tensorlayer3/tensorlayer/layers/dense/quan_dense_bn.py

187 lines
6.6 KiB
Python

#! /usr/bin/python
# -*- coding: utf-8 -*-
import tensorlayer as tl
from tensorlayer import logging
from tensorlayer.layers.core import Module
from tensorflow.python.training import moving_averages
from tensorlayer.layers.utils import (
quantize_active_overflow, quantize_weight_overflow, mean_var_with_update, w_fold, bias_fold
)
__all__ = [
'QuanDenseWithBN',
]
class QuanDenseWithBN(Module):
"""The :class:`QuanDenseWithBN` class is a quantized fully connected layer with BN, which weights are 'bitW' bits and the output of the previous layer
are 'bitA' bits while inferencing.
# TODO The QuanDenseWithBN only supports TensorFlow backend.
Parameters
----------
n_units : int
The number of units of this layer.
act : activation function
The activation function of this layer.
decay : float
A decay factor for `ExponentialMovingAverage`.
Suggest to use a large value for large dataset.
epsilon : float
Eplison.
is_train : boolean
Is being used for training or inference.
beta_init : initializer or None
The initializer for initializing beta, if None, skip beta.
Usually you should not skip beta unless you know what happened.
gamma_init : initializer or None
The initializer for initializing gamma, if None, skip gamma.
bitW : int
The bits of this layer's parameter
bitA : int
The bits of the output of previous layer
use_gemm : boolean
If True, use gemm instead of ``tf.matmul`` for inferencing. (TODO).
W_init : initializer
The initializer for the the weight matrix.
W_init_args : dictionary
The arguments for the weight matrix initializer.
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 : a str
A unique layer name.
Examples
---------
>>> import tensorlayer as tl
>>> net = tl.layers.Input([50, 256])
>>> layer = tl.layers.QuanDenseWithBN(128, act='relu', name='qdbn1')(net)
>>> net = tl.layers.QuanDenseWithBN(256, act='relu', name='qdbn2')(net)
"""
def __init__(
self,
n_units=100,
act=None,
decay=0.9,
epsilon=1e-5,
is_train=False,
bitW=8,
bitA=8,
gamma_init=tl.initializers.truncated_normal(stddev=0.05),
beta_init=tl.initializers.truncated_normal(stddev=0.05),
use_gemm=False,
W_init=tl.initializers.truncated_normal(stddev=0.05),
W_init_args=None,
in_channels=None,
name=None, # 'quan_dense_with_bn',
):
super(QuanDenseWithBN, self).__init__(act=act, W_init_args=W_init_args, name=name)
self.n_units = n_units
self.decay = decay
self.epsilon = epsilon
self.is_train = is_train
self.bitW = bitW
self.bitA = bitA
self.gamma_init = gamma_init
self.beta_init = beta_init
self.use_gemm = use_gemm
self.W_init = W_init
self.in_channels = in_channels
if self.in_channels is not None:
self.build((None, self.in_channels))
self._built = True
logging.info(
"QuanDenseLayerWithBN %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.__class__.__name__ if self.act is not None else 'No Activation'
s = ('{classname}(n_units={n_units}, ' + actstr)
s += ', bitW={bitW}, bitA={bitA}'
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 self.in_channels is None and len(inputs_shape) != 2:
raise Exception("The input dimension must be rank 2, please reshape or flatten it")
if self.in_channels is None:
self.in_channels = inputs_shape[1]
if self.use_gemm:
raise Exception("TODO. The current version use tf.matmul for inferencing.")
n_in = inputs_shape[-1]
self.W = self._get_weights("weights", shape=(n_in, self.n_units), init=self.W_init)
para_bn_shape = (self.n_units, )
if self.gamma_init:
self.scale_para = self._get_weights("gamm_weights", shape=para_bn_shape, init=self.gamma_init)
else:
self.scale_para = None
if self.beta_init:
self.offset_para = self._get_weights("beta_weights", shape=para_bn_shape, init=self.beta_init)
else:
self.offset_para = None
self.moving_mean = self._get_weights(
"moving_mean", shape=para_bn_shape, init=tl.initializers.constant(1.0), trainable=False
)
self.moving_variance = self._get_weights(
"moving_variacne", shape=para_bn_shape, init=tl.initializers.constant(1.0), trainable=False
)
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
x = inputs
inputs = quantize_active_overflow(inputs, self.bitA)
mid_out = tl.ops.matmul(x, self.W)
mean, variance = tl.ops.moments(x=mid_out, axes=list(range(len(mid_out.get_shape()) - 1)))
update_moving_mean = moving_averages.assign_moving_average(
self.moving_mean, mean, self.decay, zero_debias=False
) # if zero_debias=True, has bias
update_moving_variance = moving_averages.assign_moving_average(
self.moving_variance, variance, self.decay, zero_debias=False
) # if zero_debias=True, has bias
if self.is_train:
mean, var = mean_var_with_update(update_moving_mean, update_moving_variance, mean, variance)
else:
mean, var = self.moving_mean, self.moving_variance
_w_fold = w_fold(self.W, self.scale_para, var, self.epsilon)
W = quantize_weight_overflow(_w_fold, self.bitW)
outputs = tl.ops.matmul(inputs, W)
if self.beta_init:
_bias_fold = bias_fold(self.offset_para, self.scale_para, mean, var, self.epsilon)
outputs = tl.ops.bias_add(outputs, _bias_fold)
else:
outputs = outputs
if self.act:
outputs = self.act(outputs)
else:
outputs = outputs
return outputs