ModelConfig.proto 18.7 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Z
zhangjinchao01 已提交
2 3 4 5 6 7 8 9 10 11 12 13

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
Y
Yu Yang 已提交
14
syntax = "proto2";
Z
zhangjinchao01 已提交
15 16 17 18 19 20 21 22

import "ParameterConfig.proto";

package paddle;

/**
 * Various structs for the configuration of a neural network
 */
Y
Yu Yang 已提交
23

Z
zhangjinchao01 已提交
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

message ExternalConfig {
  repeated string layer_names = 1;
  repeated string input_layer_names = 2;
  repeated string output_layer_names = 3;
}

message ActivationConfig {
  // identity: f(x) = x
  // sigmoid: f(x) = 1 / (1 + exp(-x))
  // logistic: f(x) = (1 - exp(-x)) / (1+ exp(-x))
  // softmax: y_i = f(x_i) = exp(x_i) / (\sum_i exp(x_i))
  // relu: y = max(0, x)
  required string type = 1;
};

message ConvConfig {
  // filter_size = 5, says that this layer will use
  // filters of size 5x5 pixels.
  required uint32 filter_size = 1;

  // The image data dimensionality.
  // This value must be either 1, 2, 3, or a multiple of 4.
  required uint32 channels = 2;

  // stride = 1, indicates that the distance between
  // successive filter applications should be 1 pixel.
  required uint32 stride = 3;

  // padding = 4, instructs the net to implicitly
  // pad the images with a 4-pixel border of zeros.
  required uint32 padding = 4;

  // If groups = 4 together with the filters = 32 parameter,
  // they state that this convolutional layer is to have 4
  // groups of 32 filters. Each filter will connect to 8
  // input channels.
  required uint32 groups = 5;
  required uint32 filter_channels = 6;

  // The size of output feature map.
  required uint32 output_x = 7;

  // The size of input feature map.
  required uint32 img_size = 8;

  // caffe mode for output size coherence
  required bool caffe_mode = 9 [default = true];

  // if filter_size_y is set , this convolutional layer will use
  // filters of size filter_size * filter_size_y pixels.
  // if filter_size_y is not set, this convolutional layer will use
  // filters of size filter_size * filter_size
  required uint32 filter_size_y = 10;
  required uint32 padding_y = 11;
  required uint32 stride_y = 12;
L
Luo Tao 已提交
80 81

  // if not set, use output_x
L
Luo Tao 已提交
82
  optional uint32 output_y = 13;
L
Luo Tao 已提交
83 84

  // if not set, use img_size
L
Luo Tao 已提交
85
  optional uint32 img_size_y = 14;
Z
zhangjinchao01 已提交
86 87 88 89 90 91 92 93 94 95 96 97
}

message PoolConfig {
  // max or avg pooling
  required string pool_type = 1;
  required uint32 channels = 2;

  // Defines the size of the pooling region in
  // the x (equivalently, y) dimension.
  required uint32 size_x = 3;

  // Tell the net where in the input image to start the pooling.
98 99
  // start is deprecated now.
  optional uint32 start = 4;
Z
zhangjinchao01 已提交
100 101

  // Defines the stride size between successive pooling squares.
D
dangqingqing 已提交
102
  required uint32 stride = 5 [default = 1];
Z
zhangjinchao01 已提交
103 104 105 106 107 108 109 110 111 112 113 114

  // The size of output feature map.
  required uint32 output_x = 6;

  // The size of input feature map.
  required uint32 img_size = 7;

  // padding = 4, instructs the net to implicitly
  // pad the images with a 4-pixel border of zeros.
  optional uint32 padding = 8 [default = 0];

  // if not set, use size_x
D
dangqingqing 已提交
115
  optional uint32 size_y = 9;
Z
zhangjinchao01 已提交
116 117

  // if not set, use stride
D
dangqingqing 已提交
118
  optional uint32 stride_y = 10;
Z
zhangjinchao01 已提交
119 120

  // if not set, use output_x
D
dangqingqing 已提交
121
  optional uint32 output_y = 11;
Z
zhangjinchao01 已提交
122 123

  // if not set, use img_size
D
dangqingqing 已提交
124
  optional uint32 img_size_y = 12;
Z
zhangjinchao01 已提交
125 126

  // if not set, use padding
D
dangqingqing 已提交
127
  optional uint32 padding_y = 13;
Z
zhangjinchao01 已提交
128 129
}

Q
qijun 已提交
130
message SppConfig {
L
Luo Tao 已提交
131 132 133
  required ImageConfig image_conf = 1;
  required string pool_type = 2;
  required uint32 pyramid_height = 3;
Q
qijun 已提交
134 135
}

Z
zhangjinchao01 已提交
136 137 138 139 140 141 142 143 144 145 146 147 148
message NormConfig {
  // rnorm or cmrnorm
  required string norm_type = 1;
  required uint32 channels = 2;

  // rnorm: this defines the size of the local regions
  // used for response normalization.
  // cmrnorm: The size parameter indicates how many
  // nearby maps to use for normalization.
  required uint32 size = 3;

  // the parameters for normalization
  // u = u / (1+scale*sum(u^2 in window))^pow
Y
Yu Yang 已提交
149 150
  required double scale = 4;
  required double pow = 5;
Z
zhangjinchao01 已提交
151 152 153 154 155 156 157 158 159 160 161 162

  // The size of output feature map.
  required uint32 output_x = 6;

  // The size of input feature map.
  required uint32 img_size = 7;

  // normalize with fixed window or sliding window
  // u = u / (1+scale*sum(u^2 in window))^pow
  // fixed window: shared a fixed window for each value
  // sliding window: have a different window for each value
  optional bool blocked = 8;
L
Luo Tao 已提交
163 164

  // if not set, use output_x
L
Luo Tao 已提交
165
  optional uint32 output_y = 9;
L
Luo Tao 已提交
166 167

  // if not set, use img_size
L
Luo Tao 已提交
168
  optional uint32 img_size_y = 10;
Z
zhangjinchao01 已提交
169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191
}

message BlockExpandConfig {
  required uint32 channels = 1;

  required uint32 stride_x = 2;
  required uint32 stride_y = 3;

  required uint32 padding_x = 4;
  required uint32 padding_y = 5;

  required uint32 block_x = 6;
  required uint32 block_y = 7;

  // The size of output feature map.
  required uint32 output_x = 8;
  required uint32 output_y = 9;

  // The size of input feature map.
  required uint32 img_size_x = 10;
  required uint32 img_size_y = 11;
}

192
message MaxOutConfig {
L
Luo Tao 已提交
193
  required ImageConfig image_conf = 1;
194 195 196
  required uint32 groups = 2;
}

197 198 199 200
message RowConvConfig {
  required uint32 context_length = 1;
}

Z
zhangjinchao01 已提交
201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217
message ProjectionConfig {
  required string type = 1;
  required string name = 2;
  required uint64 input_size = 3;
  required uint64 output_size = 4;

  // For ShiftProjection
  optional int32 context_start = 5;
  optional int32 context_length = 6;
  optional bool trainable_padding = 7 [default = false];

  // For convolution
  optional ConvConfig conv_conf = 8;
  optional int32 num_filters = 9;

  // For IdentityOffsetProjection
  optional uint64 offset = 11 [default = 0];
Q
qijun 已提交
218 219 220

  // For pool
  optional PoolConfig pool_conf = 12;
Z
zhangjinchao01 已提交
221 222 223 224 225 226 227 228 229
}

message OperatorConfig {
  required string type = 1;
  repeated int32 input_indices = 2;
  repeated uint64 input_sizes = 3;
  required uint64 output_size = 4;

  // For DotMulOperator
Y
Yu Yang 已提交
230
  optional double dotmul_scale = 5 [default = 1.0];
Z
zhangjinchao01 已提交
231 232 233 234 235 236

  // For ConvOperator
  optional ConvConfig conv_conf = 6;
  optional int32 num_filters = 7;
}

L
liaogang 已提交
237
message BilinearInterpConfig {
L
liaogang 已提交
238
  // The size of input feature map.
L
Luo Tao 已提交
239
  required ImageConfig image_conf = 1;
L
liaogang 已提交
240
  // The size of output feature map.
L
Luo Tao 已提交
241 242
  required uint32 out_size_x = 2;
  required uint32 out_size_y = 3;
L
liaogang 已提交
243
}
Z
zhangjinchao01 已提交
244 245 246 247 248 249 250 251

message ImageConfig {
  // The image data dimensionality.
  // This value must be either 1, 2, 3, or a multiple of 4.
  required uint32 channels = 2;

  // The size of input feature map.
  required uint32 img_size = 8;
252
  optional uint32 img_size_y = 9;
Z
zhangjinchao01 已提交
253 254
}

Y
yuan 已提交
255 256 257 258 259 260 261
message PriorBoxConfig {
  repeated uint32 min_size = 1;
  repeated uint32 max_size = 2;
  repeated float aspect_ratio = 3;
  repeated float variance = 4;
}

D
dangqingqing 已提交
262 263 264 265 266 267 268
message PadConfig {
  required ImageConfig image_conf = 1;
  repeated uint32 pad_c = 2;
  repeated uint32 pad_h = 3;
  repeated uint32 pad_w = 4;
}

269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291
message MultiBoxLossConfig {
  required uint32 num_classes = 1;
  required float overlap_threshold = 2;
  required float neg_pos_ratio = 3;
  required float neg_overlap = 4;
  required uint32 background_id = 5;
  required uint32 input_num = 6;
  optional uint32 height = 7 [default = 1];
  optional uint32 width = 8 [default = 1];
}

message DetectionOutputConfig {
  required uint32 num_classes = 1;
  required float nms_threshold = 2;
  required uint32 nms_top_k = 3;
  required uint32 background_id = 4;
  required uint32 input_num = 5;
  required uint32 keep_top_k = 6;
  required float confidence_threshold = 7;
  optional uint32 height = 8 [default = 1];
  optional uint32 width = 9 [default = 1];
}

G
guosheng 已提交
292 293 294 295 296
message ClipConfig {
  required float clip_threshold_low = 1;
  required float clip_threshold_high = 2;
}

Z
zhangjinchao01 已提交
297 298 299 300 301 302 303 304 305 306 307 308
message LayerInputConfig {
  required string input_layer_name = 1;
  optional string input_parameter_name = 2;
  optional ConvConfig conv_conf = 3;
  optional PoolConfig pool_conf = 4;
  optional NormConfig norm_conf = 5;
  optional ProjectionConfig proj_conf = 6;
  optional BlockExpandConfig block_expand_conf = 7;
  optional ImageConfig image_conf = 8;
  // If the input layer has multi-output.
  // Set the argument name.
  optional string input_layer_argument = 9;
L
liaogang 已提交
309
  optional BilinearInterpConfig bilinear_interp_conf = 10;
L
liaogang 已提交
310
  optional MaxOutConfig maxout_conf = 11;
Q
qijun 已提交
311
  optional SppConfig spp_conf = 12;
Y
yuan 已提交
312
  optional PriorBoxConfig priorbox_conf = 13;
D
dangqingqing 已提交
313
  optional PadConfig pad_conf = 14;
314
  optional RowConvConfig row_conv_conf = 15;
315 316
  optional MultiBoxLossConfig multibox_loss_conf = 16;
  optional DetectionOutputConfig detection_output_conf = 17;
G
guosheng 已提交
317
  optional ClipConfig clip_conf = 18;
Z
zhangjinchao01 已提交
318 319 320
}

message LayerConfig {
Y
Yu Yang 已提交
321

Z
zhangjinchao01 已提交
322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337
  required string name = 1;
  required string type = 2;
  optional uint64 size = 3;
  //optional ActivationConfig activation = 4;
  optional string active_type = 4;
  repeated LayerInputConfig inputs = 5;
  optional string bias_parameter_name = 6;

  // This number must be a multiple of 16.
  optional uint32 num_filters = 7;

  // indicates that the biases of every filter in this layer
  // should be shared amongst all applications of that filter
  // (which is how convnets are usually trained). Setting this to
  // false will untie the biases, yielding a separate bias for
  // every location at which the filter is applied.
338
  optional bool shared_biases = 8 [default = false];
Z
zhangjinchao01 已提交
339 340 341 342 343 344 345 346 347

  // Valid values are ones that divide the area of the output
  // grid in this convolutional layer. For example if this layer
  // produces 32-channel 20x20 output grid, valid values of
  // partialSum are ones which divide 20*20 = 400.
  // I'll update this comments when confirmed
  optional uint32 partial_sum = 9;

  // for dropout
Y
Yu Yang 已提交
348
  optional double drop_rate = 10;
Z
zhangjinchao01 已提交
349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371

  // for HierarchicalSoftmaxLayer and NCELayer
  // the number of classes
  optional uint32 num_classes = 11;

  // the gpu device which the Layer's data in.
  // Only used by ParallelNeuralNetork. Ignored otherwise.
  optional int32 device = 12 [default = -1];

  // for recurrent layer. If true, the recurrence runs from the end to the beginning.
  optional bool reversed = 13 [default = false];

  // for lstmemory layer. Different types of nodes have different activation type.
  optional string active_gate_type  = 14;
  optional string active_state_type = 15;

  // For NCELayer
  // The number of random negative labels for each sample
  optional int32 num_neg_samples = 16 [default = 10];

  // For NCELayer
  // The distribution for generating the random negative labels.
  // A uniform distribution will be used if not provided
Y
Yu Yang 已提交
372
  repeated double neg_sampling_dist = 17 [packed = true];
Z
zhangjinchao01 已提交
373 374 375

  // For MaxLayer
  // default: output VALUE of MaxLayer. set this flag to true for output INDEX
Y
Yu Yang 已提交
376
  // INDEX will be put in Argument::value as double values.
Z
zhangjinchao01 已提交
377 378 379 380 381
  optional bool output_max_index = 19 [default = false];

  /// The filed number 20 have been deprecated.

  // For self-normalized estimation
Y
Yu Yang 已提交
382
  optional double softmax_selfnorm_alpha = 21 [default = 0.1];
Z
zhangjinchao01 已提交
383 384 385 386 387 388 389 390 391 392

  /// The filed numbers 22 and 23 have been deprecated.

  // for MDLstmLayer
  repeated bool directions = 24;

  // for CTCLayer
  optional bool norm_by_times = 25;

  // for CostLayers
Y
Yu Yang 已提交
393
  optional double coeff = 26 [default = 1.0];
Z
zhangjinchao01 已提交
394 395 396 397 398 399

  // for AverageLayer
  // can be set to: 'average', 'sum' or 'squarerootn'
  optional string average_strategy = 27;

  // for error clipping
Y
Yu Yang 已提交
400
  optional double error_clipping_threshold = 28 [default = 0.0];
Z
zhangjinchao01 已提交
401 402 403 404 405 406 407 408 409

  // for operators used by mixed layer
  repeated OperatorConfig operator_confs = 29;

  // for lambdaCost
  optional int32 NDCG_num = 30;
  optional int32 max_sort_size = 31;

  // for SlopeInterceptLayer
Y
Yu Yang 已提交
410 411
  optional double slope = 32;
  optional double intercept = 33;
Z
zhangjinchao01 已提交
412 413

  // for CosSimVecMatLayer and CosSimLayer
Y
Yu Yang 已提交
414
  optional double cos_scale = 34;
Z
zhangjinchao01 已提交
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

  // for DataNormLayer
  // can be set to: 'z-score', 'min-max' or 'decimal-scaling'
  optional string data_norm_strategy = 36;

  // for bos/eos id
  optional uint32 bos_id = 37;
  optional uint32 eos_id = 38;

  // for max id layer
  optional uint32 beam_size = 39;

  // for seqlastins layer, whether select first instead last
  optional bool select_first = 40 [default = false];

  // for seqlastins layer, AverageLayer, MaxLayer and ExpandLayer
  // can be set to: 'non-seq','seq'
  optional string trans_type = 41 [default = 'non-seq'];

  // to indicate whether selective_fc layer
  // is used in sequence generation or not
  optional bool selective_fc_pass_generation = 42 [default = false];

  // to indicate whether selective_fc layer take its last input to
  // selected several columns and only compute the multiplications
  // between the input matrices and the selected columns of
  // the parameter matrices of this layer.
  // if set false, selective_fc degrades into fc.
  optional bool has_selected_colums = 43 [default = true];

  // this parameter is for speed consideration.
  // if number of the selected columns is less than
  // sample number * selective_fc output size * selective_fc_mull_mull_ratio
  // sparse multiplication is used, otherwise, using full multiplication.
Y
Yu Yang 已提交
449
  optional double selective_fc_full_mul_ratio = 44 [default = 0.02];
Z
zhangjinchao01 已提交
450 451 452 453 454 455 456 457 458 459 460

  // to indicate how many threads selective_fc use to to accelate
  // the plain_mul period
  // leave empty or set to 0 to disable multi-thread accleleration
  optional uint32 selective_fc_parallel_plain_mul_thread_num = 45 [default = 0];

  // for batch normalization layer
  // if set use_global_stats true, will use the loaded mean and variance.
  optional bool use_global_stats = 46;

  // use to compute moving mean and variance.
Y
Yu Yang 已提交
461
  optional double moving_average_fraction = 47 [default = 0.9];
462 463 464

  // bias size
  optional uint32 bias_size = 48 [default = 0];
H
Haichao-Zhang 已提交
465

466
  // this parameter can be used as a user-defined parameter when necessary,
H
Haichao-Zhang 已提交
467
  // without changing the proto file.
468
  // e.g., when a new layer with a user-defined parameter is implemented,
H
Haichao-Zhang 已提交
469 470
  // it can be used to pass that parameter, without modifying the proto file.
  // string type is used for flexibility: different types can be converted
471
  // to string and reinterpreted in the user's own layer implementation.
H
Haichao-Zhang 已提交
472
  optional string user_arg = 49;
473

L
Luo Tao 已提交
474 475 476
  // to indicate rectangle image data
  optional uint64 height = 50;
  optional uint64 width = 51;
H
Haichao-Zhang 已提交
477

L
Liu Yiqun 已提交
478 479
  // blank label used in ctc loss
  optional uint32 blank = 52 [default = 0];
480

481
  // stride parameter for seqlastins layer, AverageLayer, MaxLayer, which
482 483 484
  // controls the scope of pooling operation. can be set > 0.
  // leave empty or set to -1 to disable this stride pooling.
  optional int32 seq_pool_stride = 53 [default = -1];
485 486 487 488 489 490

  // for crop layer
  optional int32 axis = 54 [default = 2];
  repeated uint32 offset = 55;
  repeated uint32 shape = 56;

Z
zhangjinchao01 已提交
491 492 493 494 495 496 497 498
}

message EvaluatorConfig {
  required string name = 1;
  required string type = 2;
  repeated string input_layers = 3;

  // Used by ChunkEvaluator
499 500 501 502
  // one of "IOB", "IOE", "IOBES"
  optional string chunk_scheme = 4;
  // number of chunk types other than "other"
  optional int32 num_chunk_types = 5;
Z
zhangjinchao01 已提交
503 504 505

  // Used by PrecisionRecallEvaluator and ClassificationErrorEvaluator
  // For multi binary labels: true if output > classification_threshold
Y
Yu Yang 已提交
506
  optional double classification_threshold = 6 [default = 0.5];
Z
zhangjinchao01 已提交
507 508 509 510 511 512 513 514 515 516 517 518 519 520
  // The positive label. -1 means average precision and recall
  optional int32 positive_label = 7 [default = -1];

  // load dict from this file
  optional string dict_file = 8;

  // dump result in this file
  optional string result_file = 9;

  // top # results for max id printer
  optional int32 num_results = 10 [default = 1];

  // whether to delimit the sequence in the seq_text_printer
  optional bool delimited = 11 [default = true];
521

P
Peng Li 已提交
522 523 524
  // Used by ChunkEvaluator
  // chunk of these types are not counted
  repeated int32 excluded_chunk_types = 12;
L
Liang Zhao 已提交
525 526 527 528

  // Used by ClassificationErrorEvaluator
  // top # classification error
  optional int32 top_k = 13 [default = 1];
Y
yangyaming 已提交
529 530 531 532 533 534 535 536 537

  // Used by DetectionMAPEvaluator
  optional double overlap_threshold = 14 [default = 0.5];

  optional int32 background_id = 15 [default = 0];

  optional bool evaluate_difficult = 16 [default = false];

  optional string ap_type = 17 [default = "11point"];
Z
zhangjinchao01 已提交
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
}

message LinkConfig {
  required string layer_name = 1;
  required string link_name = 2;
  // If true, this link has sub-sequence
  optional bool has_subseq = 3 [default = false];
}

message MemoryConfig {
  required string layer_name = 1;
  required string link_name = 2;

  optional string boot_layer_name = 3;
  optional string boot_bias_parameter_name = 4;
  optional string boot_bias_active_type = 5;
  optional uint32 boot_with_const_id = 7;

  // memory is a sequence, initailized by a sequence boot layer
  optional bool is_sequence = 6 [default = false];
}

message GeneratorConfig {
  required uint32 max_num_frames = 1;
  required string eos_layer_name = 2;
  optional int32 num_results_per_sample = 3 [default = 1];

  // for beam search
  optional int32 beam_size = 4 [default = 1];

  optional bool log_prob = 5 [default = true];
}

message SubModelConfig {
  required string name = 1;
  repeated string layer_names = 2; // selected layers in sub model
  repeated string input_layer_names = 3;
  repeated string output_layer_names = 4;
  repeated string evaluator_names = 5;

  optional bool is_recurrent_layer_group = 6 [default = false];

  // If true, the recurrence runs from the end to the beginning.
  optional bool reversed = 7 [default = false];

  // name and link name of memory
  repeated MemoryConfig memories = 8;

  // if use recurrent layer group, all layers in submodel will postfix by
  // "_in_"+submodel.name, so we add a name pair to link between
  // root model and layer group,
  // note that these in/out layers are not input/output of the network.
  repeated LinkConfig in_links = 9;
  repeated LinkConfig out_links = 10;

  optional GeneratorConfig generator = 11;
594 595 596

  // the id of inlink which share info with outlinks, used in recurrent layer group
  optional int32 target_inlinkid = 12;
Z
zhangjinchao01 已提交
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
}

message ModelConfig {
  // type of the model.
  // Currently, "nn", "recurrent_nn" and "recursive_nn" are supported
  required string type = 1 [default = "nn"];

  // layers should be ordered in such a way that the forward propagation
  // can be correctly executed by going from the first layer to the last layer
  repeated LayerConfig layers = 2;

  repeated ParameterConfig parameters = 3;

  // Input layers should have the same order as the data streams provided
  // by the data provider. The type of input layers should be "data"
  repeated string input_layer_names = 4;

  // For training, the type of a output layer is usually cost layer.
  // For prediction, they should be the actual output layers.
  repeated string output_layer_names = 5;

  repeated EvaluatorConfig evaluators = 6;

  repeated SubModelConfig sub_models = 8;

  // For External Machine, defining how to split a neural network
  // into multiple parts.
  optional ExternalConfig external_config = 9;
};