PaddleAPI.h 20.5 KB
Newer Older
Z
zhangjinchao01 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve.

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. */


#pragma once

#include <stddef.h>
#include <stdint.h>
#include <string>
#include <vector>
#include "paddle/utils/GlobalConstants.h"
23
#include "paddle/utils/TypeDefs.h"
Z
zhangjinchao01 已提交
24

L
lipeng17 已提交
25
/// Import PaddlePaddle's enumeration into global namespace.
Z
zhangjinchao01 已提交
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
using namespace paddle::enumeration_wrapper;  // NOLINT

#define DISABLE_COPY_AND_ASSIGN(classname) \
  classname(const classname& other);       \
  classname& operator=(const classname& other)

/**
 * @brief Initialize paddle.
 *
 * In python, this method should be invoked as
 * @code
 *  import sys
 *  import paddle
 *  paddle.initPaddle(sys.argv)
 *  or you can change arguments as any list of str.
 * @endcode
 */
void initPaddle(int argc, char** argv);

/// Return true if this py_paddle is compiled in GPU Version
bool isGpuVersion();

/// The Error of IO Operation. Such as file not found, etc.
class IOError {};

/// Out of range error
class RangeError {};

/// Not support Error, such as access GPU memory directly, etc.
class UnsupportError {};

/// This type will map to python's list of float.
struct FloatArray {
L
liaogang 已提交
59
  const float* buf;
Z
zhangjinchao01 已提交
60 61
  const size_t length;
  bool needFree;  // true if the buf is dynamic alloced.
L
liaogang 已提交
62
  FloatArray(const float* b, const size_t l);
Z
zhangjinchao01 已提交
63 64 65 66 67 68 69 70 71 72 73 74
};

/// This type will map to python's list of int
struct IntArray {
  const int* buf;
  const size_t length;
  bool needFree;
  IntArray(const int* b, const size_t l, bool f = false);
};

/// This type will map to python's list of (int, float)
struct IntWithFloatArray {
L
liaogang 已提交
75
  const float* valBuf;
Z
zhangjinchao01 已提交
76 77 78
  const int* idxBuf;
  const size_t length;
  bool needFree;
L
liaogang 已提交
79
  IntWithFloatArray(const float* v, const int* i, size_t l, bool f = false);
Z
zhangjinchao01 已提交
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
};

enum SparseValueType { SPARSE_NON_VALUE = 0, SPARSE_VALUE = 1 };

enum SparseFormatType { SPARSE_CSR = 0, SPARSE_CSC = 1 };

/**
 * In Python, -1UL is hard to write. So define a const value used by python
 * side.
 */
const size_t NO_SPARSE_ID = -1UL;

struct MatrixPrivate;
class Matrix {
  Matrix();  // User Cannot Create Matrix.
  DISABLE_COPY_AND_ASSIGN(Matrix);
  static Matrix* createByPaddleMatrixPtr(void* sharedPtr);

public:
  virtual ~Matrix();

  /**
   * Create A Matrix with height,width, which is filled by zero.
   */
  static Matrix* createZero(size_t height, size_t width, bool useGpu = false);

  /**
   * Create Sparse Matrix.
   *
   * After create sparse, sparseCopyFrom can be used to fill matrix.
   *
   * @param nnz  Number of non zero values.
   *
   * @note the default sparse type is SPARSE_CSR.
   */
  static Matrix* createSparse(size_t height, size_t width, size_t nnz,
                              bool isNonVal = true, bool trans = false,
                              bool useGpu = false);

  /**
   * Create Dense Matrix.
   *
   * @param data  list of float should be passed in python.
   * @note        the value will be copy into a new matrix.
   */
L
liaogang 已提交
125
  static Matrix* createDense(const std::vector<float>& data, size_t height,
Z
zhangjinchao01 已提交
126 127 128 129 130 131 132 133 134 135 136
                             size_t width, bool useGpu = false);

  /**
   *  Create Cpu Dense Matrix from numpy matrix, dtype=float32
   *
   *  @param data  a numpy matrix.
   *  @param dim1  dimension of data.
   *  @param dim2  dimension of data.
   *  @param copy  true if copy into a new matrix, false will create
   *               matrix inplace.
   */
L
liaogang 已提交
137
  static Matrix* createCpuDenseFromNumpy(float* data, int dim1, int dim2,
Z
zhangjinchao01 已提交
138 139 140
                                         bool copy = false);

  /// Create Gpu Dense Matrix from numpy matrix, dtype=float32
L
liaogang 已提交
141
  static Matrix* createGpuDenseFromNumpy(float* data, int dim1, int dim2);
Z
zhangjinchao01 已提交
142 143 144 145 146 147 148 149 150 151 152 153 154 155 156

  /**
   * Cast to numpy matrix.
   *
   * @note    This method take no parameter in python.
   * @note    This method in python will return a numpy matrix, not void.
   * @note    Only CpuDenseMatrix is supported.
   *
   * Example:
   * @code
   * import paddle
   * m = paddle.Matrix.createZero(10,2)
   * numpy_mat = m.toNumpyMat()
   * @endcode
   */
L
liaogang 已提交
157
  void toNumpyMatInplace(float** view_data, int* dim1,
Z
zhangjinchao01 已提交
158 159 160
                         int* dim2) throw(UnsupportError);

  /// Copy To numpy mat.
L
liaogang 已提交
161
  void copyToNumpyMat(float** view_m_data, int* dim1,
Z
zhangjinchao01 已提交
162 163 164
                      int* dim2) throw(UnsupportError);

  /// Copy From Numpy Mat
L
liaogang 已提交
165
  void copyFromNumpyMat(float* data, int dim1, int dim2) throw(UnsupportError,
Z
zhangjinchao01 已提交
166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183
                                                               RangeError);

  /// return true if this matrix is sparse.
  bool isSparse() const;

  SparseValueType getSparseValueType() const throw(UnsupportError);

  SparseFormatType getSparseFormat() const throw(UnsupportError);

  IntArray getSparseRowCols(size_t i) const throw(UnsupportError, RangeError);

  IntWithFloatArray getSparseRowColsVal(size_t i) const
      throw(UnsupportError, RangeError);

  size_t getHeight() const;

  size_t getWidth() const;

L
liaogang 已提交
184
  float get(size_t x, size_t y) const throw(RangeError);
Z
zhangjinchao01 已提交
185

L
liaogang 已提交
186
  void set(size_t x, size_t y, float val) throw(RangeError, UnsupportError);
Z
zhangjinchao01 已提交
187 188 189 190 191 192 193 194 195 196 197

  /// return type is list of float
  FloatArray getData() const;

  /**
   * Copy from rows, cols, values.
   *
   * if sparse_nonvalue, the values should be []
   */
  void sparseCopyFrom(const std::vector<int>& rows,
                      const std::vector<int>& cols,
L
liaogang 已提交
198 199
                      const std::vector<float>& values =
                          std::vector<float>()) throw(UnsupportError);
Z
zhangjinchao01 已提交
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

  bool isGpu() const;

private:
  void* getSharedPtr() const;

  MatrixPrivate* m;
  friend class Trainer;
  friend class GradientMachine;
  friend class Arguments;
};

struct VectorPrivate;
class Vector {
  DISABLE_COPY_AND_ASSIGN(Vector);
  Vector();
  static Vector* createByPaddleVectorPtr(void* ptr);

  void* getSharedPtr();

public:
  ~Vector();

  /// Create Vector filled with zero.
  static Vector* createZero(size_t sz, bool useGpu = false);

  /**
   * Create Vector from list of float.
   *
   * It will create a new vector, and copy data into it.
   */
L
liaogang 已提交
231
  static Vector* create(const std::vector<float>& data, bool useGpu = false);
Z
zhangjinchao01 已提交
232 233 234 235 236 237

  /**
   * Create Cpu Vector from numpy array, which dtype=float32
   *
   * If copy is false, it will create vector inplace.
   */
L
liaogang 已提交
238
  static Vector* createCpuVectorFromNumpy(float* data, int dim,
Z
zhangjinchao01 已提交
239 240 241
                                          bool copy = false);

  /// Create Gpu Vector from numpy array, which dtype=float32
L
liaogang 已提交
242
  static Vector* createGpuVectorFromNumpy(float* data, int dim);
Z
zhangjinchao01 已提交
243 244

  /// Cast to numpy array inplace.
L
liaogang 已提交
245
  void toNumpyArrayInplace(float** view_data, int* dim1) throw(UnsupportError);
Z
zhangjinchao01 已提交
246 247

  /// Copy to numpy array.
L
liaogang 已提交
248
  void copyToNumpyArray(float** view_m_data, int* dim1);
Z
zhangjinchao01 已提交
249 250

  /// Copy from numpy array.
L
liaogang 已提交
251
  void copyFromNumpyArray(float* data, int dim);
Z
zhangjinchao01 已提交
252 253

  /// __getitem__ in python
L
liaogang 已提交
254
  float get(const size_t idx) const throw(RangeError, UnsupportError);
Z
zhangjinchao01 已提交
255 256

  /// __setitem__ in python
L
liaogang 已提交
257
  void set(const size_t idx, float val) throw(RangeError, UnsupportError);
Z
zhangjinchao01 已提交
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

  /// Return is GPU vector or not.
  bool isGpu() const;

  /// __len__ in python
  size_t getSize() const;

private:
  VectorPrivate* m;

private:
  friend class Parameter;
  friend class ParameterOptimizer;
  friend struct ParameterTraverseCallbackPrivate;
};

struct IVectorPrivate;
class IVector {
  IVector();
  DISABLE_COPY_AND_ASSIGN(IVector);
  static IVector* createByPaddleVectorPtr(void* ptr);

public:
  /// Create IVector filled with zero
  static IVector* createZero(size_t sz, bool useGpu = false);

  /**
   * Create IVector from list of int.
   * It will create a new vector, and copy data into it.
   */
  static IVector* create(const std::vector<int>& data, bool useGpu = false);

  /**
   * Create Cpu IVector from numpy array, which dtype=int32
   *
   * If copy is false, it will create vector inplace
   */
  static IVector* createCpuVectorFromNumpy(int* data, int dim,
                                           bool copy = false);
  /**
   * Create Gpu IVector from numpy array, which dtype=int32
   */
  static IVector* createGpuVectorFromNumy(int* data, int dim);

  /// Cast to numpy array inplace.
  void toNumpyArrayInplace(int** view_data, int* dim1) throw(UnsupportError);

  /// Copy to numpy array.
  void copyToNumpyArray(int** view_m_data, int* dim1);

  /// Copy from numpy array.
  void copyFromNumpyArray(int* data, int dim);

  virtual ~IVector();

  /// Return a list of int, the memory is alloced and copied.
  IntArray getData() const;

  /// This method will map to python [] method.
  int& operator[](const size_t idx) throw(RangeError, UnsupportError);

  const int& operator[](const size_t idx) const
      throw(RangeError, UnsupportError);

  inline int get(const size_t idx) const throw(RangeError, UnsupportError) {
    return (*this)[idx];
  }

  inline void set(const size_t idx, int val) throw(RangeError, UnsupportError) {
    (*this)[idx] = val;
  }

  /// Return true if it is gpu vector.
  bool isGpu() const;

  /// This method will map to python __len__();
  size_t getSize() const;

private:
  void* getSharedPtr() const;

  friend class Arguments;
  IVectorPrivate* m;
};

struct ArgumentsPrivate;

/// The Arguments is actual a std::vector<paddle::Argument> in paddle.
class Arguments {
private:
  Arguments();  // Internal Create.
  DISABLE_COPY_AND_ASSIGN(Arguments);

public:
  /**
   * Create a arguments with size.
   * Note that it can be zero.
   */
  static Arguments* createArguments(size_t slotNum);

  void resize(size_t slotNum);

  virtual ~Arguments();

  /**
   * Return the slot number that aguments contains.
   *
   * It is actually the vector's size
   */
  size_t getSlotNum() const;

  /**
   * The get functions of Arguments
   *
   * the param idx is the slot id
   */
  Matrix* getSlotValue(size_t idx) const throw(RangeError);
  IVector* getSlotIds(size_t idx) const throw(RangeError);
  Matrix* getSlotIn(size_t idx) const throw(RangeError);
  IVector* getSlotSequenceStartPositions(size_t idx) const throw(RangeError);
378
  IVector* getSlotSubSequenceStartPositions(size_t idx) const throw(RangeError);
Z
zhangjinchao01 已提交
379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394
  IVector* getSlotSequenceDim(size_t idx) const throw(RangeError);
  // End Of get functions of Arguments

  int64_t getBatchSize(size_t idx = 0) const throw(RangeError);

  /**
   * The set functions of Arguments.
   *
   * The param idx is the slot id.
   * The other param is the input Matrix or vector.
   */
  void setSlotValue(size_t idx, Matrix* mat) throw(RangeError);
  void setSlotIn(size_t idx, Matrix* mat) throw(RangeError);
  void setSlotIds(size_t idx, IVector* vec) throw(RangeError);
  void setSlotSequenceStartPositions(size_t idx,
                                     IVector* vec) throw(RangeError);
395
  void setSlotSubSequenceStartPositions(size_t idx,
Y
yuyang18 已提交
396
                                        IVector* vec) throw(RangeError);
Z
zhangjinchao01 已提交
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
  void setSlotSequenceDim(size_t idx, IVector* vec) throw(RangeError);

private:
  static Arguments* createByPaddleArgumentVector(void* ptr);
  void* getInternalArgumentsPtr() const;

private:
  ArgumentsPrivate* m;
  friend class Trainer;
  friend class GradientMachine;
  friend class SequenceGenerator;
};

enum GradientMatchineCreateMode {
  CREATE_MODE_NORMAL = 0,
  CREATE_MODE_TESTING = 4
};

struct ParameterConfigPrivate;
class ParameterConfig {
  DISABLE_COPY_AND_ASSIGN(ParameterConfig);
  ParameterConfig();

  /**
   * Internal methods
   */
  static ParameterConfig* createParameterConfigFromParameterSharedPtr(
      void* ptr);
  static ParameterConfig* createParameterConfigFromParameterPtr(void* ptr);
  void* getRawPtr();

public:
  ~ParameterConfig();

  /**
   * return proto buf string.
   */
  std::string toProtoString() const;

private:
  ParameterConfigPrivate* m;

private:
  friend class Parameter;
  friend class ParameterOptimizer;
  friend struct ParameterTraverseCallbackPrivate;
};

struct OptimizationConfigPrivate;
class OptimizationConfig {
  DISABLE_COPY_AND_ASSIGN(OptimizationConfig);
  OptimizationConfig();

public:
  static OptimizationConfig* createFromProtoString(const std::string& str);
  ~OptimizationConfig();

  /**
   * return protobuf string.
   */
  std::string toProtoString();

private:
  OptimizationConfigPrivate* m;

  friend class TrainerConfig;
  friend class ParameterOptimizer;
E
emailweixu 已提交
464
  friend class Trainer;
Z
zhangjinchao01 已提交
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
};

struct ParameterPrivate;
class Parameter {
private:
  Parameter();
  DISABLE_COPY_AND_ASSIGN(Parameter);

public:
  virtual ~Parameter();

  /**
   * get parameter name
   */
  std::string getName() const;

  /**
   * get buf in Parameter
   */
  Vector* getBuf(ParameterType type);

  /**
   * get id
   */
  size_t getID() const;

  ParameterConfig* getConfig();

private:
  static Parameter* createFromRawPtr(void* ptr);
  static Parameter* createFromSharedPtr(void* ptr);

private:
  ParameterPrivate* m;
  friend class UpdateCallbackWrapper;
  friend class GradientMachine;
};

struct ModelConfigPrivate;
/**
 * You can only get model config from TrainerConfig.
 *
 * It is used by GradientMachine.
 */
class ModelConfig {
private:
  ModelConfig();
  DISABLE_COPY_AND_ASSIGN(ModelConfig);

public:
  virtual ~ModelConfig();

private:
  ModelConfigPrivate* m;
  friend class TrainerConfig;
  friend struct TrainerConfigPrivate;
  friend class GradientMachine;
};

struct TrainerConfigPrivate;
/**
 * To get TrainerConfig from file.
 *
 * It is used by GradientMachine.
 */
class TrainerConfig {
private:
  TrainerConfig();
  DISABLE_COPY_AND_ASSIGN(TrainerConfig);

public:
  virtual ~TrainerConfig();

  static TrainerConfig* createFromTrainerConfigFile(
      const std::string& configPath);
E
emailweixu 已提交
540
  static TrainerConfig* createFromProtoString(const std::string& str);
Z
zhangjinchao01 已提交
541 542 543 544 545 546 547

  ModelConfig* getModelConfig() const;

  OptimizationConfig* getOptimizationConfig() const;

private:
  TrainerConfigPrivate* m;
E
emailweixu 已提交
548
  friend class Trainer;
Z
zhangjinchao01 已提交
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
};

/**
 * The callback in backword.
 *
 * You can inherit this class in python.
 *
 * @code
 * class UpdateCallbackInPython(paddle.UpdateCallback):
 *   def __init__(self):
 *     paddle.UpdateCallback.__init__(self)
 *
 *   def apply(self, param):
 *     assert isinstance(param, paddle.Parameter)
 * @endcode
 */
class UpdateCallback {
public:
  virtual ~UpdateCallback();
  virtual void apply(Parameter* p);
};

struct ParameterTraverseCallbackPrivate;
class ParameterTraverseCallback {
  DISABLE_COPY_AND_ASSIGN(ParameterTraverseCallback);
  ParameterTraverseCallback();

public:
  ~ParameterTraverseCallback();

  void apply(const std::vector<Vector*>& vecs, const ParameterConfig& config,
             size_t sparseId);

private:
  ParameterTraverseCallbackPrivate* m;
  friend class ParameterOptimizer;
};

/**
 * The ParameterOptimizer Wrapper Class.
 *
 * Basically same as common/ParameterOptimizer.h
 */
struct ParameterOptimizerPrivate;
class ParameterOptimizer {
  DISABLE_COPY_AND_ASSIGN(ParameterOptimizer);
  ParameterOptimizer();

public:
  static ParameterOptimizer* create(OptimizationConfig* config);

  ~ParameterOptimizer();

  void init(size_t numRows, const ParameterConfig* config);

  void startPass();

  void finishPass();

  void startBatch(size_t numSamplesProcessed);

  void finishBatch();

  void update(const std::vector<Vector*>& vecs, const ParameterConfig& conf,
              size_t sparseId = NO_SPARSE_ID);

  std::vector<int> getParameterTypes() const;

  ParameterTraverseCallback* needSpecialTraversal(
      const ParameterConfig& config) const;

private:
  ParameterOptimizerPrivate* m;
};

class SequenceGenerator;

struct GradientMachinePrivate;
class GradientMachine {
private:
  GradientMachine();
  DISABLE_COPY_AND_ASSIGN(GradientMachine);

public:
  virtual ~GradientMachine();

  /**
   * Create By ProtoStr.
   *
   * The ProtoStr can be generate by python's protobuf code.
   */
  static GradientMachine* createByConfigProtoStr(
      const std::string& protoStr,
      GradientMatchineCreateMode mode = CREATE_MODE_NORMAL,
      const std::vector<int>& parameterTypes = defaultParamTypes);

  /**
   * Create by ModelConfig object.
   *
   * To get ModelConfig, you can get TrainerConfig from config file, then get
   * model config by TrainerConfig
   */
  static GradientMachine* createByModelConfig(
      ModelConfig* conf, GradientMatchineCreateMode mode = CREATE_MODE_NORMAL,
      const std::vector<int>& parameterTypes = defaultParamTypes);

  /**
   * The forward stage of GradientMachine.
   *
   * @note  the outArgs could be zero length arguemnts.
   * @note  THIS METHOD IS VERY USEFULL FOR PREDICT FROM TRAINED MODEL.
   */
  void forward(const Arguments& inArgs, Arguments* outArgs, PassType passType);

  /**
   * The backward stage of GradientMachine.
   *
   * @note  Currently the ParameterUpdater is not wrapped in SWIG, so backward
   * cannot actually train a network. But you can write a update callback to
   * change the parameter or implement a ParameterUpdater in python side.
   */
  void backward(const UpdateCallback& callback = UpdateCallback());

  /**
   * Combine forward/backward
   */
  void forwardBackward(const Arguments& inArgs, Arguments* outArgs,
                       PassType passType,
                       const UpdateCallback& callback = UpdateCallback());

  void loadParameters(const std::string& path);

  size_t getParameterSize() const;
  Parameter* getParameter(size_t i) throw(RangeError);

  void randParameters();

  Matrix* getLayerOutput(const std::string& layerName) const
      throw(UnsupportError);

  /**
   * Create a sequence generator.
   *
   * @note  It just like a paddle_gen_sequence.
   */
  SequenceGenerator* asSequenceGenerator(
      const std::vector<std::string>& dict = std::vector<std::string>(),
      size_t begin_id = 0UL, size_t end_id = 0UL, size_t max_length = 100UL,
      size_t beam_size = -1UL);

private:
  GradientMachinePrivate* m;

  static GradientMachine* createFromPaddleModelPtr(
E
emailweixu 已提交
703
      const void* confPtr, GradientMatchineCreateMode mode,
Z
zhangjinchao01 已提交
704 705 706 707
      const std::vector<int>& types);

  // Not to use c++ 11 init-list, so we use static var as function default arg.
  static std::vector<int> defaultParamTypes;
E
emailweixu 已提交
708
  friend class Trainer;
Z
zhangjinchao01 已提交
709 710 711 712 713 714 715
};

struct TrainerPrivate;
class Trainer {
private:
  TrainerPrivate* m;
  Trainer();
E
emailweixu 已提交
716
  Trainer(TrainerConfig* optConfig, GradientMachine* gm);
Z
zhangjinchao01 已提交
717 718 719 720 721 722 723 724
  DISABLE_COPY_AND_ASSIGN(Trainer);

public:
  virtual ~Trainer();

  /// Create A Trainer By TrainerConfig. using paddle command line.
  static Trainer* createByCommandLine() throw(IOError);

E
emailweixu 已提交
725 726 727 728
  static Trainer* create(TrainerConfig* optConfig, GradientMachine* gm)
      throw(IOError);

  /// Start training
Z
zhangjinchao01 已提交
729
  void startTrain();
E
emailweixu 已提交
730 731

  /// Finish training
Z
zhangjinchao01 已提交
732 733
  void finishTrain();

E
emailweixu 已提交
734
  /// Start a pass.
Z
zhangjinchao01 已提交
735 736
  void startTrainPass();

E
emailweixu 已提交
737 738
  /// Finish a pass
  void finishTrainPass();
Z
zhangjinchao01 已提交
739 740 741 742 743 744

  /**
   * Train one batch,
   *
   * @return true if all batch finished.
   */
E
emailweixu 已提交
745
  bool trainOneBatch(size_t batchSize);
Z
zhangjinchao01 已提交
746

E
emailweixu 已提交
747
  void trainOneDataBatch(size_t batchSize, const Arguments& args);
Z
zhangjinchao01 已提交
748

E
emailweixu 已提交
749 750 751
  void startTestPeriod();
  void testOneDataBatch(size_t batchSize, const Arguments& args);
  void finishTestPeriod();
Z
zhangjinchao01 已提交
752

E
emailweixu 已提交
753
  void forwardOneBatch(size_t batchSize);
Z
zhangjinchao01 已提交
754

E
emailweixu 已提交
755
  Arguments* getForwardOutput();
Z
zhangjinchao01 已提交
756 757 758 759

  Matrix* getLayerOutput(const std::string& layerName);
};

E
emailweixu 已提交
760
/// the N-Best results generated from one input sequence.
Z
zhangjinchao01 已提交
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
class ISequenceResults {
public:
  virtual ~ISequenceResults();

  /// Number of result.
  virtual size_t getSize() const = 0;

  /**
   * Get sentence from dictionary.
   *
   * @param id  the index of result.
   * @param split  if true, the return sentence will be splited with ' ' by
   *               each word. Default is false.
   */
  virtual std::string getSentence(size_t id, bool split = false) const
      throw(RangeError) = 0;
  virtual std::vector<int> getSequence(size_t id) const throw(RangeError) = 0;
  virtual float getScore(size_t id) const throw(RangeError) = 0;
};

struct SequenceGeneratorPrivate;
class SequenceGenerator {
  DISABLE_COPY_AND_ASSIGN(SequenceGenerator);
  SequenceGenerator();

public:
  virtual ~SequenceGenerator();

  /**
   * Generate Sequence by input.
   *
   * @note  The inArgs is just one sequence of data.
   * @note  The return will get a N-best generate result by inArgs.
   *        Sort by score.
   */
  ISequenceResults* generateSequence(const Arguments& inArgs) const;

  void setDict(const std::vector<std::string>& dict);
  void setBos(size_t bos);
  void setEos(size_t eos);
  void setMaxLength(size_t maxlength);
  void setBeamSize(size_t beamSize);

private:
  static SequenceGenerator* createByGradientMachineSharedPtr(void* ptr);
  friend class GradientMachine;

private:
  SequenceGeneratorPrivate* m;
};