paddle数据Tensor

1.根据输入数据构建loader

　　

src = fluid.data(name="src", shape=[None, None], dtype="int64")
    src_sequence_length = fluid.data(name="src_sequence_length",shape=[None],dtype="int64")
    inputs = [src, src_sequence_length]
loader = fluid.io.DataLoader.from_generator(feed_list=inputs,capacity=10,
                                                iterable=True,use_double_buffer=True)

loader.set_batch_generator(reader,places=places)

 

之后loader会将输入的Tensor自动转化为输入需要的lodTensor

 

2.program

train_prog = fluid.Program()
startup_prog = fluid.Program()
with fluid.program_guard(train_prog, startup_prog):
    with fluid.unique_name.guard():
        pass

默认情况下网络是在default_main_program中，如果要自定义program必须使用with结构