未验证 提交 3b543756 编写于 作者: C Cao Ying 提交者: GitHub

Merge pull request #7241 from will-am/fix_doc_paramattr

Fix doc of ParameterAttribute.
......@@ -58,12 +58,12 @@ def is_compatible_with(x, Type):
class HookAttribute(object):
"""
Hook Attribute object. As a member of ParameterAttribute class, the hook is an auxiliary operation that occurs
Hook Attribute object. As a member of ParameterAttribute class, the hook is an auxiliary operation that occurs
during training process of a layer with parameters, such as img_conv layer, fc layer.
:param type: Hook type, currently supported types:
:param type: Hook type, currently supported types:
'pruning' : user specify a sparsity_ratio before training started, and the
network will prune the parameters based on the sparsity_ratio.
network will prune the parameters based on the sparsity_ratio.
eg: The definition of Hook object can be hk = HookAttribute('pruning', 0.6)
The specific usage can be paddle.layer.img_conv(input=img, filter_size=3,
num_channels=3, num_filters=64,
......@@ -71,10 +71,10 @@ class HookAttribute(object):
The pruning details can be found https://arxiv.org/pdf/1506.02626.pdf
:type type: string
:param sparsity_ratio: Must be specified if hook type is 'pruning',
:param sparsity_ratio: Must be specified if hook type is 'pruning',
it represents the ratio of the zero elements to be set by the Parameter.
:type sparsity_ratio: float or None
"""
def __init__(self, type, sparsity_ratio=None):
......@@ -130,10 +130,12 @@ class ParameterAttribute(object):
:param sparse_update: Enable sparse update for this parameter. It will
enable both local and remote sparse update.
:type sparse_update: bool
:param update_hooks: A HookAttribute object.
:type update_hooks: HookAttribute
:param initializer: If not None, it should be a callable object which accepts
a parameter name and returns numpy array for the initial
value of the parameter
:param initializer: callable object
:type initializer: callable object
"""
def __init__(self,
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册