forked from TensorLayer/tensorlayer3
126 lines
4.1 KiB
Python
126 lines
4.1 KiB
Python
#! /usr/bin/python
|
|
# -*- coding: utf-8 -*-
|
|
|
|
import numbers
|
|
import tensorlayer as tl
|
|
from tensorlayer import logging
|
|
from tensorlayer.layers.core import Module
|
|
|
|
__all__ = [
|
|
'DropconnectDense',
|
|
]
|
|
|
|
|
|
class DropconnectDense(Module):
|
|
"""
|
|
The :class:`DropconnectDense` class is :class:`Dense` with DropConnect
|
|
behaviour which randomly removes connections between this layer and the previous
|
|
layer according to a keeping probability.
|
|
|
|
Parameters
|
|
----------
|
|
keep : float
|
|
The keeping probability.
|
|
The lower the probability it is, the more activations are set to zero.
|
|
n_units : int
|
|
The number of units of this layer.
|
|
act : activation function
|
|
The activation function of this layer.
|
|
W_init : weights initializer
|
|
The initializer for the weight matrix.
|
|
b_init : biases initializer
|
|
The initializer for the bias vector.
|
|
in_channels: int
|
|
The number of channels of the previous layer.
|
|
If None, it will be automatically detected when the layer is forwarded for the first time.
|
|
name : str
|
|
A unique layer name.
|
|
|
|
Examples
|
|
--------
|
|
>>> net = tl.layers.Input([10, 784], name='input')
|
|
>>> net = tl.layers.DropconnectDense(keep=0.8, n_units=800, act=tl.ReLU, name='relu1')(net)
|
|
>>> output shape :(10, 800)
|
|
>>> net = tl.layers.DropconnectDense(keep=0.5, n_units=800, act=tl.ReLU, name='relu2')(net)
|
|
>>> output shape :(10, 800)
|
|
>>> net = tl.layers.DropconnectDense(keep=0.5, n_units=10, name='output')(net)
|
|
>>> output shape :(10, 10)
|
|
|
|
References
|
|
----------
|
|
- `Wan, L. (2013). Regularization of neural networks using dropconnect <http://machinelearning.wustl.edu/mlpapers/papers/icml2013_wan13>`__
|
|
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
keep=0.5,
|
|
n_units=100,
|
|
act=None,
|
|
W_init=tl.initializers.truncated_normal(stddev=0.05),
|
|
b_init=tl.initializers.constant(value=0.0),
|
|
in_channels=None,
|
|
name=None, # 'dropconnect',
|
|
):
|
|
super().__init__(name, act=act)
|
|
|
|
if isinstance(keep, numbers.Real) and not (keep > 0 and keep <= 1):
|
|
raise ValueError("keep must be a scalar tensor or a float in the " "range (0, 1], got %g" % keep)
|
|
|
|
self.keep = keep
|
|
self.n_units = n_units
|
|
self.W_init = W_init
|
|
self.b_init = b_init
|
|
self.in_channels = in_channels
|
|
|
|
if self.in_channels is not None:
|
|
self.build((None, self.in_channels))
|
|
self._built = True
|
|
|
|
logging.info(
|
|
"DropconnectDense %s: %d %s" %
|
|
(self.name, n_units, self.act.__class__.__name__ if self.act is not None else 'No Activation')
|
|
)
|
|
|
|
def __repr__(self):
|
|
actstr = self.act.__name__ if self.act is not None else 'No Activation'
|
|
s = ('{classname}(n_units={n_units}, ' + actstr)
|
|
s += ', keep={keep}'
|
|
if self.in_channels is not None:
|
|
s += ', in_channels=\'{in_channels}\''
|
|
if self.name is not None:
|
|
s += ', name=\'{name}\''
|
|
s += ')'
|
|
return s.format(classname=self.__class__.__name__, **self.__dict__)
|
|
|
|
def build(self, inputs_shape):
|
|
if len(inputs_shape) != 2:
|
|
raise Exception("The input dimension must be rank 2")
|
|
|
|
if self.in_channels is None:
|
|
self.in_channels = inputs_shape[1]
|
|
|
|
n_in = inputs_shape[-1]
|
|
self.W = self._get_weights("weights", shape=(n_in, self.n_units), init=self.W_init)
|
|
if self.b_init:
|
|
self.b = self._get_weights("biases", shape=(self.n_units), init=self.b_init)
|
|
|
|
self.dropout = tl.ops.Dropout(keep=self.keep)
|
|
self.matmul = tl.ops.MatMul()
|
|
self.bias_add = tl.ops.BiasAdd()
|
|
|
|
def forward(self, inputs):
|
|
if self._forward_state == False:
|
|
if self._built == False:
|
|
self.build(tl.get_tensor_shape(inputs))
|
|
self._built = True
|
|
self._forward_state = True
|
|
|
|
W_dropcon = self.dropout(self.W)
|
|
outputs = self.matmul(inputs, W_dropcon)
|
|
if self.b_init:
|
|
outputs = self.bias_add(outputs, self.b)
|
|
if self.act:
|
|
outputs = self.act(outputs)
|
|
return outputs
|