I have writen a sample code to test split op based on the develop branch:

import paddle.fluid as fluid
input = fluid.layers.data(shape=[-1, 4], dtype="float32")
x0, x1 = fluid.layers.split(input, num_or_sections=2, dim=0)
print x0.shape
print x1.shape

but no error message found. Which version of paddle do you use?

I have writen a sample code to test split op based on the develop branch:

import paddle.fluid as fluid
input = fluid.layers.data(shape=[-1, 4], dtype="float32")
x0, x1 = fluid.layers.split(input, num_or_sections=2, dim=0)
print x0.shape
print x1.shape

but no error message found. Which version of paddle do you use?

you should test with num_or_sections=[2, 2]

import paddle.fluid as fluid
input = fluid.layers.data(shape=[-1, 4], dtype="float32", name='data')
x0, x1 = fluid.layers.split(input, num_or_sections=[2, 2], dim=0)
print x0.shape
print x1.shape

reference: https://www.paddlepaddle.org.cn/documentation/docs/zh/api_cn/layers_cn/split_cn.html#split
if your batch size is fixed as 4=2+2, you would better fix the batch size in data

reference: https://www.paddlepaddle.org.cn/documentation/docs/zh/api_cn/layers_cn/split_cn.html#split
if your batch size is fixed as 4=2+2, you would better fix the batch size in data

i use pyreader to build network. if i fix the batch size in the pyreader, it will produce tensor with 5 dimension. Because it will append -1 to the shapes.

test code：

import numpy as np
import paddle
import paddle.fluid as fluid


queue_capacity = 64
batch_size = -1
#batch_size = 8
image_shape = [3, 224, 224]
py_reader = fluid.layers.py_reader(
    capacity=queue_capacity,
    shapes=[[batch_size] + image_shape, [batch_size, 1]],
    lod_levels=[0, 0],
    dtypes=["float32", "int64"],
    use_double_buffer=True)
image, label = fluid.layers.read_file(py_reader)

conv = fluid.layers.conv2d(image, 16, 3, 2, 1)
output, _ = fluid.layers.split(conv, num_or_sections=[4, 4], dim=0)

print output.shape
fluid.layers.Print(output)

def reader():
    for _ in xrange(100):
        image = np.random.rand(3, 224, 224)
        label = 0
        yield (image, label)

train_reader = paddle.batch(reader, batch_size=8, drop_last=True)
py_reader.decorate_paddle_reader(train_reader)
py_reader.start()
fetch_list = [output.name]

place = fluid.CUDAPlace(0)
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())

res = exe.run(
    fluid.default_main_program(),
    fetch_list=fetch_list
)
print res[0].shape

In summary, user need to specify a batch size with pyreader, but the pyreader always appends a dimension if the shape[0] is not -1, which is not as expected

which

right.

https://www.paddlepaddle.org.cn/documentation/docs/zh/api_cn/io_cn/DataLoader_cn.html#dataloader
suggest DataLoader