ParameterList

  • class paddle.fluid.dygraph.ParameterList(parameters=None)[源代码]

参数列表容器。此容器的行为类似于Python列表，但它包含的参数将被正确地注册和添加。

  • 参数：
    • parameters (iterable，可选) - 可迭代的Parameters。

返回：无

代码示例

  1. import paddle.fluid as fluid
  2. import numpy as np
  3.  
  4. class MyLayer(fluid.Layer):
  5.     def __init__(self, num_stacked_param):
  6.         super(MyLayer, self).__init__()
  7.         # 使用可迭代的 Parameters 创建 ParameterList
  8.         self.params = fluid.dygraph.ParameterList(
  9.             [fluid.layers.create_parameter(
  10.                 shape=[2, 2], dtype='float32')] * num_stacked_param)
  11.     def forward(self, x):
  12.         for i, p in enumerate(self.params):
  13.             tmp = self._helper.create_variable_for_type_inference('float32')
  14.             self._helper.append_op(
  15.                 type="mul",
  16.                 inputs={"X": x,
  17.                         "Y": p},
  18.                 outputs={"Out": tmp},
  19.                 attrs={"x_num_col_dims": 1,
  20.                        "y_num_col_dims": 1})
  21.             x = tmp
  22.         return x
  23.  
  24. data_np = np.random.uniform(-1, 1, [5, 2]).astype('float32')
  25. with fluid.dygraph.guard():
  26.     x = fluid.dygraph.to_variable(data_np)
  27.     num_stacked_param = 4
  28.     model = MyLayer(num_stacked_param)
  29.     print(len(model.params))  # 4
  30.     res = model(x)
  31.     print(res.shape)  # [5, 2]
  32.     replaced_param = fluid.layers.create_parameter(shape=[2, 3], dtype='float32')
  33.     model.params[num_stacked_param - 1] = replaced_param  # 替换最后一个参数
  34.     res = model(x)
  35.     print(res.shape)  # [5, 3]
  36.     model.params.append(fluid.layers.create_parameter(shape=[3, 4], dtype='float32'))  # 添加参数
  37.     print(len(model.params))  # 5
  38.     res = model(x)
  39.     print(res.shape)  # [5, 4]