Blame view
egs/wsj/s5/steps/libs/nnet3/xconfig/basic_layers.py
62.2 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 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 |
# Copyright 2016 Johns Hopkins University (Dan Povey) # 2016 Vijayaditya Peddinti # 2017 Google Inc. (vpeddinti@google.com) # 2017 Vimal Manohar # Apache 2.0. """ This module contains the parent class from which all layers are inherited and some basic layer definitions. """ from __future__ import print_function from __future__ import division import math import re import sys import libs.nnet3.xconfig.utils as xutils import libs.common as common_lib class XconfigLayerBase(object): """ A base-class for classes representing layers of xconfig files. """ def __init__(self, first_token, key_to_value, all_layers): """ first_token: first token on the xconfig line, e.g. 'affine-layer'.f key_to_value: dictionary with parameter values { 'name':'affine1', 'input':'Append(0, 1, 2, ReplaceIndex(ivector, t, 0))', 'dim=1024' }. The only required and 'special' values that are dealt with directly at this level, are 'name' and 'input'. The rest are put in self.config and are dealt with by the child classes' init functions. all_layers: An array of objects inheriting XconfigLayerBase for all previously parsed layers. """ self.layer_type = first_token if 'name' not in key_to_value: raise RuntimeError("Expected 'name' to be specified.") self.name = key_to_value['name'] if not xutils.is_valid_line_name(self.name): raise RuntimeError("Invalid value: name={0}".format( key_to_value['name'])) # It is possible to have two layers with a same name in 'all_layer', if # the layer type for one of them is 'existing'. # Layers of type 'existing' are corresponding to the component-node names # in the existing model, which we are adding layers to them. # 'existing' layers are not presented in any config file, and new layer # with the same name can exist in 'all_layers'. # e.g. It is possible to have 'output-node' with name 'output' in the # existing model, which is added to all_layers using layer type 'existing', # and 'output-node' of type 'output-layer' with the same name 'output' in # 'all_layers'. for prev_layer in all_layers: if (self.name == prev_layer.name and prev_layer.layer_type is not 'existing'): raise RuntimeError("Name '{0}' is used for more than one " "layer.".format(self.name)) self.config = {} # the following, which should be overridden in the child class, sets # default config parameters in self.config. self.set_default_configs() # The following is not to be reimplemented in child classes; # it sets the config values to those specified by the user, and # parses any Descriptors. self.set_configs(key_to_value, all_layers) # This method, sets the derived default config values # i.e., config values when not specified can be derived from # other values. It can be overridden in the child class. self.set_derived_configs() # the following, which should be overridden in the child class, checks # that the config parameters that have been set are reasonable. self.check_configs() def set_configs(self, key_to_value, all_layers): """ Sets the config variables. We broke this code out of __init__ for clarity. the child-class constructor will deal with the configuration values in a more specific way. """ # First check that there are no keys that don't correspond to any config # parameter of this layer, and if so, raise an exception with an # informative message saying what configs are allowed. for key, value in key_to_value.items(): if key != 'name': if key not in self.config: configs = ' '.join([('{0}->"{1}"'.format(x, y) if isinstance(y, str) else '{0}->{1}'.format(x, y)) for x, y in self.config.items()]) raise RuntimeError("Configuration value {0}={1} was not " "expected in layer of type {2}; allowed " "configs with their defaults: {3}" "" .format(key, value, self.layer_type, configs)) for key, value in key_to_value.items(): if key != 'name': assert key in self.config # we checked above. self.config[key] = xutils.convert_value_to_type(key, type(self.config[key]), value) self.descriptors = dict() self.descriptor_dims = dict() # Parse Descriptors and get their dims and their 'final' string form. # in self.descriptors[key] for key in self.get_input_descriptor_names(): if key not in self.config: raise RuntimeError("{0}: object of type {1} needs to override" " get_input_descriptor_names()." "".format(sys.argv[0], str(type(self)))) descriptor_string = self.config[key] # input string. assert isinstance(descriptor_string, str) desc = self.convert_to_descriptor(descriptor_string, all_layers) desc_dim = self.get_dim_for_descriptor(desc, all_layers) desc_norm_str = desc.str() # desc_output_str contains the "final" component names, those that # appear in the actual config file (i.e. not names like # 'layer.auxiliary_output'); that's how it differs from desc_norm_str. # Note: it's possible that the two strings might be the same in # many, even most, cases-- it depends whether # output_name(self, auxiliary_output) # returns self.get_name() + '.' + auxiliary_output # when auxiliary_output is not None. # That's up to the designer of the layer type. desc_output_str = self.get_string_for_descriptor(desc, all_layers) self.descriptors[key] = {'string': desc, 'normalized-string': desc_norm_str, 'final-string': desc_output_str, 'dim': desc_dim} # the following helps to check the code by parsing it again. desc2 = self.convert_to_descriptor(desc_norm_str, all_layers) desc_norm_str2 = desc2.str() # if the following ever fails we'll have to do some debugging. if desc_norm_str != desc_norm_str2: raise RuntimeError("Likely code error: '{0}' != '{1}'" "".format(desc_norm_str, desc_norm_str2)) def str(self): """Converts 'this' to a string which could be printed to an xconfig file; in xconfig_to_configs.py we actually expand all the lines to strings and write it as xconfig.expanded as a reference (so users can see any defaults). """ list_of_entries = ['{0} name={1}'.format(self.layer_type, self.name)] for key, value in sorted(self.config.items()): if isinstance(value, str) and re.search('=', value): # the value is a string that contains an '=' sign, so we need to # enclose it in double-quotes, otherwise we woudldn't be able to # parse from that output. if re.search('"', value): print("Warning: config '{0}={1}' contains both double-quotes " "and equals sign; it will not be possible to parse it " "from the file.".format(key, value), file=sys.stderr) list_of_entries.append('{0}="{1}"'.format(key, value)) else: list_of_entries.append('{0}={1}'.format(key, value)) return ' '.join(list_of_entries) def __str__(self): return self.str() def normalize_descriptors(self): """Converts any config variables in self.config which correspond to Descriptors, into a 'normalized form' derived from parsing them as Descriptors, replacing things like [-1] with the actual layer names, and regenerating them as strings. We stored this when the object was initialized, in self.descriptors; this function just copies them back to the config. """ for key, desc_str_dict in self.descriptors.items(): self.config[key] = desc_str_dict['normalized-string'] def convert_to_descriptor(self, descriptor_string, all_layers): """Convenience function intended to be called from child classes, converts a string representing a descriptor ('descriptor_string') into an object of type Descriptor, and returns it. It needs 'self' and 'all_layers' (where 'all_layers' is a list of objects of type XconfigLayerBase) so that it can work out a list of the names of other layers, and get dimensions from them. """ prev_names = xutils.get_prev_names(all_layers, self) tokens = xutils.tokenize_descriptor(descriptor_string, prev_names) pos = 0 (descriptor, pos) = xutils.parse_new_descriptor(tokens, pos, prev_names) # note: 'pos' should point to the 'end of string' marker # that terminates 'tokens'. if pos != len(tokens) - 1: raise RuntimeError("Parsing Descriptor, saw junk at end: {0}" "".format(' '.join(tokens[pos:-1]))) return descriptor def get_dim_for_descriptor(self, descriptor, all_layers): """Returns the dimension of a Descriptor object. This is a convenience function used in set_configs. """ layer_to_dim_func = \ lambda name: xutils.get_dim_from_layer_name(all_layers, self, name) return descriptor.dim(layer_to_dim_func) def get_string_for_descriptor(self, descriptor, all_layers): """Returns the 'final' string form of a Descriptor object, as could be used in config files. This is a convenience function provided for use in child classes; """ layer_to_string_func = \ lambda name: xutils.get_string_from_layer_name(all_layers, self, name) return descriptor.config_string(layer_to_string_func) def get_name(self): """Returns the name of this layer, e.g. 'affine1'. It does not necessarily correspond to a component name. """ return self.name ###### Functions that might be overridden by the child class: ##### def set_default_configs(self): """Child classes should override this. """ raise Exception("Child classes must override set_default_configs().") def set_derived_configs(self): """This is expected to be called after set_configs and before check_configs(). """ if 'dim' in self.config and self.config['dim'] <= 0: self.config['dim'] = self.descriptors['input']['dim'] def check_configs(self): """child classes should override this. """ pass def get_input_descriptor_names(self): """This function, which may be (but usually will not have to be) overridden by child classes, returns a list of names of the input descriptors expected by this component. Typically this would just return ['input'] as most layers just have one 'input'. However some layers might require more inputs (e.g. cell state of previous LSTM layer in Highway LSTMs). It is used in the function 'normalize_descriptors()'. This implementation will work for layer types whose only Descriptor-valued config is 'input'. If a child class adds more inputs, or does not have an input (e.g. the XconfigInputLayer), it should override this function's implementation to something like: `return ['input', 'input2']` """ return ['input'] def auxiliary_outputs(self): """Returns a list of all auxiliary outputs that this layer supports. These are either 'None' for the regular output, or a string (e.g. 'projection' or 'memory_cell') for any auxiliary outputs that the layer might provide. Most layer types will not need to override this. """ return [None] def output_name(self, auxiliary_output=None): """Called with auxiliary_output is None, this returns the component-node name of the principal output of the layer (or if you prefer, the text form of a descriptor that gives you such an output; such as Append(some_node, some_other_node)). The 'auxiliary_output' argument is a text value that is designed for extensions to layers that have additional auxiliary outputs. For example, to implement a highway LSTM you need the memory-cell of a layer, so you might allow auxiliary_output='memory_cell' for such a layer type, and it would return the component node or a suitable Descriptor: something like 'lstm3.c_t' """ raise Exception("Child classes must override output_name()") def output_dim(self, auxiliary_output=None): """The dimension that this layer outputs. The 'auxiliary_output' parameter is for layer types which support auxiliary outputs. """ raise Exception("Child classes must override output_dim()") def get_full_config(self): """This function returns lines destined for the 'full' config format, as would be read by the C++ programs. Since the program xconfig_to_configs.py writes several config files, this function returns a list of pairs of the form (config_file_basename, line), e.g. something like [ ('init', 'input-node name=input dim=40'), ('ref', 'input-node name=input dim=40') ] which would be written to config_dir/init.config and config_dir/ref.config. """ raise Exception("Child classes must override get_full_config()") class XconfigInputLayer(XconfigLayerBase): """This class is for lines like 'input name=input dim=40' or 'input name=ivector dim=100' in the config file. """ def __init__(self, first_token, key_to_value, prev_names=None): assert first_token == 'input' XconfigLayerBase.__init__(self, first_token, key_to_value, prev_names) def set_default_configs(self): self.config = {'dim': -1} def check_configs(self): if self.config['dim'] <= 0: raise RuntimeError("Dimension of input-layer '{0}'" "should be positive.".format(self.name)) def get_input_descriptor_names(self): return [] # there is no 'input' field in self.config. def output_name(self, auxiliary_outputs=None): # there are no auxiliary outputs as this layer will just pass the input assert auxiliary_outputs is None return self.name def output_dim(self, auxiliary_outputs=None): # there are no auxiliary outputs as this layer will just pass the input assert auxiliary_outputs is None return self.config['dim'] def get_full_config(self): # unlike other layers the input layers need to be printed in # 'init.config' (which initializes the neural network prior to the LDA) ans = [] for config_name in ['init', 'ref', 'final']: ans.append((config_name, 'input-node name={0} dim={1}'.format(self.name, self.config['dim']))) return ans class XconfigTrivialOutputLayer(XconfigLayerBase): """ This class is for lines like 'output name=output input=Append(input@-1, input@0, input@1, ReplaceIndex(ivector, t, 0))' This is for outputs that are not really output "layers" (there is no affine transform or nonlinearity), they just directly map to an output-node in nnet3. Parameters of the class, and their defaults: input='[-1]' : Descriptor giving the input of the layer. objective-type=linear : the only other choice currently is 'quadratic', for use in regression problems output-delay=0 : Can be used to shift the frames on the output, equivalent to delaying labels by this many frames (positive value increases latency in online decoding but may help if you're using unidirectional LSTMs. """ def __init__(self, first_token, key_to_value, prev_names=None): assert first_token == 'output' 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, 'objective-type': 'linear', 'output-delay': 0} def check_configs(self): if self.config['objective-type'] != 'linear' and \ self.config['objective-type'] != 'quadratic': raise RuntimeError("In output, objective-type has" " invalid value {0}" "".format(self.config['objective-type'])) def output_name(self, auxiliary_outputs=None): # there are no auxiliary outputs as this layer will just pass the output # of the previous layer assert auxiliary_outputs is None return self.name def output_dim(self, auxiliary_outputs=None): assert auxiliary_outputs is None # note: each value of self.descriptors is (descriptor, dim, normalized-string, output-string). return self.descriptors['input']['dim'] def get_full_config(self): # the input layers need to be printed in 'init.config' (which # initializes the neural network prior to the LDA), in 'ref.config', # which is a version of the config file used for getting left and right # context (it doesn't read anything for the LDA-like transform). # In 'full.config' we write everything, this is just for reference, # and also for cases where we don't use the LDA-like transform. ans = [] # note: each value of self.descriptors is (descriptor, dim, # normalized-string, output-string). # by 'output-string' we mean a string that can appear in # config-files, i.e. it contains the 'final' names of nodes. descriptor_final_str = self.descriptors['input']['final-string'] objective_type = self.config['objective-type'] output_delay = self.config['output-delay'] if output_delay != 0: descriptor_final_str = ( 'Offset({0}, {1})'.format(descriptor_final_str, output_delay)) for config_name in ['ref', 'final']: ans.append((config_name, 'output-node name={0} input={1} ' 'objective={2}'.format( self.name, descriptor_final_str, objective_type))) return ans class XconfigOutputLayer(XconfigLayerBase): """This class is for lines like 'output-layer name=output dim=4257 input=Append(input@-1, input@0, input@1, ReplaceIndex(ivector, t, 0))' By default this includes a log-softmax component. The parameters are initialized to zero, as this empirically tends to be the best approach for output layers. Parameters of the class, and their defaults: input='[-1]' : Descriptor giving the input of the layer. dim=None : Output dimension of layer, will normally equal the number of pdfs. bottleneck-dim=None : Bottleneck dimension of layer: if supplied, instead of an affine component we'll have a linear then affine, so a linear bottleneck, with the linear part constrained to be orthonormal. include-log-softmax=true : setting it to false will omit the log-softmax component- useful for chain models. objective-type=linear : the only other choice currently is 'quadratic', for use in regression problems learning-rate-factor=1.0 : Learning rate factor for the final affine component, multiplies the standard learning rate. normally you'll leave this as-is, but for xent regularization output layers for chain models you'll want to set learning-rate-factor=(0.5/xent_regularize), normally learning-rate-factor=5.0 since xent_regularize is normally 0.1. max-change=1.5 : Can be used to change the max-change parameter in the affine component; this affects how much the matrix can change on each iteration. l2-regularize=0.0: Set this to a nonzero value (e.g. 1.0e-05) to add l2 regularization on the parameter norm for the affine component. output-delay=0 : Can be used to shift the frames on the output, equivalent to delaying labels by this many frames (positive value increases latency in online decoding but may help if you're using unidirectional LSTMs. ng-affine-options='' : Can be used supply non-default options to the affine layer (intended for the natural gradient but can be an arbitrary string to be added to the config line. e.g. 'update-period=2'.). ng-linear-options='' : Options, like ng-affine-options, that are passed to the LinearComponent, only in bottleneck layers (i.e. if bottleneck-dim is supplied). """ def __init__(self, first_token, key_to_value, prev_names=None): assert first_token == 'output-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, 'bottleneck-dim': -1, 'orthonormal-constraint': 1.0, # orthonormal-constraint only matters if bottleneck-dim is set. 'include-log-softmax': True, # this would be false for chain models 'objective-type': 'linear', # see Nnet::ProcessOutputNodeConfigLine in # nnet-nnet.cc for other options 'output-delay': 0, 'ng-affine-options': '', 'ng-linear-options': '', # only affects bottleneck output layers. # The following are just passed through to the affine # component, and (in the bottleneck case) the linear # component. 'learning-rate-factor': '', # effective default: 1.0 'l2-regularize': '', # effective default: 0.0 'max-change': 1.5, # The following are passed through to the affine component only. # It tends to be beneficial to initialize the output layer with # zero values, unlike the hidden layers. 'param-stddev': 0.0, 'bias-stddev': 0.0, } def check_configs(self): if self.config['dim'] <= -1: raise RuntimeError("In output-layer, dim has invalid value {0}" "".format(self.config['dim'])) if self.config['objective-type'] != 'linear' and \ self.config['objective-type'] != 'quadratic': raise RuntimeError("In output-layer, objective-type has" " invalid value {0}" "".format(self.config['objective-type'])) if self.config['orthonormal-constraint'] <= 0.0: raise RuntimeError("output-layer does not support negative (floating) " "orthonormal constraint; use a separate linear-component " "followed by batchnorm-component.") def auxiliary_outputs(self): auxiliary_outputs = ['affine'] if self.config['include-log-softmax']: auxiliary_outputs.append('log-softmax') return auxiliary_outputs def output_name(self, auxiliary_output=None): if auxiliary_output is None: # Note: nodes of type output-node in nnet3 may not be accessed in # Descriptors, so calling this with auxiliary_outputs=None doesn't # make sense. raise RuntimeError("Outputs of output-layer may not be used by other" " layers") if auxiliary_output in self.auxiliary_outputs(): return '{0}.{1}'.format(self.name, auxiliary_output) else: raise RuntimeError("Unknown auxiliary output name {0}" "".format(auxiliary_output)) def output_dim(self, auxiliary_output=None): if auxiliary_output is None: # Note: nodes of type output-node in nnet3 may not be accessed in # Descriptors, so calling this with auxiliary_outputs=None doesn't # make sense. raise RuntimeError("Outputs of output-layer may not be used by other" " layers") return self.config['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 LSTM initialization # so 'ref' and 'final' configs are the same. ans.append((config_name, line)) return ans def _generate_config(self): configs = [] # note: each value of self.descriptors is (descriptor, dim, # normalized-string, output-string). # by 'descriptor_final_string' we mean a string that can appear in # config-files, i.e. it contains the 'final' names of nodes. descriptor_final_string = self.descriptors['input']['final-string'] input_dim = self.descriptors['input']['dim'] output_dim = self.config['dim'] bottleneck_dim = self.config['bottleneck-dim'] objective_type = self.config['objective-type'] include_log_softmax = self.config['include-log-softmax'] output_delay = self.config['output-delay'] affine_options = self.config['ng-affine-options'] for opt in [ 'learning-rate-factor', 'l2-regularize', 'max-change', 'param-stddev', 'bias-stddev' ]: if self.config[opt] != '': affine_options += ' {0}={1}'.format(opt, self.config[opt]) cur_node = descriptor_final_string cur_dim = input_dim if bottleneck_dim >= 0: if bottleneck_dim == 0 or bottleneck_dim >= input_dim or bottleneck_dim >= output_dim: raise RuntimeError("Bottleneck dim has value that does not make sense: {0}".format( bottleneck_dim)) # This is the bottleneck case (it doesn't necessarily imply we # will be using the features from the bottleneck; it's just a factorization # of the matrix into two pieces without a nonlinearity in between). # We don't include the l2-regularize option because it's useless # given the orthonormality constraint. linear_options = self.config['ng-linear-options'] for opt in [ 'learning-rate-factor', 'l2-regularize', 'max-change' ]: if self.config[opt] != '': linear_options += ' {0}={1}'.format(opt, self.config[opt]) # note: by default the LinearComponent uses natural gradient. line = ('component name={0}.linear type=LinearComponent ' 'orthonormal-constraint={1} param-stddev={2} ' 'input-dim={3} output-dim={4} max-change=0.75 {5}' ''.format(self.name, self.config['orthonormal-constraint'], self.config['orthonormal-constraint'] / math.sqrt(input_dim), input_dim, bottleneck_dim, linear_options)) configs.append(line) line = ('component-node name={0}.linear component={0}.linear input={1}' ''.format(self.name, cur_node)) configs.append(line) cur_node = '{0}.linear'.format(self.name) cur_dim = bottleneck_dim line = ('component name={0}.affine' ' type=NaturalGradientAffineComponent' ' input-dim={1} output-dim={2} {3}' ''.format(self.name, cur_dim, output_dim, affine_options)) configs.append(line) line = ('component-node name={0}.affine' ' component={0}.affine input={1}' ''.format(self.name, cur_node)) configs.append(line) cur_node = '{0}.affine'.format(self.name) if include_log_softmax: line = ('component name={0}.log-softmax' ' type=LogSoftmaxComponent dim={1}' ''.format(self.name, output_dim)) configs.append(line) line = ('component-node name={0}.log-softmax' ' component={0}.log-softmax input={1}' ''.format(self.name, cur_node)) configs.append(line) cur_node = '{0}.log-softmax'.format(self.name) if output_delay != 0: cur_node = 'Offset({0}, {1})'.format(cur_node, output_delay) line = ('output-node name={0} input={1} ' 'objective={2}'.format( self.name, cur_node, objective_type)) configs.append(line) return configs class XconfigBasicLayer(XconfigLayerBase): """This class is for parsing lines like 'relu-renorm-layer name=layer1 dim=1024 input=Append(-3,0,3)' or: 'sigmoid-layer name=layer1 dim=1024 input=Append(-3,0,3)' which specify addition of an affine component and a sequence of non-linearities. Here, the name of the layer itself dictates the sequence of nonlinearities that are applied after the affine component; the name should contain some combination of 'relu', 'renorm', 'sigmoid' and 'tanh', and these nonlinearities will be added along with the affine component. The dimension specified is the output dim; the input dim is worked out from the input descriptor. This class supports only nonlinearity types that do not change the dimension; we can create another layer type to enable the use p-norm and similar dimension-reducing nonlinearities. See other configuration values below. Parameters of the class, and their defaults: input='[-1]' [Descriptor giving the input of the layer.] dim=-1 [Output dimension of layer, e.g. 1024] bottleneck-dim=-1 [If you set this, a linear bottleneck is added, so we project to first bottleneck-dim then to dim. The first of the two matrices is constrained to be orthonormal.] 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]. add-log-stddev=False [If true, the log of the stddev of the output of renorm layer is appended as an additional dimension of the layer's output] l2-regularize=0.0 [Set this to a nonzero value (e.g. 1.0e-05) to add l2 regularization on the parameter norm for this component. """ def __init__(self, first_token, key_to_value, prev_names=None): 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, 'bottleneck-dim': -1, # Deprecated! Use tdnnf-layer for # factorized TDNNs, or prefinal-layer # for bottlenecks just before the output. 'self-repair-scale': 1.0e-05, 'target-rms': 1.0, 'ng-affine-options': '', 'ng-linear-options': '', # only affects bottleneck layers. 'dropout-proportion': 0.5, # dropout-proportion only # affects layers with # 'dropout' in the name 'dropout-per-dim': False, # if dropout-per-dim=true, the dropout # mask is shared across time. 'dropout-per-dim-continuous': False, # if you set this, it's # like dropout-per-dim but with a # continuous-valued (not zero-one) mask. 'add-log-stddev': False, # the following are not really inspected by this level of # code, just passed through to the affine component if # their value is not ''. 'bias-stddev': '', 'l2-regularize': '', 'learning-rate-factor': '', 'max-change': 0.75 } def check_configs(self): if self.config['dim'] < 0: raise RuntimeError("dim has invalid value {0}".format(self.config['dim'])) b = self.config['bottleneck-dim'] if b >= 0 and (b >= self.config['dim'] or b == 0): raise RuntimeError("bottleneck-dim has an invalid value {0}".format(b)) 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'] != '' and self.config['learning-rate-factor'] <= 0.0): raise RuntimeError("learning-rate-factor has invalid value {0}" .format(self.config['learning-rate-factor'])) def output_name(self, auxiliary_output=None): # at a later stage we might want to expose even the pre-nonlinearity # vectors assert auxiliary_output is 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 output_dim(self, auxiliary_output=None): output_dim = self.config['dim'] # If not set, the output-dim defaults to the input-dim. if output_dim <= 0: self.config['dim'] = self.descriptors['input']['dim'] return 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): output_dim = self.output_dim() self_repair_scale = self.config['self-repair-scale'] target_rms = self.config['target-rms'] affine_options = self.config['ng-affine-options'] for opt_name in [ 'max-change', 'learning-rate-factor', 'bias-stddev', 'l2-regularize' ]: value = self.config[opt_name] if value != '': affine_options += ' {0}={1}'.format(opt_name, value) # The output of the affine component needs to have one dimension fewer in order to # get the required output dim, if the final 'renorm' component has 'add-log-stddev' set # (since in that case it increases the dimension by one). if self.config['add-log-stddev']: output_dim -= 1 if not self.layer_type.split('-')[-2] == "renorm": raise RuntimeError("add-log-stddev cannot be true unless " "there is a final 'renorm' component.") configs = [] cur_dim = input_dim cur_node = input_desc # First the affine node (or linear then affine, if bottleneck). if self.config['bottleneck-dim'] > 0: # The 'bottleneck-dim' option is deprecated and may eventually be # removed. Best to use tdnnf-layer if you want factorized TDNNs. # This is the bottleneck case (it doesn't necessarily imply we # will be using the features from the bottleneck; it's just a factorization # of the matrix into two pieces without a nonlinearity in between). # We don't include the l2-regularize option because it's useless # given the orthonormality constraint. linear_options = self.config['ng-linear-options'] for opt_name in [ 'max-change', 'learning-rate-factor' ]: value = self.config[opt_name] if value != '': linear_options += ' {0}={1}'.format(opt_name, value) bottleneck_dim = self.config['bottleneck-dim'] # note: by default the LinearComponent uses natural gradient. line = ('component name={0}.linear type=LinearComponent ' 'input-dim={1} orthonormal-constraint=1.0 output-dim={2} {3}' ''.format(self.name, input_dim, bottleneck_dim, linear_options)) configs.append(line) line = ('component-node name={0}.linear component={0}.linear input={1}' ''.format(self.name, cur_node)) configs.append(line) cur_node = '{0}.linear'.format(self.name) cur_dim = bottleneck_dim line = ('component name={0}.affine type=NaturalGradientAffineComponent' ' input-dim={1} output-dim={2} {3}' ''.format(self.name, cur_dim, output_dim, affine_options)) configs.append(line) line = ('component-node name={0}.affine component={0}.affine input={1}' ''.format(self.name, cur_node)) configs.append(line) cur_node = '{0}.affine'.format(self.name) for i, nonlinearity in enumerate(nonlinearities): if nonlinearity == 'relu': line = ('component name={0}.{1} type=RectifiedLinearComponent dim={2}' ' self-repair-scale={3}' ''.format(self.name, nonlinearity, output_dim, self_repair_scale)) elif nonlinearity == 'sigmoid': line = ('component name={0}.{1}' ' type=SigmoidComponent dim={2}' ' self-repair-scale={3}' ''.format(self.name, nonlinearity, output_dim, self_repair_scale)) elif nonlinearity == 'tanh': line = ('component name={0}.{1}' ' type=TanhComponent dim={2}' ' self-repair-scale={3}' ''.format(self.name, nonlinearity, output_dim, self_repair_scale)) elif nonlinearity == 'renorm': add_log_stddev = "false" if i == len(nonlinearities) - 1: add_log_stddev = ("true" if self.config['add-log-stddev'] else "false") line = ('component name={0}.{1}' ' type=NormalizeComponent dim={2}' ' target-rms={3}' ' add-log-stddev={4}' ''.format(self.name, nonlinearity, output_dim, target_rms, add_log_stddev)) elif nonlinearity == 'batchnorm': line = ('component name={0}.{1}' ' type=BatchNormComponent dim={2} target-rms={3}' ''.format(self.name, nonlinearity, output_dim, target_rms)) elif nonlinearity == 'so': line = ('component name={0}.{1}' ' type=ScaleAndOffsetComponent dim={2} max-change=0.5 ' ''.format(self.name, nonlinearity, output_dim)) elif nonlinearity == 'dropout': if not (self.config['dropout-per-dim'] or self.config['dropout-per-dim-continuous']): line = ('component name={0}.{1} type=DropoutComponent ' 'dim={2} dropout-proportion={3}'.format( self.name, nonlinearity, output_dim, self.config['dropout-proportion'])) else: continuous_opt='continuous=true' if self.config['dropout-per-dim-continuous'] else '' line = ('component name={0}.dropout type=GeneralDropoutComponent ' 'dim={1} dropout-proportion={2} {3}'.format( self.name, output_dim, self.config['dropout-proportion'], continuous_opt)) 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 class XconfigFixedAffineLayer(XconfigLayerBase): """ This class is for lines like 'fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=foo/bar/lda.mat' The output dimension of the layer may be specified via 'dim=xxx', but if not specified, the dimension defaults to the same as the input. Note: we don't attempt to read that file at the time the config is created, because in the recipes, that file is created after the config files. See other configuration values below. Parameters of the class, and their defaults: input='[-1]' [Descriptor giving the input of the layer.] dim=None [Output dimension of layer; defaults to the same as the input dim.] affine-transform-file='' [Must be specified.] delay=0 [Optional delay for the output-node in init.config] """ def __init__(self, first_token, key_to_value, prev_names=None): assert first_token == 'fixed-affine-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, 'affine-transform-file': '', 'delay': 0, 'write-init-config': True} def check_configs(self): if self.config['affine-transform-file'] is None: raise RuntimeError("affine-transform-file must be set.") def output_name(self, auxiliary_output=None): # Fixed affine layer computes only one vector, there are no intermediate # vectors. assert auxiliary_output is None return self.name def output_dim(self, auxiliary_output=None): output_dim = self.config['dim'] # If not set, the output-dim defaults to the input-dim. if output_dim <= 0: output_dim = self.descriptors['input']['dim'] return output_dim def get_full_config(self): ans = [] # note: each value of self.descriptors is (descriptor, dim, # normalized-string, output-string). # by 'descriptor_final_string' we mean a string that can appear in # config-files, i.e. it contains the 'final' names of nodes. descriptor_final_string = self.descriptors['input']['final-string'] input_dim = self.descriptors['input']['dim'] output_dim = self.output_dim() transform_file = self.config['affine-transform-file'] if self.config['write-init-config']: if self.config['delay'] != 0: line = 'component name={0}.delayed type=NoOpComponent dim={1}'.format(self.name, input_dim) ans.append(('init', line)) line = 'component-node name={0}.delayed component={0}.delayed input={1}'.format(self.name, descriptor_final_string) ans.append(('init', line)) line = 'output-node name=output input=Offset({0}.delayed, {1})'.format(self.name, self.config['delay']) ans.append(('init', line)) else: # to init.config we write an output-node with the name 'output' and # with a Descriptor equal to the descriptor that's the input to this # layer. This will be used to accumulate stats to learn the LDA transform. line = 'output-node name=output input={0}'.format(descriptor_final_string) ans.append(('init', line)) # write the 'real' component to final.config line = 'component name={0} type=FixedAffineComponent matrix={1}'.format( self.name, transform_file) ans.append(('final', line)) # write a random version of the component, with the same dims, to ref.config line = 'component name={0} type=FixedAffineComponent input-dim={1} output-dim={2}'.format( self.name, input_dim, output_dim) ans.append(('ref', line)) # the component-node gets written to final.config and ref.config. line = 'component-node name={0} component={0} input={1}'.format( self.name, descriptor_final_string) ans.append(('final', line)) ans.append(('ref', line)) return ans class XconfigAffineLayer(XconfigLayerBase): """ This class is for lines like 'affine-layer name=affine input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0))' The output dimension of the layer may be specified via 'dim=xxx', but if not specified, the dimension defaults to the same as the input. Note: we don't attempt to read that file at the time the config is created, because in the recipes, that file is created after the config files. See other configuration values below. Parameters of the class, and their defaults: input='[-1]' [Descriptor giving the input of the layer.] dim=None [Output dimension of layer; defaults to the same as the input dim.] l2-regularize=0.0 [Set this to a nonzero value (e.g. 1.0e-05) to add l2 regularization on the parameter norm for the affine component.] """ def __init__(self, first_token, key_to_value, prev_names=None): assert first_token == 'affine-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. # use None for optional parameters as we want to default to the C++ defaults # C++ component provides more options but I will just expose these for now # Note : The type of the parameter is determined based on the value assigned # so please use decimal point if your parameter is a float self.config = {'input': '[-1]', 'dim': -1, 'param-stddev': -1.0, # this has to be initialized to 1/sqrt(input_dim) 'bias-stddev': 1.0, 'bias-mean': 0.0, 'max-change': 0.75, 'l2-regularize': 0.0, 'learning-rate-factor': 1.0, 'ng-affine-options': ''} def set_derived_configs(self): super(XconfigAffineLayer, self).set_derived_configs() if self.config['param-stddev'] < 0: self.config['param-stddev'] = 1.0 / math.sqrt(self.descriptors['input']['dim']) def check_configs(self): if self.config['dim'] <= 0: raise RuntimeError("dim specified is invalid") def output_name(self, auxiliary_output=None): # affine layer computes only one vector, there are no intermediate # vectors. assert auxiliary_output is None return self.name def output_dim(self, auxiliary_output=None): output_dim = self.config['dim'] # If not set, the output-dim defaults to the input-dim. if output_dim <= 0: output_dim = self.descriptors['input']['dim'] return output_dim def get_full_config(self): ans = [] # note: each value of self.descriptors is (descriptor, dim, # normalized-string, output-string). # by 'descriptor_final_string' we mean a string that can appear in # config-files, i.e. it contains the 'final' names of nodes. descriptor_final_string = self.descriptors['input']['final-string'] input_dim = self.descriptors['input']['dim'] output_dim = self.output_dim() option_string = '' for key in ['param-stddev', 'bias-stddev', 'bias-mean', 'max-change', 'l2-regularize']: option_string += ' {0}={1}'.format(key, self.config[key]) option_string += self.config['ng-affine-options'] conf_lines = [] # write the 'real' component to final.config conf_lines.append('component name={n} type=NaturalGradientAffineComponent ' 'input-dim={i} output-dim={o} {opts}'.format(n=self.name, i=input_dim, o=output_dim, opts=option_string)) # the component-node gets written to final.config and ref.config. conf_lines.append('component-node name={0} component={0} input={1}'.format(self.name, descriptor_final_string)) # the config is same for both final and ref configs for conf_name in ['final', 'ref']: for line in conf_lines: ans.append((conf_name, line)) return ans class XconfigIdctLayer(XconfigLayerBase): """ This class is for lines like 'idct-layer name=idct dim=40 cepstral-lifter=22 affine-transform-file=foo/bar/idct.mat' This is used to convert input MFCC-features to Filterbank featurs. The affine transformation is written out to the file specified via 'affine-transform-file=xxx'. The output dimension of the layer may be specified via 'dim=xxx', but if not specified, the dimension defaults to the same as the input. See other configuration values below. Parameters of the class, and their defaults: input='[-1]' [Descriptor giving the input of the layer.] dim=None [Output dimension of layer; defaults to the same as the input dim.] cepstral-lifter=22 [Apply liftering co-efficient.] affine-transform-file='' [Must be specified.] include-in-init=false [You should set this to true if this precedes a `fixed-affine-layer` that is to be initialized via LDA] """ def __init__(self, first_token, key_to_value, prev_names=None): assert first_token == 'idct-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, 'cepstral-lifter': 22.0, 'affine-transform-file': '', 'include-in-init': False} def check_configs(self): if self.config['affine-transform-file'] is None: raise RuntimeError("affine-transform-file must be set.") def output_name(self, auxiliary_output=None): # Fixed affine layer computes only one vector, there are no intermediate # vectors. assert auxiliary_output is None return self.name def output_dim(self, auxiliary_output=None): output_dim = self.config['dim'] # If not set, the output-dim defaults to the input-dim. if output_dim <= 0: output_dim = self.descriptors['input']['dim'] return 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)) if self.config['include-in-init']: ans.append(('init', line)) return ans def _generate_config(self): # note: each value of self.descriptors is (descriptor, dim, # normalized-string, output-string). # by 'descriptor_final_string' we mean a string that can appear in # config-files, i.e. it contains the 'final' names of nodes. descriptor_final_string = self.descriptors['input']['final-string'] input_dim = self.descriptors['input']['dim'] output_dim = self.output_dim() transform_file = self.config['affine-transform-file'] idct_mat = common_lib.compute_idct_matrix( input_dim, output_dim, self.config['cepstral-lifter']) # append a zero column to the matrix, this is the bias of the fixed # affine component for n in range(0, output_dim): idct_mat[n].append(0) common_lib.write_kaldi_matrix(transform_file, idct_mat) configs = [] # write the 'real' component to final.config line = 'component name={0} type=FixedAffineComponent matrix={1}'.format( self.name, transform_file) configs.append(line) line = 'component-node name={0} component={0} input={1}'.format( self.name, descriptor_final_string) configs.append(line) return configs class XconfigExistingLayer(XconfigLayerBase): """ This class is used to internally convert component-nodes in an existing model into lines like 'existing name=tdnn1.affine dim=40'. Layers of this type are not presented in any actual xconfig or config files, but are created internally for all component nodes in an existing neural net model to use as input to other layers in xconfig. (i.e. get_model_component_info function, which is called in steps/nnet3/xconfig_to_configs.py, parses the name and dimension of component-nodes used in the existing model using the nnet3-info and returns a list of 'existing' layers.) This class is useful in cases like transferring existing model and using {input, output, component}-nodes in this model as input to new layers. """ def __init__(self, first_token, key_to_value, prev_names=None): assert first_token == 'existing' XconfigLayerBase.__init__(self, first_token, key_to_value, prev_names) def set_default_configs(self): self.config = { 'dim': -1} def check_configs(self): if self.config['dim'] <= 0: raise RuntimeError("Dimension of existing-layer '{0}'" "should be positive.".format(self.name)) def get_input_descriptor_names(self): return [] # there is no 'input' field in self.config. def output_name(self, auxiliary_outputs=None): # there are no auxiliary outputs as this layer will just pass the input assert auxiliary_outputs is None return self.name def output_dim(self, auxiliary_outputs=None): # there are no auxiliary outputs as this layer will just pass the input assert auxiliary_outputs is None return self.config['dim'] def get_full_config(self): # unlike other layers the existing layers should not to be printed in # any '*.config' ans = [] return ans class XconfigSpecAugmentLayer(XconfigLayerBase): """This class is for parsing lines like 'spec-augment-layer name=spec-augment freq-max-proportion=0.5 time-zeroed-proportion=0.2 time-mask-max-frames=10' which will produce a component of type GeneralDropoutComponent (to do the frequency-domain part) and then one of type SpecaugmentTimeMaskComponent (to do the time part). Parameters of the class, and their defaults: input='[-1]' [Descriptor giving the input of the layer.] freq-max-proportion=0.5 [The maximum proportion of the frequency space that might be zeroed out] time-zeroed-proportion=0.2 [The proportion of time frames that will be zeroed out] time-mask-max-frames=20 [The maximum length of a zeroed region in the time axis, in frames.] include-in-init=false [You should set this to true if this precedes a `fixed-affine-layer` that is to be initialized via LDA] """ def __init__(self, first_token, key_to_value, prev_names=None): XconfigLayerBase.__init__(self, first_token, key_to_value, prev_names) def set_default_configs(self): self.config = {'input': '[-1]', 'freq-max-proportion': 0.5, 'time-zeroed-proportion': 0.2, 'time-mask-max-frames': 20, 'include-in-init': False} def check_configs(self): assert (self.config['freq-max-proportion'] > 0.0 and self.config['freq-max-proportion'] < 1.0 and self.config['time-zeroed-proportion'] > 0.0 and self.config['time-zeroed-proportion'] < 1.0 and self.config['time-mask-max-frames'] >= 1) def output_name(self, auxiliary_output=None): assert auxiliary_output is None return '{0}.time-mask'.format(self.name) 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 = [] 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)) if self.config['include-in-init']: ans.append(('init', line)) return ans def _generate_config(self): # 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'] freq_max_proportion = self.config['freq-max-proportion'] time_zeroed_proportion = self.config['time-zeroed-proportion'] time_mask_max_frames = self.config['time-mask-max-frames'] configs = [] line = ('component name={0}.freq-mask type=GeneralDropoutComponent dim={1} specaugment-max-proportion={2}'.format( self.name, input_dim, freq_max_proportion)) configs.append(line) line = ('component-node name={0}.freq-mask component={0}.freq-mask input={1}'.format( self.name, input_desc)) configs.append(line) line = ('component name={0}.time-mask type=SpecAugmentTimeMaskComponent dim={1} ' 'zeroed-proportion={2} time-mask-max-frames={3}'.format( self.name, input_dim, time_zeroed_proportion, time_mask_max_frames)) configs.append(line) line = ('component-node name={0}.time-mask component={0}.time-mask input={0}.freq-mask'.format( self.name)) configs.append(line) return configs def test_layers(): # for some config lines that should be printed the same way as they # are read, check that this is the case. for x in ['input name=input dim=30']: assert str(config_line_to_object(x, [])) == x |