Blame view
egs/wsj/s5/steps/libs/nnet3/xconfig/convolution.py
59.7 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 |
# Copyright 2018 Johns Hopkins University (Author: Dan Povey)
# 2016 Vijayaditya Peddinti
# Apache 2.0.
""" This module has the implementation of convolutional 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 lines like the following:
#
# conv-batchnorm-layer name=conv2 height-in=40 height-out=40 \
# num-filters-out=64 height-offsets=-1,0,1 time-offsets=-1,0,1 \
# required-time-offsets=0
# or (with NormalizeLayer instead of batch-norm, and with subsampling on the height axis):
# conv-renorm-layer name=conv3 height-in=40 height-out=20 \
# height-subsample-out=2 num-filters-out=128 height-offsets=-1,0,1 \
# time-offsets=-1,0,1 required-time-offsets=0
#
# You don't specify subsampling on the time axis explicitly, it's implicit
# in the 'time-offsets' which are the same as the splicing indexes in a TDNN,
# and which, unlike the height offsets, operate relative to a fixed clock,
# so that after subsampling by a factor of 2, we'd expect all time-offsets
# of subsequent layers to be a factor of 2. You don't specify the input
# num-filters either; it's worked out from the input height and the input dim.
#
# The layer-name encodes the use (or not) of batch normalization, so that if you
# want to skip batch normalization you could just call it 'conv-layer'.
#
# If batch-normalization is used, it's *spatial* batch-normalization, meaning
# that the offset and scale is specific to the output filter, but shared across
# all time and height offsets.
#
# Most of the configuration values mirror same-named values in class
# TimeHeightConvolutionComponent, and for a deeper understanding of what's going
# on you should look at the comment by its declaration, in
# src/nnet3/nnet-convolutional-component.h.
#
# Parameters of the class, and their defaults if they have defaults:
#
# input='[-1]' Descriptor giving the input of the layer.
# height-in The height of the input image, e.g. 40 if the input
# is MFCCs. The num-filters-in is worked out as
# (dimension of input) / height-in. If the preceding
# layer is a convolutional layer, height-in should be
# the same as the height-out of the preceding layer.
# height-subsample-out=1 The height subsampling factor, will be e.g. 2 if you
# want to subsample by a factor of 2 on the height
# axis.
# height-out The height of the output image. This will normally
# be <= (height-in / height-subsample-out).
# Zero-padding on the height axis may be implied by a
# combination of this and height-offsets-in, e.g. if
# height-out==height-in and height-subsample-out=1
# and height-offsets=-2,-1,0,1 then we'd be padding
# by 2 pixels on the bottom and 1 on the top; see
# comments in nnet-convolutional-layers.h for more
# details.
# height-offsets The offsets on the height axis that define what
# inputs require for each output pixel; will
# often be something like -1,0,1 (if zero-padding
# on height axis) or 0,1,2 otherwise. These are
# comparable to TDNN splicing offsets; e.g. if
# height-offsets=-1,0,1 then height 10 at the output
# would take input from heights 9,10,11 at the input.
# num-filters-out The number of output filters. The output dimension
# of this layer is num-filters-out * height-out; the
# filter dim varies the fastest (filter-stride == 1).
# time-offsets The input offsets on the time axis; these are
# interpreted just like the splicing indexes in TDNNs.
# E.g. if time-offsets=-2,0,2 then time 100 at the
# output would require times 98,100,102 at the input.
# required-time-offsets The subset of 'time-offsets' that are required in
# order to produce an output; if the set has fewer
# elements than 'time-offsets' then it implies some
# kind of zero-padding on the time axis is allowed.
# Defaults to the same as 'time-offsets'. For speech
# tasks we recommend not to set this, as the normal
# padding approach is to pad with copies of the
# first/last frame, which is handled automatically in
# the calling code.
# target-rms=1.0 Only applicable if the layer type is
# conv-batchnorm-layer or
# conv-normalize-layer. This will affect the
# scaling of the output features (larger -> larger),
# and sometimes we set target-rms=0.5 for the layer
# prior to the final layer to make the final layer
# train more slowly.
# self-repair-scale=2.0e-05 This affects the ReLu's. It is a scale on the
# 'self-repair' mechanism that nudges the inputs to the
# ReLUs into the appropriate range in cases where
# the unit is active either too little of the time
# (<10%) or too much of the time (>90%).
#
# The following initialization and natural-gradient related options are, if
# provided, passed through to the config file; if not, they are left at the
# defaults in the code. See nnet-convolutional-component.h for more information.
#
# param-stddev, bias-stddev, max-change, learning-rate-factor (float)
# use-natural-gradient (bool)
# rank-in, rank-out (int)
# num-minibatches-history (float)
# alpha-in, alpha-out (float)
# the following is also passed into the convolution components, if specified:
# l2-regularize (float)
class XconfigConvLayer(XconfigLayerBase):
def __init__(self, first_token, key_to_value, prev_names = None):
for operation in first_token.split('-')[:-1]:
assert operation in ['conv', 'renorm', 'batchnorm', 'relu',
'noconv', 'dropout', 'so']
XconfigLayerBase.__init__(self, first_token, key_to_value, prev_names)
def set_default_configs(self):
self.config = {'input':'[-1]',
'height-in':-1,
'height-subsample-out':1,
'height-out':-1,
'height-offsets':'',
'num-filters-out':-1,
'time-offsets':'',
'required-time-offsets':'',
'target-rms':1.0,
'self-repair-scale': 2.0e-05,
'self-repair-lower-threshold': 0.05,
# the following are not really inspected by this level of
# code, just passed through (but not if left at '').
'param-stddev':'', 'bias-stddev':'',
'max-change': 0.75, 'learning-rate-factor':'',
'use-natural-gradient':'',
'rank-in':'', 'rank-out':'', 'num-minibatches-history':'',
'alpha-in':'', 'alpha-out':'', 'l2-regularize':'',
'dropout-proportion': 0.5}
def set_derived_configs(self):
# sets 'num-filters-in'.
input_dim = self.descriptors['input']['dim']
height_in = self.config['height-in']
if height_in <= 0:
raise RuntimeError("height-in must be specified");
if input_dim % height_in != 0:
raise RuntimeError("Input dimension {0} is not a multiple of height-in={1}".format(
input_dim, height_in))
self.config['num-filters-in'] = input_dim // height_in
# Check whether 'str' is a sorted, unique, nonempty list of integers, like -1,0,1.,
# returns true if so.
def check_offsets_var(self, str):
try:
a = [ int(x) for x in str.split(",") ]
if len(a) == 0:
return False
for i in range(len(a) - 1):
if a[i] >= a[i+1]:
return False
return True
except:
return False
def check_configs(self):
# Do some basic checking of the configs. The component-level code does
# some more thorough checking, but if you set the height-out too small it
# prints it as a warning, which the user may not see, so at a minimum we
# want to check for that here.
height_subsample_out = self.config['height-subsample-out']
height_in = self.config['height-in']
height_out = self.config['height-out']
if height_subsample_out <= 0:
raise RuntimeError("height-subsample-out has invalid value {0}.".format(
height_subsample_out))
# we already checked height-in in set_derived_configs.
if height_out <= 0:
raise RuntimeError("height-out has invalid value {0}.".format(
height_out))
if height_out * height_subsample_out > height_in:
raise RuntimeError("The combination height-in={0}, height-out={1} and "
"height-subsample-out={2} does not look right "
"(height-out too large).".format(
height_in, height_out, height_subsample_out))
height_offsets = self.config['height-offsets']
time_offsets = self.config['time-offsets']
required_time_offsets = self.config['required-time-offsets']
if not 'noconv' in self.layer_type.split('-'):
# only check height-offsets, time-offsets and required-time-offsets if there
# is actually a convolution in this layer.
if not self.check_offsets_var(height_offsets):
raise RuntimeError("height-offsets={0} is not valid".format(height_offsets))
if not self.check_offsets_var(time_offsets):
raise RuntimeError("time-offsets={0} is not valid".format(time_offsets))
if required_time_offsets != "" and not self.check_offsets_var(required_time_offsets):
raise RuntimeError("required-time-offsets={0} is not valid".format(
required_time_offsets))
if height_out * height_subsample_out < \
height_in - len(height_offsets.split(',')):
raise RuntimeError("The combination height-in={0}, height-out={1} and "
"height-subsample-out={2} and height-offsets={3} "
"does not look right (height-out too small).")
if self.config['target-rms'] <= 0.0:
raise RuntimeError("Config value target-rms={0} is not valid".format(
self.config['target_rms']))
def auxiliary_outputs(self):
return []
def output_name(self, auxiliary_output = None):
assert auxiliary_output is None
# note: the [:-1] is to remove the '-layer'.
operations = self.layer_type.split('-')[:-1]
if operations[-1] == 'noconv':
operations = operations[:-1]
assert len(operations) >= 1
last_operation = operations[-1]
assert last_operation in ['relu', 'conv', 'renorm', 'batchnorm', 'dropout', 'so']
# we'll return something like 'layer1.batchnorm'.
return '{0}.{1}'.format(self.name, last_operation)
def output_dim(self, auxiliary_output = None):
assert auxiliary_output is None
return self.config['num-filters-out'] * self.config['height-out']
def get_full_config(self):
ans = []
config_lines = self._generate_cnn_config()
for line in config_lines:
for config_name in ['ref', 'final']:
# we do not support user specified matrices in CNN initialization
# so 'ref' and 'final' configs are the same.
ans.append((config_name, line))
return ans
# convenience function to generate the CNN config
def _generate_cnn_config(self):
configs = []
name = self.name
# These 3 variables will be updated as we add components.
cur_num_filters = self.config['num-filters-in']
cur_height = self.config['height-in']
cur_descriptor = self.descriptors['input']['final-string']
# note: the [:-1] is to remove the '-layer'.
operations = self.layer_type.split('-')[:-1]
if operations[-1] == 'noconv':
operations = operations[:-1]
# e.g.:
# operations = [ 'conv', 'relu', 'batchnorm' ]
# or:
# operations = [ 'relu', 'conv', 'renorm' ]
for operation in operations:
if operation == 'conv':
a = []
for opt_name in [
'param-stddev', 'bias-stddev', 'use-natural-gradient',
'max-change', 'rank-in', 'rank-out', 'num-minibatches-history',
'alpha-in', 'alpha-out', 'num-filters-in', 'num-filters-out',
'height-in','height-out', 'height-subsample-out',
'height-offsets', 'time-offsets', 'required-time-offsets',
'learning-rate-factor', 'l2-regularize' ]:
value = self.config[opt_name]
if value != '':
a.append('{0}={1}'.format(opt_name, value))
conv_opts = ' '.join(a)
configs.append('component name={0}.conv type=TimeHeightConvolutionComponent '
'{1}'.format(name, conv_opts))
configs.append('component-node name={0}.conv component={0}.conv '
'input={1}'.format(name, cur_descriptor))
cur_num_filters = self.config['num-filters-out']
cur_height = self.config['height-out']
elif operation == 'batchnorm':
configs.append('component name={0}.batchnorm type=BatchNormComponent dim={1} '
'block-dim={2} target-rms={3}'.format(
name, cur_num_filters * cur_height, cur_num_filters,
self.config['target-rms']))
configs.append('component-node name={0}.batchnorm component={0}.batchnorm '
'input={1}'.format(name, cur_descriptor))
elif operation == 'renorm':
configs.append('component name={0}.renorm type=NormalizeComponent '
'dim={1} target-rms={2}'.format(
name, cur_num_filters * cur_height,
self.config['target-rms']))
configs.append('component-node name={0}.renorm component={0}.renorm '
'input={1}'.format(name, cur_descriptor))
elif operation == 'relu':
configs.append('component name={0}.relu type=RectifiedLinearComponent '
'dim={1} block-dim={2} self-repair-scale={3} '
'self-repair-lower-threshold={4}'.format(
name, cur_num_filters * cur_height, cur_num_filters,
self.config['self-repair-scale'],
self.config['self-repair-lower-threshold']))
configs.append('component-node name={0}.relu component={0}.relu '
'input={1}'.format(name, cur_descriptor))
elif operation == 'dropout':
configs.append('component name={0}.dropout type=DropoutComponent '
'dim={1} dropout-proportion={2}'.format(
name, cur_num_filters * cur_height,
self.config['dropout-proportion']))
configs.append('component-node name={0}.dropout component={0}.dropout '
'input={1}'.format(name, cur_descriptor))
elif operation == 'so':
configs.append('component name={0}.so type=ScaleAndOffsetComponent '
'dim={1} block-dim={2}'.format(
name, cur_num_filters * cur_height, cur_num_filters))
configs.append('component-node name={0}.so component={0}.so '
'input={1}'.format(name, cur_descriptor))
else:
raise RuntimeError("Un-handled operation type: " + operation)
cur_descriptor = '{0}.{1}'.format(name, operation)
return configs
# This class is for lines like the following:
#
# res-block name=res1 num-filters=64 height=32 time-period=1
#
# It implements a residual block as in ResNets, with pre-activation, and with
# some small differences-- basically, instead of adding the input to the output,
# we put a convolutional layer in there but initialize it to the unit matrix and
# if you want you can give it a relatively small (or even zero) learning rate
# and max-change. And there is batch-norm in that path also.
#
# The number of filters is the same on the input and output; it is actually
# redundant to write it in the config file, because given that we know the
# height, we can work it out from the dimension of the input (as dimension =
# height * num-filters). But we allow it to be specified anyway, for clarity.
#
# Note: the res-block does not support subsampling or changing the number of
# filters. If you want to do that, we recommend that you should do it with a
# single relu-batchnorm-conv-layer.
#
# Here are the most important configuration values, with defaults shown if
# defaults exist:
#
# input='[-1]' Descriptor giving the input of the layer.
# height The input and output height of the image, e.g. 40. Note: the width
# is associated with the time dimension and is dealt with
# implicitly, so it's not specified here.
# num-filters The number of filters on the input and output, e.g. 64.
# It does not have to be specified; if it is not specified,
# we work it out from the input dimension.
# num-bottleneck-filters If specified then this will be a 'bottleneck'
# ResBlock, in which there is a 1x1 convolution from
# num-filters->num-bottleneck-filters, a 3x3 convolution
# from num-bottleneck-filters->num-bottleneck-filters, and
# a 1x1 convolution from num-bottleneck-filters->num-filters.
#
# time-period=1 Think of this as the stride in the time dimension. At the
# input of the network will always have time-period=1; then
# after subsampling once in time we'd have time-period=2; then
# after subsampling again we'd have time-period=4. Because of
# the way nnet3 works, subsampling on the time axis is an
# implicit, not explicit, operation.
# height-period=1 This will almost always be left at the default (1). It is
# analogous to time-period, but because the height, unlike the
# time, is explicitly subsampled, in normal topologies this should
# be left at 1.
#
# bypass-source=noop
# The output of this component is Sum(convolution, x), and
# this option controls what 'x' is. There are 3 options
# here: 'noop', 'input', 'relu' or 'batchnorm'. 'noop' is
# equivalent to 'input' in what it computes; it just
# inserts a 'noop' component in order to make the
# computation more efficient. For both 'noop' and
# 'input', x is the input to this component. If
# bypass-source=relu then we use the relu of the
# input; if 'batchnorm', then we use the relu+batchnorm of
# the input.
# allow-zero-padding=true By default this will allow zero-padding in the time
# dimension, meaning that you don't need extra frames at
# the input to compute the output. There may be ASR
# applications where you want to pad in the time dimension
# with repeats of the first or last frame (as we do for
# TDNNs), where it would be appropriate to write
# allow-zero-padding=false. Note: the way we have
# set it up, it does zero-padding on the height axis
# regardless
#
# Less important config variables:
# self-repair-scale=2.0e-05 This affects the ReLu's. It is a scale on the
# 'self-repair' mechanism that nudges the inputs to the
# ReLUs into the appropriate range in cases where
# the unit is active either too little of the time
# (<10%) or too much of the time (>90%).
# max-change=0.75 Max-parameter-change constant (per minibatch)
# used for convolutional components.
#
#
# The following natural-gradient-related configuration variables are passed in
# to the convolution components, if specified:
# use-natural-gradient (bool)
# rank-in, rank-out (int)
# num-minibatches-history (float)
# alpha-in, alpha-out (float)
# the following is also passed into the convolution components, if specified:
# l2-regularize (float)
#
class XconfigResBlock(XconfigLayerBase):
def __init__(self, first_token, key_to_value, prev_names = None):
assert first_token == 'res-block'
XconfigLayerBase.__init__(self, first_token, key_to_value, prev_names)
def set_default_configs(self):
self.config = {'input':'[-1]',
'height':-1,
'num-filters':-1,
'num-bottleneck-filters':-1,
'time-period':1,
'height-period':1,
'self-repair-scale': 2.0e-05,
'self-repair-lower-threshold1': 0.05,
'self-repair-lower-threshold2': 0.05,
'self-repair-lower-threshold3': 0.05,
'max-change': 0.75,
'allow-zero-padding': True,
'bypass-source' : 'noop',
# the following are not really inspected by this level of
# code, just passed through (but not if left at '').
'param-stddev':'', 'bias-stddev':'',
'use-natural-gradient':'',
'rank-in':'', 'rank-out':'',
'num-minibatches-history':'',
'alpha-in':'', 'alpha-out':'', 'l2-regularize':'' }
def set_derived_configs(self):
# set 'num-filters' or check it..
input_dim = self.descriptors['input']['dim']
height = self.config['height']
cur_num_filters = self.config['num-filters']
if cur_num_filters == -1:
if input_dim % height != 0:
raise RuntimeError("Specified image height {0} does not "
"divide the input dim {1}".format(
height, input_dim))
self.config['num-filters'] = input_dim / height
elif input_dim != cur_num_filters * height:
raise RuntimeError("Expected the input-dim to equal "
"height={0} * num-filters={1} = {2}, but "
"it is {3}".format(
height, cur_num_filters,
height * cur_num_filters,
input_dim));
def check_configs(self):
# we checked the dimensions in set_derived_configs.
if not self.config['bypass-source'] in [
'input', 'noop', 'relu', 'batchnorm' ]:
raise RuntimeError("Expected direct-convolution-source to "
"be input, relu or batchnorm, got: {1}".format(
self.config['direct-convolution-source']))
def auxiliary_outputs(self):
return []
def output_name(self, auxiliary_output = None):
bypass_source = self.config['bypass-source']
b = self.config['num-bottleneck-filters']
conv = ('{0}.conv2' if b <= 0 else '{0}.conv3').format(self.name)
if bypass_source == 'input':
residual = self.descriptors['input']['final-string']
elif bypass_source == 'noop':
# we let the noop be the sum of the convolutional part and the
# input, so just return the output of the no-op component.
return '{0}.noop'.format(self.name)
elif bypass_source == 'relu':
residual = '{0}.relu1'.format(self.name)
else:
assert bypass_source == 'batchnorm'
residual = '{0}.batchnorm1'.format(self.name)
return 'Sum({0}, {1})'.format(conv, residual)
def output_dim(self, auxiliary_output = None):
assert auxiliary_output is None
input_dim = self.descriptors['input']['dim']
return input_dim
def get_full_config(self):
ans = []
b = self.config['num-bottleneck-filters']
if b <= 0:
config_lines = self._generate_normal_resblock_config()
else:
config_lines = self._generate_bottleneck_resblock_config()
for line in config_lines:
for config_name in ['ref', 'final']:
# we do not support user specified matrices in CNN initialization
# so 'ref' and 'final' configs are the same.
ans.append((config_name, line))
return ans
# _generate_normal_resblock_config is a convenience function to generate the
# res-block config (the non-bottleneck version).
#
# The main path inside the res-block in the non-bottleneck case is as
# follows:
#
# input -> relu1 -> batchnorm1 -> conv1 -> relu2 -> batchnorm2 -> conv2
#
# We put the relu before the batchnorm because we think it makes more sense;
# because the Torch people seemed to find that this works better
# (https://github.com/gcr/torch-residual-networks/issues/5);
# and because in our batchnorm component we haven't implemented the beta and
# gamma; these would be essential to having it work before relu, but
# when before a convolution or linear component, they add no extra modeling
# power.
#
# The output of the res-block can be the sum of the last convolutional
# component (conv2), with the input. However, the option ('bypass-source')
# controls whether we sum with the raw input, or its relu or relu+batchnorm.
# If the term is going to be the raw input, we give the option ('noop') and
# to cache the output sum via a NoOpComponent)-- because due to how nnet3
# works, if we didn't do this, redundant summing operations would take
# place.
def _generate_normal_resblock_config(self):
configs = []
name = self.name
num_filters = self.config['num-filters']
assert self.config['num-bottleneck-filters'] == -1
height = self.config['height']
input_descriptor = self.descriptors['input']['final-string']
allow_zero_padding = self.config['allow-zero-padding']
height_period = self.config['height-period']
time_period = self.config['time-period']
# input -> relu1 -> batchnorm1 -> conv1 -> relu2 -> batchnorm2 -> conv2
cur_descriptor = input_descriptor
for n in [1, 2]:
# the ReLU
configs.append('component name={0}.relu{1} type=RectifiedLinearComponent '
'dim={2} block-dim={3} self-repair-scale={4} '
'self-repair-lower-threshold={5}'.format(
name, n, num_filters * height, num_filters,
self.config['self-repair-scale'],
self.config['self-repair-lower-threshold{0}'.format(n)]))
configs.append('component-node name={0}.relu{1} component={0}.relu{1} '
'input={2}'.format(name, n, cur_descriptor))
cur_descriptor = '{0}.relu{1}'.format(name, n)
# the batch-norm
configs.append('component name={0}.batchnorm{1} type=BatchNormComponent dim={2} '
'block-dim={3}'.format(
name, n, num_filters * height,
num_filters))
configs.append('component-node name={0}.batchnorm{1} component={0}.batchnorm{1} '
'input={2}'.format(name, n, cur_descriptor))
cur_descriptor = '{0}.batchnorm{1}'.format(name, n)
# the convolution.
a = []
for opt_name in [
'param-stddev', 'bias-stddev', 'use-natural-gradient',
'max-change', 'rank-in', 'rank-out', 'num-minibatches-history',
'alpha-in', 'alpha-out', 'l2-regularize' ]:
value = self.config[opt_name]
if value != '':
a.append('{0}={1}'.format(opt_name, value))
conv_opts = ('height-in={h} height-out={h} height-offsets=-{hp},0,{hp} '
'time-offsets=-{p},0,{p} '
'num-filters-in={f} num-filters-out={f} {r} {o}'.format(
h=height, hp=height_period, p=time_period, f=num_filters,
r=('required-time-offsets=0' if allow_zero_padding else ''),
o=' '.join(a)))
configs.append('component name={0}.conv{1} type=TimeHeightConvolutionComponent '
'{2}'.format(name, n, conv_opts))
configs.append('component-node name={0}.conv{1} component={0}.conv{1} '
'input={2}'.format(name, n, cur_descriptor))
cur_descriptor = '{0}.conv{1}'.format(name, n)
if self.config['bypass-source'] == 'noop':
dim = self.descriptors['input']['dim']
configs.append('component name={0}.noop dim={1} type=NoOpComponent'.format(
name, dim))
configs.append('component-node name={0}.noop component={0}.noop '
'input=Sum({1}, {0}.conv2)'.format(name,
input_descriptor))
# Note: the function 'output_name' is responsible for returning the
# descriptor corresponding to the output of the network.
return configs
# _generate_bottleneck_resblock_config is a convenience function to generate the
# res-block config (this is the bottleneck version, where there is
# a 3x3 kernel with a smaller number of filters than at the input and output,
# sandwiched between two 1x1 kernels.
#
# The main path inside the res-block in the bottleneck case is as follows:
#
# input -> relu1 -> batchnorm1 -> conv1 -> relu2 -> batchnorm2 -> conv2 ->
# relu3 -> batchnorm3 -> conv3
#
# power.
#
# The output of the res-block can be the sum of the last convolutional
# component (conv3), with the input. However we give the option
# ('bypass-source') to sum with the raw input, or its relu or
# relu+batchnorm. If the term is going to be the raw input, we give the
# option ('noop') and to cache the output sum via a NoOpComponent)-- because
# due to how nnet3 works, if we didn't do this, redundant summing operations
# would take place.
def _generate_bottleneck_resblock_config(self):
configs = []
name = self.name
num_filters = self.config['num-filters']
num_bottleneck_filters = self.config['num-bottleneck-filters']
assert num_bottleneck_filters > 0
height = self.config['height']
input_descriptor = self.descriptors['input']['final-string']
allow_zero_padding = self.config['allow-zero-padding']
height_period = self.config['height-period']
time_period = self.config['time-period']
# input -> relu1 -> batchnorm1 -> conv1 -> relu2 -> batchnorm2 -> conv2
cur_descriptor = input_descriptor
cur_num_filters = num_filters
for n in [1, 2, 3]:
# the ReLU
configs.append('component name={0}.relu{1} type=RectifiedLinearComponent '
'dim={2} block-dim={3} self-repair-scale={4} '
'self-repair-lower-threshold={5}'.format(
name, n, cur_num_filters * height, cur_num_filters,
self.config['self-repair-scale'],
self.config['self-repair-lower-threshold{0}'.format(n)]))
configs.append('component-node name={0}.relu{1} component={0}.relu{1} '
'input={2}'.format(name, n, cur_descriptor))
cur_descriptor = '{0}.relu{1}'.format(name, n)
# the batch-norm
configs.append('component name={0}.batchnorm{1} type=BatchNormComponent dim={2} '
'block-dim={3}'.format(
name, n, cur_num_filters * height,
cur_num_filters))
configs.append('component-node name={0}.batchnorm{1} component={0}.batchnorm{1} '
'input={2}'.format(name, n, cur_descriptor))
cur_descriptor = '{0}.batchnorm{1}'.format(name, n)
# the convolution.
a = []
for opt_name in [
'param-stddev', 'bias-stddev', 'use-natural-gradient',
'max-change', 'rank-in', 'rank-out', 'num-minibatches-history',
'alpha-in', 'alpha-out', 'l2-regularize' ]:
value = self.config[opt_name]
if value != '':
a.append('{0}={1}'.format(opt_name, value))
height_offsets = ('-{hp},0,{hp}'.format(hp=height_period) if n == 2 else '0')
time_offsets = ('-{t},0,{t}'.format(t=time_period) if n == 2 else '0')
next_num_filters = (num_filters if n == 3 else num_bottleneck_filters)
conv_opts = ('height-in={h} height-out={h} height-offsets={ho} time-offsets={to} '
'num-filters-in={fi} num-filters-out={fo} {r} {o}'.format(
h=height, ho=height_offsets, to=time_offsets,
fi=cur_num_filters, fo=next_num_filters,
r=('required-time-offsets=0' if allow_zero_padding else ''),
o=' '.join(a)))
configs.append('component name={0}.conv{1} type=TimeHeightConvolutionComponent '
'{2}'.format(name, n, conv_opts))
configs.append('component-node name={0}.conv{1} component={0}.conv{1} '
'input={2}'.format(name, n, cur_descriptor))
cur_descriptor = '{0}.conv{1}'.format(name, n)
cur_num_filters = next_num_filters
if self.config['bypass-source'] == 'noop':
dim = self.descriptors['input']['dim']
configs.append('component name={0}.noop dim={1} type=NoOpComponent'.format(
name, dim))
configs.append('component-node name={0}.noop component={0}.noop '
'input=Sum({1}, {0}.conv3)'.format(name,
input_descriptor))
# Note: the function 'output_name' is responsible for returning the
# descriptor corresponding to the output of the network.
return configs
# This class is for lines like the following:
#
# res2-block name=res1 num-filters=64 height=32 time-period=1
#
# It is a residual block with post-activations, which does not support
# downsampling (strided convolution) or changing the number of filters;
# for that, see res2-downsample-block.
# It's a pretty standard res-block, more standard than "res-block" (XconfigResBlock).
#
# The number of filters is the same on the input and output; it is actually
# redundant to write it in the config file, because given that we know the
# height, we can work it out from the dimension of the input (as dimension =
# height * num-filters). But we allow it to be specified anyway, for clarity.
#
# Here are the most important configuration values, with defaults shown if
# defaults exist:
#
# input='[-1]' Descriptor giving the input of the layer.
# height The input and output height of the image, e.g. 40. Note: the width
# is associated with the time dimension and is dealt with
# implicitly, so it's not specified here.
# num-filters The number of filters on the input and output, e.g. 64.
# It does not have to be specified; if it is not specified,
# we work it out from the input dimension.
# num-bottleneck-filters If specified then this will be a 'bottleneck'
# ResBlock, in which there is a 1x1 convolution from
# num-filters->num-bottleneck-filters, a 3x3 convolution
# from num-bottleneck-filters->num-bottleneck-filters, and
# a 1x1 convolution from num-bottleneck-filters->num-filters.
# time-period=1 Think of this as the stride in the time dimension. At the
# input of the network will always have time-period=1; then
# after subsampling once in time we'd have time-period=2; then
# after subsampling again we'd have time-period=4. Because of
# the way nnet3 works, subsampling on the time axis is an
# implicit, not explicit, operation.
# allow-zero-padding=true By default this will allow zero-padding in the time
# dimension, meaning that you don't need extra frames at
# the input to compute the output. There may be ASR
# applications where you want to pad in the time dimension
# with repeats of the first or last frame (as we do for
# TDNNs), where it would be appropriate to write
# allow-zero-padding=false. Note: the way we have
# set it up, it does zero-padding on the height axis
# regardless
#
# Less important config variables:
# self-repair-scale=2.0e-05 This affects the ReLu's. It is a scale on the
# 'self-repair' mechanism that nudges the inputs to the
# ReLUs into the appropriate range in cases where
# the unit is active either too little of the time
# (<10%) or too much of the time (>90%).
# max-change=0.75 Max-parameter-change constant (per minibatch)
# used for convolutional components.
#
#
# The following natural-gradient-related configuration variables are passed in
# to the convolution components, if specified:
# use-natural-gradient (bool)
# rank-in, rank-out (int)
# num-minibatches-history (float)
# alpha-in, alpha-out (float)
# the following is also passed into the convolution components, if specified:
# l2-regularize (float)
class XconfigRes2Block(XconfigLayerBase):
def __init__(self, first_token, key_to_value, prev_names = None):
assert first_token == 'res2-block'
XconfigLayerBase.__init__(self, first_token, key_to_value, prev_names)
def set_default_configs(self):
self.config = {'input':'[-1]',
'height':-1, # sets height-in and height-out
'height-in':-1,
'height-out':-1,
'num-filters':-1, # interpreted as num-filters-out.
'num-bottleneck-filters':-1,
'time-period':1,
'self-repair-scale': 2.0e-05,
'self-repair-lower-threshold1': 0.05,
'self-repair-lower-threshold2': 0.05,
'self-repair-lower-threshold3': 0.05,
'max-change': 0.75,
'allow-zero-padding': True,
# the following are not really inspected by this level of
# code, just passed through (but not if left at '').
'param-stddev':'', 'bias-stddev':'',
'use-natural-gradient':'',
'rank-in':'', 'rank-out':'',
'num-minibatches-history':'',
'alpha-in':'', 'alpha-out':'',
'l2-regularize':'' }
def set_derived_configs(self):
input_dim = self.descriptors['input']['dim']
if not ((self.config['height'] > 0 and self.config['height-in'] == -1 and
self.config['height-out'] == -1) or
(self.config['height-out'] > 0 and self.config['height-in'] > 0)):
raise RuntimeError("You must specify height, or height-in and height-out, for res2-block.")
if not (self.config['height-in'] > 0 and self.config['height-out'] > 0):
height = self.config['height']
if not height > 0:
raise RuntimeError("You must specify either height, or height-in and height-out, for "
"res2-block.")
self.config['height-in'] = height
self.config['height-out'] = height
height_in = self.config['height-in']
if input_dim % height_in != 0:
raise RuntimeError("Specified input image height {0} does not "
"divide the input dim {1}".format(
height_in, input_dim))
self.config['num-filters'] = input_dim / height
def check_configs(self):
if self.config['num-filters'] == -1:
raise RuntimeError("You must specify num-filters for res2-block.")
def auxiliary_outputs(self):
return []
def output_name(self, auxiliary_output = None):
b = self.config['num-bottleneck-filters']
return ('{0}.relu2' if b <= 0 else '{0}.relu3').format(self.name)
def output_dim(self, auxiliary_output = None):
assert auxiliary_output is None
return self.config['height-out'] * self.config['num-filters']
def get_full_config(self):
ans = []
b = self.config['num-bottleneck-filters']
if b <= 0:
config_lines = self._generate_normal_resblock_config()
else:
config_lines = self._generate_bottleneck_resblock_config()
for line in config_lines:
for config_name in ['ref', 'final']:
# we do not support user specified matrices in CNN initialization
# so 'ref' and 'final' configs are the same.
ans.append((config_name, line))
return ans
# _generate_normal_resblock_config is a convenience function to generate the
# res-block config (the non-bottleneck version).
#
# The main path inside the res-block in the non-bottleneck case is as
# follows:
#
# input -> conv1 -> batchnorm1 -> scaleoffset1 -> relu1 -> conv2 -> batchnorm2 -> scaleoffset2 -> relu2
#
# where the 'scaleoffsetN' are ScaleAndOffsetComponent, which conventionally would be
# considered part of the BatchNorm.
#
# The relu2 actually sees the sum of the input and 'scaleoffset2'-- which gives us the bypass
# connection.
def _generate_normal_resblock_config(self):
configs = []
name = self.name
assert self.config['num-bottleneck-filters'] == -1
input_dim = self.descriptors['input']['dim']
height_in = self.config['height-in']
height_out = self.config['height-out']
time_period_out = self.config['time-period']
if not input_dim % height_in == 0:
raise RuntimeError("input-dim {0} does not divide height-in {1}".format(
input_dim, height_in))
num_filters_in = input_dim / height_in
num_filters_out = self.config['num-filters']
if height_out != height_in:
if height_out < height_in / 2 - 1 or height_out > height_in / 2 + 1:
raise RuntimeError("Expected height-out to be about half height-in, or the same: "
"height-in={0} height-out={1}".format(height_in, height_out))
if not time_period_out % 2 == 0:
raise RuntimeError("Expected time-period to be a multiple of 2 if you are subsampling "
"on height.")
time_period_in = time_period_out / 2
height_subsample = 2
else:
time_period_in = time_period_out
height_subsample = 1
cur_time_period = time_period_in
cur_num_filters = num_filters_in
cur_height = height_in
input_descriptor = self.descriptors['input']['final-string']
allow_zero_padding = self.config['allow-zero-padding']
if height_subsample == 1 and num_filters_in == num_filters_out:
bypass_descriptor = input_descriptor
else:
bypass_descriptor = '{0}.conv_bypass'.format(name)
cur_descriptor = input_descriptor
# get miscellaneous convolution options passed in from the xconfig line
a = []
for opt_name in [
'param-stddev', 'bias-stddev', 'use-natural-gradient',
'max-change', 'rank-in', 'rank-out', 'num-minibatches-history',
'alpha-in', 'alpha-out', 'l2-regularize' ]:
value = self.config[opt_name]
if value != '':
a.append('{0}={1}'.format(opt_name, value))
misc_conv_opts = ' '.join(a)
for n in [1, 2]:
# the convolution.
conv_opts = ('height-in={hi} height-out={ho} height-offsets=-1,0,1 '
'height-subsample-out={hs} '
'time-offsets=-{p},0,{p} '
'num-filters-in={fi} num-filters-out={fo} {r} {o}'.format(
hi=cur_height, ho=height_out,
p=cur_time_period,
hs=(height_subsample if n == 1 else 1),
fi=cur_num_filters,
fo=num_filters_out,
r=('required-time-offsets=0' if allow_zero_padding else ''),
o=misc_conv_opts))
configs.append('component name={0}.conv{1} type=TimeHeightConvolutionComponent '
'{2}'.format(name, n, conv_opts))
configs.append('component-node name={0}.conv{1} component={0}.conv{1} '
'input={2}'.format(name, n, cur_descriptor))
cur_descriptor = '{0}.conv{1}'.format(name, n)
cur_num_filters = num_filters_out
cur_height = height_out
cur_time_period = time_period_out
# the batch-norm
configs.append('component name={0}.batchnorm{1} type=BatchNormComponent dim={2} '
'block-dim={3}'.format(
name, n, cur_num_filters * cur_height,
cur_num_filters))
configs.append('component-node name={0}.batchnorm{1} component={0}.batchnorm{1} '
'input={2}'.format(name, n, cur_descriptor))
cur_descriptor = '{0}.batchnorm{1}'.format(name, n)
# the scale-and-offset
configs.append('component name={0}.scaleoffset{1} type=ScaleAndOffsetComponent dim={2} '
'block-dim={3}'.format(
name, n, cur_num_filters * cur_height,
cur_num_filters))
configs.append('component-node name={0}.scaleoffset{1} component={0}.scaleoffset{1} '
'input={2}'.format(name, n, cur_descriptor))
cur_descriptor = '{0}.scaleoffset{1}'.format(name, n)
if n == 2:
# the bypass connection
cur_descriptor = 'Sum({0}, {1})'.format(cur_descriptor, bypass_descriptor)
# the ReLU
configs.append('component name={0}.relu{1} type=RectifiedLinearComponent '
'dim={2} block-dim={3} self-repair-scale={4} '
'self-repair-lower-threshold={5}'.format(
name, n, cur_num_filters * cur_height, cur_num_filters,
self.config['self-repair-scale'],
self.config['self-repair-lower-threshold{0}'.format(n)]))
configs.append('component-node name={0}.relu{1} component={0}.relu{1} '
'input={2}'.format(name, n, cur_descriptor))
cur_descriptor = '{0}.relu{1}'.format(name, n)
if bypass_descriptor != input_descriptor:
# We need to add the 1x1 bypass convolution because we're either doing height
# subsampling or changing the number of filters.
conv_opts = ('height-in={hi} height-out={ho} height-offsets=0 '
'time-offsets=0 height-subsample-out={hs} '
'num-filters-in={fi} num-filters-out={fo} {o}'.format(
hi=height_in, ho=height_out, hs=height_subsample,
fi=num_filters_in, fo=num_filters_out, o=misc_conv_opts))
configs.append('component name={0}.conv_bypass type=TimeHeightConvolutionComponent '
'{1}'.format(name, conv_opts))
configs.append('component-node name={0}.conv_bypass component={0}.conv_bypass '
'input={1}'.format(name, input_descriptor))
# Note: the function 'output_name' is responsible for returning the
# descriptor corresponding to the output of the network, which in
# this case would be '{0}.relu2'.format(name).
return configs
# _generate_bottleneck_resblock_config is a convenience function to generate the
# res-block config (this is the bottleneck version, where there is
# a 3x3 kernel with a smaller number of filters than at the input and output,
# sandwiched between two 1x1 kernels.
#
# The main path inside the res-block in the bottleneck case is as follows:
#
# input -> conv1 -> batchnorm1 -> scaleoffset1 -> relu1 ->
# conv2 -> batchnorm2 -> scaleoffset2 -> relu2 ->
# conv3 -> batchnorm3 -> scaleoffset3 -> relu3
#
# but the relu3 takes as its input the sum of 'input' and 'scaleoffset3'.
#
def _generate_bottleneck_resblock_config(self):
configs = []
name = self.name
num_bottleneck_filters = self.config['num-bottleneck-filters']
assert num_bottleneck_filters > 0
input_dim = self.descriptors['input']['dim']
height_in = self.config['height-in']
height_out = self.config['height-out']
input_descriptor = self.descriptors['input']['final-string']
allow_zero_padding = self.config['allow-zero-padding']
time_period_out = self.config['time-period']
if not input_dim % height_in == 0:
raise RuntimeError("input-dim={0} does not divide height-in={1}".format(
input_dim, height_in))
num_filters_in = input_dim / height_in
num_filters_out = self.config['num-filters']
if height_out != height_in:
if height_out < height_in / 2 - 1 or height_out > height_in / 2 + 1:
raise RuntimeError("Expected height-out to be about half height-in, or the same: "
"height-in={0} height-out={1}".format(height_in, height_out))
height_subsample = 2
else:
height_subsample = 1
cur_descriptor = input_descriptor
cur_num_filters = num_filters_in
cur_height = height_in
if height_subsample == 1 and num_filters_in == num_filters_out:
bypass_descriptor = input_descriptor
else:
bypass_descriptor = '{0}.conv_bypass'.format(name)
# get miscellaneous convolution options passed in from the xconfig line
a = []
for opt_name in [
'param-stddev', 'bias-stddev', 'use-natural-gradient',
'max-change', 'rank-in', 'rank-out', 'num-minibatches-history',
'alpha-in', 'alpha-out', 'l2-regularize' ]:
value = self.config[opt_name]
if value != '':
a.append('{0}={1}'.format(opt_name, value))
misc_conv_opts = ' '.join(a)
for n in [1, 2, 3]:
# the convolution.
height_offsets = ('-1,0,1' if n == 2 else '0')
this_height_subsample = height_subsample if n == 1 else 1
time_offsets = ('-{t},0,{t}'.format(t=time_period_out) if n == 2 else '0')
next_num_filters = (num_filters_out if n == 3 else num_bottleneck_filters)
conv_opts = ('height-in={h_in} height-out={h_out} height-offsets={ho} time-offsets={to} '
'num-filters-in={fi} num-filters-out={fo} height-subsample-out={hs} '
'{r} {o}'.format(
h_in=cur_height, h_out=height_out,
to=time_offsets, ho=height_offsets,
hs=this_height_subsample,
fi=cur_num_filters, fo=next_num_filters,
r=('required-time-offsets=0' if allow_zero_padding else ''),
o=misc_conv_opts))
configs.append('component name={0}.conv{1} type=TimeHeightConvolutionComponent '
'{2}'.format(name, n, conv_opts))
configs.append('component-node name={0}.conv{1} component={0}.conv{1} '
'input={2}'.format(name, n, cur_descriptor))
cur_num_filters = next_num_filters
cur_height = height_out
cur_descriptor = '{0}.conv{1}'.format(name, n)
# the batch-norm
configs.append('component name={0}.batchnorm{1} type=BatchNormComponent dim={2} '
'block-dim={3}'.format(
name, n, cur_num_filters * cur_height,
cur_num_filters))
configs.append('component-node name={0}.batchnorm{1} component={0}.batchnorm{1} '
'input={2}'.format(name, n, cur_descriptor))
cur_descriptor = '{0}.batchnorm{1}'.format(name, n)
# the scale and offset
configs.append('component name={0}.scaleoffset{1} type=ScaleAndOffsetComponent dim={2} '
'block-dim={3}'.format(
name, n, cur_num_filters * cur_height,
cur_num_filters))
configs.append('component-node name={0}.scaleoffset{1} component={0}.scaleoffset{1} '
'input={2}'.format(name, n, cur_descriptor))
cur_descriptor = '{0}.scaleoffset{1}'.format(name, n)
if n == 3:
# the bypass connection
cur_descriptor = 'Sum({0}, {1})'.format(cur_descriptor, bypass_descriptor)
# the ReLU
configs.append('component name={0}.relu{1} type=RectifiedLinearComponent '
'dim={2} block-dim={3} self-repair-scale={4} '
'self-repair-lower-threshold={5}'.format(
name, n, cur_num_filters * cur_height, cur_num_filters,
self.config['self-repair-scale'],
self.config['self-repair-lower-threshold{0}'.format(n)]))
configs.append('component-node name={0}.relu{1} component={0}.relu{1} '
'input={2}'.format(name, n, cur_descriptor))
cur_descriptor = '{0}.relu{1}'.format(name, n)
if bypass_descriptor != input_descriptor:
# We need to add the 1x1 bypass convolution because we're either doing height
# subsampling or changing the number of filters.
conv_opts = ('height-in={hi} height-out={ho} height-offsets=0 '
'time-offsets=0 height-subsample-out={hs} '
'num-filters-in={fi} num-filters-out={fo} {o}'.format(
hi=height_in, ho=height_out, hs=height_subsample,
fi=num_filters_in, fo=num_filters_out, o=misc_conv_opts))
configs.append('component name={0}.conv_bypass type=TimeHeightConvolutionComponent '
'{1}'.format(name, conv_opts))
configs.append('component-node name={0}.conv_bypass component={0}.conv_bypass '
'input={1}'.format(name, input_descriptor))
# Note: the function 'output_name' is responsible for returning the
# descriptor corresponding to the output of the network, which
# in this case will be '{0}.relu3'.format(name).
return configs
# This layer just maps to a single component, a SumBlockComponent. It's for
# doing channel averaging at the end of neural networks. See scripts for
# examples of how to use it.
# An example line using this layer is:
# channel-average-layer name=channel-average input=Append(2, 4, 6, 8) dim=64
# the configuration value 'dim' is the output dimension of this layer.
# The input dimension is expected to be a multiple of 'dim'. The output
# will be the average of 'dim'-sized blocks of the input.
class ChannelAverageLayer(XconfigLayerBase):
def __init__(self, first_token, key_to_value, prev_names = None):
assert first_token == "channel-average-layer"
XconfigLayerBase.__init__(self, first_token, key_to_value, prev_names)
def set_default_configs(self):
self.config = {'input':'[-1]',
'dim': -1 }
def set_derived_configs(self):
pass
def check_configs(self):
input_dim = self.descriptors['input']['dim']
dim = self.config['dim']
if dim <= 0:
raise RuntimeError("dim must be specified and > 0.")
if input_dim % dim != 0:
raise RuntimeError("input-dim={0} is not a multiple of dim={1}".format(
input_dim, dim))
def auxiliary_outputs(self):
return []
def output_name(self, auxiliary_output = None):
assert auxiliary_output is None
return self.name
def output_dim(self, auxiliary_output = None):
assert auxiliary_output is None
return self.config['dim']
def get_full_config(self):
ans = []
config_lines = self._generate_channel_average_config()
for line in config_lines:
for config_name in ['ref', 'final']:
ans.append((config_name, line))
return ans
def _generate_channel_average_config(self):
configs = []
name = self.name
input_dim = self.descriptors['input']['dim']
input_descriptor = self.descriptors['input']['final-string']
dim = self.config['dim']
# choose the scale that makes it an average rather than a sum.
scale = dim * 1.0 / input_dim
configs.append('component name={0} type=SumBlockComponent input-dim={1} '
'output-dim={2} scale={3}'.format(name, input_dim,
dim, scale))
configs.append('component-node name={0} component={0} input={1}'.format(
name, input_descriptor))
return configs
|