Blame view
egs/wsj/s5/steps/libs/nnet3/xconfig/attention.py
11 KB
8dcb6dfcb
first commit
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 |
# Copyright 2017 Johns Hopkins University (Dan Povey)
# 2017 Hossein Hadian
# Apache 2.0.
""" This module has the implementation of attention layers.
"""
from __future__ import print_function
from __future__ import division
import math
import re
import sys
from libs.nnet3.xconfig.basic_layers import XconfigLayerBase
# This class is for parsing lines like
# 'attention-renorm-layer num-heads=10 value-dim=50 key-dim=50 time-stride=3 num-left-inputs=5 num-right-inputs=2.'
#
# Parameters of the class, and their defaults:
# input='[-1]' [Descriptor giving the input of the layer.]
# self-repair-scale=1.0e-05 [Affects relu, sigmoid and tanh layers.]
# learning-rate-factor=1.0 [This can be used to make the affine component
# train faster or slower].
# Documentation for the rest of the parameters (related to the
# attention component) can be found in nnet-attention-component.h
class XconfigAttentionLayer(XconfigLayerBase):
def __init__(self, first_token, key_to_value, prev_names = None):
# Here we just list some likely combinations.. you can just add any
# combinations you want to use, to this list.
assert first_token in ['attention-renorm-layer',
'attention-relu-renorm-layer',
'attention-relu-batchnorm-layer',
'relu-renorm-attention-layer']
XconfigLayerBase.__init__(self, first_token, key_to_value, prev_names)
def set_default_configs(self):
# note: self.config['input'] is a descriptor, '[-1]' means output
# the most recent layer.
self.config = { 'input':'[-1]',
'dim': -1,
'max-change' : 0.75,
'self-repair-scale' : 1.0e-05,
'target-rms' : 1.0,
'learning-rate-factor' : 1.0,
'ng-affine-options' : '',
'l2-regularize': 0.0,
'num-left-inputs-required': -1,
'num-right-inputs-required': -1,
'output-context': True,
'time-stride': 1,
'num-heads': 1,
'key-dim': -1,
'key-scale': 0.0,
'value-dim': -1,
'num-left-inputs': -1,
'num-right-inputs': -1,
'dropout-proportion': 0.5} # dropout-proportion only
# affects layers with
# 'dropout' in the name.
def check_configs(self):
if self.config['self-repair-scale'] < 0.0 or self.config['self-repair-scale'] > 1.0:
raise RuntimeError("self-repair-scale has invalid value {0}"
.format(self.config['self-repair-scale']))
if self.config['target-rms'] < 0.0:
raise RuntimeError("target-rms has invalid value {0}"
.format(self.config['target-rms']))
if self.config['learning-rate-factor'] <= 0.0:
raise RuntimeError("learning-rate-factor has invalid value {0}"
.format(self.config['learning-rate-factor']))
for conf in ['value-dim', 'key-dim',
'num-left-inputs', 'num-right-inputs']:
if self.config[conf] < 0:
raise RuntimeError("{0} has invalid value {1}"
.format(conf, self.config[conf]))
if self.config['key-scale'] == 0.0:
self.config['key-scale'] = 1.0 / math.sqrt(self.config['key-dim'])
def output_name(self, auxiliary_output=None):
# at a later stage we might want to expose even the pre-nonlinearity
# vectors
assert auxiliary_output == None
split_layer_name = self.layer_type.split('-')
assert split_layer_name[-1] == 'layer'
last_nonlinearity = split_layer_name[-2]
# return something like: layer3.renorm
return '{0}.{1}'.format(self.name, last_nonlinearity)
def attention_input_dim(self):
context_dim = (self.config['num-left-inputs'] +
self.config['num-right-inputs'] + 1)
num_heads = self.config['num-heads']
key_dim = self.config['key-dim']
value_dim = self.config['value-dim']
query_dim = key_dim + context_dim;
return num_heads * (key_dim + value_dim + query_dim)
def attention_output_dim(self):
context_dim = (self.config['num-left-inputs'] +
self.config['num-right-inputs'] + 1)
num_heads = self.config['num-heads']
value_dim = self.config['value-dim']
return (num_heads *
(value_dim +
(context_dim if self.config['output-context'] else 0)))
def output_dim(self, auxiliary_output = None):
return self.attention_output_dim()
def get_full_config(self):
ans = []
config_lines = self._generate_config()
for line in config_lines:
for config_name in ['ref', 'final']:
# we do not support user specified matrices in this layer
# so 'ref' and 'final' configs are the same.
ans.append((config_name, line))
return ans
def _generate_config(self):
split_layer_name = self.layer_type.split('-')
assert split_layer_name[-1] == 'layer'
nonlinearities = split_layer_name[:-1]
# by 'descriptor_final_string' we mean a string that can appear in
# config-files, i.e. it contains the 'final' names of nodes.
input_desc = self.descriptors['input']['final-string']
input_dim = self.descriptors['input']['dim']
# the child classes e.g. tdnn might want to process the input
# before adding the other components
return self._add_components(input_desc, input_dim, nonlinearities)
def _add_components(self, input_desc, input_dim, nonlinearities):
dim = self.attention_input_dim()
self_repair_scale = self.config['self-repair-scale']
target_rms = self.config['target-rms']
max_change = self.config['max-change']
ng_affine_options = self.config['ng-affine-options']
l2_regularize = self.config['l2-regularize']
learning_rate_factor=self.config['learning-rate-factor']
learning_rate_option=('learning-rate-factor={0}'.format(learning_rate_factor)
if learning_rate_factor != 1.0 else '')
l2_regularize_option = ('l2-regularize={0} '.format(l2_regularize)
if l2_regularize != 0.0 else '')
configs = []
# First the affine node.
line = ('component name={0}.affine'
' type=NaturalGradientAffineComponent'
' input-dim={1}'
' output-dim={2}'
' max-change={3}'
' {4} {5} {6}'
''.format(self.name, input_dim, dim,
max_change, ng_affine_options,
learning_rate_option, l2_regularize_option))
configs.append(line)
line = ('component-node name={0}.affine'
' component={0}.affine input={1}'
''.format(self.name, input_desc))
configs.append(line)
cur_node = '{0}.affine'.format(self.name)
for nonlinearity in nonlinearities:
if nonlinearity == 'relu':
line = ('component name={0}.{1}'
' type=RectifiedLinearComponent dim={2}'
' self-repair-scale={3}'
''.format(self.name, nonlinearity, dim,
self_repair_scale))
elif nonlinearity == 'attention':
line = ('component name={0}.{1}'
' type=RestrictedAttentionComponent'
' value-dim={2}'
' key-dim={3}'
' num-left-inputs={4}'
' num-right-inputs={5}'
' num-left-inputs-required={6}'
' num-right-inputs-required={7}'
' output-context={8}'
' time-stride={9}'
' num-heads={10}'
' key-scale={11}'
''.format(self.name, nonlinearity,
self.config['value-dim'],
self.config['key-dim'],
self.config['num-left-inputs'],
self.config['num-right-inputs'],
self.config['num-left-inputs-required'],
self.config['num-right-inputs-required'],
self.config['output-context'],
self.config['time-stride'],
self.config['num-heads'],
self.config['key-scale']))
dim = self.attention_output_dim()
elif nonlinearity == 'sigmoid':
line = ('component name={0}.{1}'
' type=SigmoidComponent dim={2}'
' self-repair-scale={3}'
''.format(self.name, nonlinearity, dim,
self_repair_scale))
elif nonlinearity == 'tanh':
line = ('component name={0}.{1}'
' type=TanhComponent dim={2}'
' self-repair-scale={3}'
''.format(self.name, nonlinearity, dim,
self_repair_scale))
elif nonlinearity == 'renorm':
line = ('component name={0}.{1}'
' type=NormalizeComponent dim={2}'
' target-rms={3}'
''.format(self.name, nonlinearity, dim,
target_rms))
elif nonlinearity == 'batchnorm':
line = ('component name={0}.{1}'
' type=BatchNormComponent dim={2}'
' target-rms={3}'
''.format(self.name, nonlinearity, dim,
target_rms))
elif nonlinearity == 'dropout':
line = ('component name={0}.{1} type=DropoutComponent '
'dim={2} dropout-proportion={3}'.format(
self.name, nonlinearity, dim,
self.config['dropout-proportion']))
else:
raise RuntimeError("Unknown nonlinearity type: {0}"
.format(nonlinearity))
configs.append(line)
line = ('component-node name={0}.{1}'
' component={0}.{1} input={2}'
''.format(self.name, nonlinearity, cur_node))
configs.append(line)
cur_node = '{0}.{1}'.format(self.name, nonlinearity)
return configs
|