

BATCH_NUM = 20
BATCH_SIZE = 16
EPOCH_NUM = 4

IMAGE_SIZE = 784
CLASS_NUM = 10

# define a random dataset
class RandomDataset(Dataset):
    def __init__(self, num_samples):
        self.num_samples = num_samples

    def __getitem__(self, idx):
        image = np.random.random([IMAGE_SIZE]).astype('float32')
        label = np.random.randint(0, CLASS_NUM - 1, (1, )).astype('int64')
        return image, label

    def __len__(self):
        return self.num_samples

dataset = RandomDataset(BATCH_NUM * BATCH_SIZE)

class SimpleNet(nn.Layer):
    def __init__(self):
        super().__init__()
        self.fc = nn.Linear(IMAGE_SIZE, CLASS_NUM)

    def forward(self, image, label=None):
        return self.fc(image)

simple_net = SimpleNet()
opt = paddle.optimizer.SGD(learning_rate=1e-3,
                            parameters=simple_net.parameters())

loader = DataLoader(dataset,
                    batch_size=BATCH_SIZE,
                    shuffle=True,
                    drop_last=True,
                    num_workers=2)

for e in range(EPOCH_NUM):
    for i, (image, label) in enumerate(loader()):
        out = simple_net(image)
        loss = F.cross_entropy(out, label)
        avg_loss = paddle.mean(loss)
        avg_loss.backward()
        opt.minimize(avg_loss)
        simple_net.clear_gradients()
        print("Epoch {} batch {}: loss = {}".format(e, i, np.mean(loss.numpy())))

当dataset中的数据不是tensor，而是str类型时，无法成功load进数据

• Paddle：

import numpy as np

import paddle
import paddle.nn as nn
import paddle.nn.functional as F
from paddle.io import Dataset, BatchSampler, DataLoader

BATCH_NUM = 20
BATCH_SIZE = 16
EPOCH_NUM = 4

IMAGE_SIZE = 784
CLASS_NUM = 10

# define a random dataset
class RandomDataset(Dataset):
    def __init__(self, num_samples):
        self.num_samples = num_samples

    def __getitem__(self, idx):                             #  改动基本就是这里，原本的np array 变成现在的 str 类型
        # image = np.random.random([IMAGE_SIZE]).astype('float32')
        # label = np.random.randint(0, CLASS_NUM - 1, (1, )).astype('int64')
        image = "This is a string format data"
        label = "Also a string format data"

        return image, label

    def __len__(self):
        return self.num_samples

dataset = RandomDataset(BATCH_NUM * BATCH_SIZE)

class SimpleNet(nn.Layer):
    def __init__(self):
        super().__init__()
        self.fc = nn.Linear(IMAGE_SIZE, CLASS_NUM)

    def forward(self, image, label=None):
        return self.fc(image)



simple_net = SimpleNet()
opt = paddle.optimizer.SGD(learning_rate=1e-3,
                            parameters=simple_net.parameters())

loader = DataLoader(dataset,
                    batch_size=BATCH_SIZE,
                    shuffle=True,
                    drop_last=True,)



for e in range(EPOCH_NUM):
    import pdb
    pdb.set_trace()
    for i, (image, label) in enumerate(loader):
        out = simple_net(image)
        loss = F.cross_entropy(out, label)
        avg_loss = paddle.mean(loss)
        avg_loss.backward()
        opt.minimize(avg_loss)
        simple_net.clear_gradients()
        print("Epoch {} batch {}: loss = {}".format(e, i, np.mean(loss.numpy())))

在内层 for 去读取 dataloader 的数据的时候，报错：

(exp) coco@coco:~$ python mytest.py
> /home/coco/mytest.py(71)<module>()
-> for i, (image, label) in enumerate(loader):
(Pdb) n
StopIteration

• Pytorch

单纯把上面 paddle 的例子中相关api换成 torch 中的 api

import numpy as np

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import Dataset, BatchSampler, DataLoader

BATCH_NUM = 20
BATCH_SIZE = 16
EPOCH_NUM = 4

IMAGE_SIZE = 784
CLASS_NUM = 10

# define a random dataset
class RandomDataset(Dataset):
    def __init__(self, num_samples):
        self.num_samples = num_samples

    def __getitem__(self, idx):
        # image = np.random.random([IMAGE_SIZE]).astype('float32')
        # label = np.random.randint(0, CLASS_NUM - 1, (1, )).astype('int64')
        image = "This is a string format data"
        label = "Also a string format data"

        return image, label

    def __len__(self):
        return self.num_samples

dataset = RandomDataset(BATCH_NUM * BATCH_SIZE)

class SimpleNet(nn.Module):
    def __init__(self):
        super().__init__()
        self.fc = nn.Linear(IMAGE_SIZE, CLASS_NUM)

    def forward(self, image, label=None):
        return self.fc(image)

simple_net = SimpleNet()
opt = torch.optim.SGD(lr=1e-3,params=simple_net.parameters())
loader = DataLoader(dataset,
                    batch_size=BATCH_SIZE,
                    shuffle=True,
                    drop_last=True,
                    num_workers=2)

for e in range(EPOCH_NUM):
    import pdb
    pdb.set_trace()
    for i, (image, label) in enumerate(loader):
        out = simple_net(image)
        loss = F.cross_entropy(out, label)
        avg_loss = paddle.mean(loss)
        avg_loss.backward()
        opt.minimize(avg_loss)
        simple_net.clear_gradients()
        print("Epoch {} batch {}: loss = {}".format(e, i, np.mean(loss.numpy())))

输出：

(exp) coco@coco:~$ python torch_mytest.py
> /home/coco/torch_mytest.py(52)<module>()
-> for i, (image, label) in enumerate(loader):
(Pdb) n
> /home/coco/torch_mytest.py(53)<module>()
-> out = simple_net(image)
(Pdb) l
 48
 49     for e in range(EPOCH_NUM):
 50         import pdb
 51         pdb.set_trace()
 52         for i, (image, label) in enumerate(loader):
 53  ->         out = simple_net(image)
 54             loss = F.cross_entropy(out, label)
 55             avg_loss = paddle.mean(loss)
 56             avg_loss.backward()
 57             opt.minimize(avg_loss)
 58             simple_net.clear_gradients()
(Pdb) image
('This is a string format data', 'This is a string format data', 'This is a string format data', 'This is a string format data', 'This is a string format data', 'This is a string format data', 'This is a string format data', 'This is a string format data', 'This is a string format data', 'This is a string format data', 'This is a string format data', 'This is a string format data', 'This is a string format data', 'This is a string format data', 'This is a string format data', 'This is a string format data')
(Pdb) label
('Also a string format data', 'Also a string format data', 'Also a string format data', 'Also a string format data', 'Also a string format data', 'Also a string format data', 'Also a string format data', 'Also a string format data', 'Also a string format data', 'Also a string format data', 'Also a string format data', 'Also a string format data', 'Also a string format data', 'Also a string format data', 'Also a string format data', 'Also a string format data')
(Pdb)

Pytorch可以正常load进数据。

上面的复现过程中仅针对数据load的部分，不考虑后续计算

但是如果把数据部分改成这样：

    def __getitem__(self, idx):
        # image = np.random.random([IMAGE_SIZE]).astype('float32')
        label = np.random.randint(0, CLASS_NUM - 1, (1, )).astype('int64')
        image = "This is a string format data"
        # label = "Also a string format data"

        return image, label

也就是有文本，也有ndarray，又可以正常读到数据。

所以看上去Paddle的dataloader遇到纯str类型数据会有有问题，而且输出仅仅是一个StopIteration，难定位问题，对用户不友好。

感觉会是这里的原因吗？：

Paddle/python/paddle/io/dataloader/dataloader_iter.py

Lines 219 to 277 in 0359685

	def _thread_loop(self, legacy_expected_place):
	# NOTE(zhiqiu): Set the expected place for new thread as the same as father thread,
	# and it will call platform::SetDeviceId() in c++ internally.
	# If we do not set cudaDeviceId in new thread, the default cudaDeviceId will be 0,
	# Which may cost hundreds of MB of GPU memory on CUDAPlace(0) if calling some cuda
	# APIs in this thread.
	core.set_current_thread_name("Dataloader_" + str(id(self)))
	_set_expected_place(legacy_expected_place)
	while not self._thread_done_event.is_set():
	try:
	indices = next(self._sampler_iter)
	# read data from dataset in mini-batch
	# with paddle.base.dygraph.guard(place=paddle.CPUPlace()):
	# read data from dataset in mini-batch
	batch = self._dataset_fetcher.fetch(
	indices, self._thread_done_event
	)
	except StopIteration:
	self._exit_thread_expectedly()
	return
	if batch is None or self._thread_done_event.is_set():
	break
	# flat batch and record structure infos
	batch, structure = _flatten_batch(batch)
	self._structure_infos.append(structure)
	if self._thread_done_event.is_set():
	break
	try:
	# pack as LoDTensorArray
	array = core.LoDTensorArray()
	for slot in batch:
	if isinstance(slot, (paddle.Tensor, core.eager.Tensor)):
	slot = slot.value().get_tensor()
	elif not isinstance(slot, core.LoDTensor):
	tmp = core.LoDTensor()
	tmp.set(slot, core.CPUPlace())
	slot = tmp
	array.append(slot)
	if self._thread_done_event.is_set():
	break
	try:
	self._blocking_queue.push(array)
	except:
	self._exit_thread_expectedly()
	except Exception as e:
	self._exit_thread_unexpectedly()
	raise e
	self._exit_thread_expectedly()

中的_flatten_batch函数里面：

Paddle/python/paddle/io/dataloader/flat.py

Lines 37 to 43 in 0359685

	if isinstance(
	field,
	(np.ndarray, paddle.Tensor, paddle.base.core.eager.Tensor),
	):
	structure.append(f'{FIELD_PREFIX}{field_idx}')
	flat_batch.append(field)
	field_idx += 1

这里用了自己递归的方式，对最内层数据(np.ndarray, paddle.Tensor, paddle.base.core.eager.Tensor) 这些类型提取并存在了flat_batch，也就是说，遇到例如纯str类型的数据时，而structure则保存了tensor和ndarray之外的信息以及给tensor和ndarray的值留了占位符，但是此时为纯文本，导致flat_batch一直都会为空，不知道是不是这里有影响emm

• 具体遇到的情景：
  在一些LLM模型中，它可能在Pytorch实现过程中没有很规范地去写processor或者collator（尤其是tokenize的过程），直接把这些模块放在了组网的过程中，这就意味着dataloader迭代得到数据会是纯文本，这就导致Paddle的dataloader出现上述问题，Paddle中必须写collator把纯文本做一下tokenize，让input_ids、attention_mask等最内层都是tensor或者np array，这样才能顺利被迭代读取。

但是至于为什么会抛出StopIteration，我还是比较困惑，希望后续如果可以成功复现，定位问题后可以在这个issue下告知呀~非常感谢！！！

其他补充信息 Additional Supplementary Information

No response

[yangjianfengo1]

已收到您的反馈，我们马上排查。