paddlepaddle中动态图查看训练参数(二)

当前版本 Paddlepaddle1.8

我的模型继承自fluid.dygraph.Layer
比如:

from net import DenseNet  # <------- 该类为已经写好的`DenseNet`
#定义网络
class MyCNN(fluid.dygraph.Layer):
    '''
    卷积神经网络
    '''
    def __init__(self):
        super(MyCNN,self).__init__()
        self.my_densenet = DenseNet(
                name_scope='DenseNet',
                layers=264,            # 使用 `DenseNet-264`
                dropout_prob=0.2,
                class_dim=5
            )
    def forward(self,inputs):
        """前向计算"""
        return self.my_densenet(inputs)

model = MyCNN()

我新建了一个MyCNN类，继承自fluid.dygraph.Layer，model为MyCNN类的实例化对象

通过上一篇博客:
https://blog.csdn.net/HaoZiHuang/article/details/107175271
我们知道了可以用model.parameters方法来查看参数，但是他是列表不太方便

通过github issue:
https://github.com/PaddlePaddle/Paddle/issues/25719
查阅到还有另一个方法model.state_dict来获取，顾名思义，这次返回的字典，他的键是model.parameters()元素的name属性，那这样就方便多了

煮个栗子:

>>> para_dict = cnn.state_dict()
>>> para_dict.keys()
odict_keys(['my_densenet.conv1.weight', 
'my_densenet.bb_0_0.denseblock.block.bn_conv1._conv.weight', 
'my_densenet.bb_0_0.denseblock.block.bn_conv1._batch_norm.weight', 
'my_densenet.bb_0_0.denseblock.block.bn_conv1._batch_norm.bias', 
'my_densenet.bb_0_0.denseblock.block.bn_conv1._batch_norm._mean', 
'my_densenet.bb_0_0.denseblock.block.bn_conv1._batch_norm._variance', 
'my_densenet.bb_0_0.denseblock.block.bn_conv2._conv.weight', 
.....

>>> # 可以看到的键是每个`parameter`的`name`熟悉
>>> para_dict['my_densenet.bb_0_0.denseblock.block.bn_conv1._conv.weight']
name conv2d_1.w_0, dtype: VarType.FP32 shape: [128, 64, 1, 1] 	lod: {}
	dim: 128, 64, 1, 1
	layout: NCHW
	dtype: float
	data: [0.229974 -0.0597555 0.167087 0.124216 0.00420646 -0.0976657 0.168332 ......

可以看到和model.parameters的元素打印出来格式相同