nnet-component-itf.h
35.4 KB
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
// nnet3/nnet-component-itf.h
// Copyright 2015 Johns Hopkins University (author: Daniel Povey)
// 2015 Guoguo Chen
// 2015 Xiaohui Zhang
// See ../../COPYING for clarification regarding multiple authors
//
// 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
//
// THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
// WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
// MERCHANTABLITY OR NON-INFRINGEMENT.
// See the Apache 2 License for the specific language governing permissions and
// limitations under the License.
#ifndef KALDI_NNET3_NNET_COMPONENT_ITF_H_
#define KALDI_NNET3_NNET_COMPONENT_ITF_H_
#include <iostream>
#include "nnet3/nnet-common.h"
#include "nnet3/nnet-parse.h"
#include "base/kaldi-error.h"
namespace kaldi {
namespace nnet3 {
// enum used to store various binary component properties.
// We give it a name ComponentProperties, but don't use this
// type for the bitmasks: instead use int32 for this type, e.g.
// int32 properties = kSimpleComponent|kBackpropNeedsOutput.
enum ComponentProperties {
kSimpleComponent = 0x001, // true if number of rows of input equals number of rows
// of output and this component doesn't care about the indexes
// (i.e. it maps each row of input to each row of output without
// regard to the index values). Will normally be true.
kUpdatableComponent = 0x002, // true if the component has parameters that can
// be updated. Components that return this flag
// must be dynamic_castable to type
// UpdatableComponent (but components of type
// UpdatableComponent do not have to return this
// flag, e.g. if this instance is not really
// updatable).
kPropagateInPlace = 0x004, // true if we can do the propagate operation in-place
// (input and output matrices are the same).
// Note: if doing backprop, you'd also need to check
// that the kBackpropNeedsInput property is not true.
kPropagateAdds = 0x008, // true if the Propagate function adds to, rather
// than setting, its output, for non-in-place
// propagation. The Component chooses whether to add
// or set, and the calling code has to accommodate
// it.
kReordersIndexes = 0x010, // true if the ReorderIndexes function might reorder
// the indexes (otherwise we can skip calling it).
// Must not be set for simple components.
kBackpropAdds = 0x020, // true if the Backprop function adds to, rather than
// setting, the "in_deriv" output for non-in-place
// backprop. The Component chooses whether to add or
// set, and the calling code has to accommodate it.
kBackpropNeedsInput = 0x040, // true if backprop operation needs access to
// forward-pass input.
kBackpropNeedsOutput = 0x080, // true if backprop operation needs access to
// forward-pass output (e.g. true for Sigmoid).
kBackpropInPlace = 0x100, // true if we can do the backprop operation in-place
// (input and output matrices may be the same).
kStoresStats = 0x200, // true if the StoreStats operation stores
// statistics e.g. on average node activations and
// derivatives of the nonlinearity, (as it does for
// Tanh, Sigmoid, ReLU and Softmax).
kInputContiguous = 0x400, // true if the component requires its input data (and
// input derivatives) to have Stride()== NumCols().
kOutputContiguous = 0x800, // true if the component requires its input data (and
// output derivatives) to have Stride()== NumCols().
kUsesMemo = 0x1000, // true if the component returns a void* pointer from its
// Propagate() function that needs to be passed into the
// corresponding Backprop function.
kRandomComponent = 0x2000 // true if the component has some kind of
// randomness, like DropoutComponent (these should
// inherit from class RandomComponent.
};
// This is a base class for a helper-class of class Component, which is used to
// store any pre-computed indexes it needs for its forward and backward
// computations. For components which are not "Simple" components (i.e. the
// kSimpleComponent property is false), and which may therefore "care" about
// which index the input and output matrix's rows represent (i.e. about
// which "struct Index" each row corresponds to), their CreateIndexes() function
// will be called prior to Propagate() and Backprop(), to create an object which
// must be a child class of class ComponentPrecomputedIndexes, where they
// can store any indexes that they need.
class ComponentPrecomputedIndexes {
public:
virtual ComponentPrecomputedIndexes *Copy() const = 0;
virtual void Write(std::ostream &os, bool binary) const = 0;
virtual void Read(std::istream &os, bool binary) = 0;
virtual std::string Type() const = 0;
static ComponentPrecomputedIndexes* ReadNew(std::istream &is, bool binary);
// cpi stands for component_precomputed_indexes
static ComponentPrecomputedIndexes* NewComponentPrecomputedIndexesOfType(
const std::string &cpi_type);
virtual ~ComponentPrecomputedIndexes() { }
};
class IndexSet; // Forward declaration; declared in nnet-computation-graph.h.
/// Abstract base-class for neural-net components.
class Component {
public:
/// \brief Propagate function.
/// \param [in] indexes A pointer to some information output by this class's
/// PrecomputeIndexes function (will be NULL for simple components,
/// i.e. those that don't do things like splicing).
/// \param [in] in The input to this component. Num-columns == InputDim().
/// \param [out] out The output of this component. Num-columns == OutputDim().
/// Note: output of this component will be added to the initial value of
/// "out" if Properties()&kPropagateAdds != 0; otherwise the output will
/// be set and the initial value ignored. Each Component chooses whether
/// it is more convenient implementation-wise to add or set, and the
/// calling code has to deal with it.
/// \return Normally returns NULL, but may return a non-NULL value for
/// components which have the flag kUsesMemo set. This value will
/// be passed into the corresponding Backprop routine.
virtual void* Propagate(const ComponentPrecomputedIndexes *indexes,
const CuMatrixBase<BaseFloat> &in,
CuMatrixBase<BaseFloat> *out) const = 0;
/// \brief Backprop function; depending on which of the arguments 'to_update'
/// and 'in_deriv' are non-NULL, this can compute input-data derivatives
/// and/or perform model update.
///
/// \param [in] debug_info The component name, to be printed out in any
/// warning messages.
/// \param [in] indexes A pointer to some information output by this
/// class's PrecomputeIndexes function (will be NULL for simple
/// components, i.e. those that don't do things like splicing).
/// \param [in] in_value The matrix that was given as input to the
/// Propagate function. Will be ignored (and may be empty) if
/// Properties()&kBackpropNeedsInput == 0.
/// \param [in] out_value The matrix that was output from the Propagate
/// function. Will be ignored (and may be empty) if
/// Properties()&kBackpropNeedsOutput == 0
/// \param [in] out_deriv The derivative at the output of this component.
/// \param [in] memo This will normally be NULL, but for component
/// types that set the flag kUsesMemo, this will be the return value
/// of the Propagate() function that corresponds to this Backprop()
/// function. Ownership of any pointers is not transferred to the
/// Backprop function; DeleteMemo() will be called to delete it.
/// \param [out] to_update If model update is desired, the Component
/// to be updated, else NULL. Does not have to be identical to this.
/// If supplied, you can assume that
/// to_update->Properties() & kUpdatableComponent is nonzero.
/// \param [out] in_deriv The derivative at the input of this component,
/// if needed (else NULL). If Properties()&kBackpropInPlace, may be
/// the same matrix as out_deriv. If Properties()&kBackpropAdds, this
/// is added to by the Backprop routine, else it is set. The component
/// code chooses which mode to work in, based on convenience.
virtual void Backprop(const std::string &debug_info,
const ComponentPrecomputedIndexes *indexes,
const CuMatrixBase<BaseFloat> &in_value,
const CuMatrixBase<BaseFloat> &out_value,
const CuMatrixBase<BaseFloat> &out_deriv,
void *memo,
Component *to_update, // may be NULL; may be identical
// to "this" or different.
CuMatrixBase<BaseFloat> *in_deriv) const = 0;
/// \brief This function may store stats on average activation values, and for
/// some component types, the average value of the derivative of the
/// nonlinearity. It only does something for those components that
/// have nonzero Properties()&kStoresStats.
///
/// \param [in] in_value The input to the Propagate() function. Note: if
/// the component sets the flag kPropagateInPlace, this should not
/// be used; the empty matrix will be provided here if in-place
/// propagation was used.
/// \param [in] out_value The output of the Propagate() function.
/// \param [in] memo The 'memo' returned by the Propagate() function; this
/// will usually be NULL.
virtual void StoreStats(const CuMatrixBase<BaseFloat> &in_value,
const CuMatrixBase<BaseFloat> &out_value,
void *memo) { }
/// \brief Components that provide an implementation of StoreStats should also
/// provide an implementation of ZeroStats(), to set those stats to
/// zero. Other components that store other types of statistics
/// (e.g. regarding gradient clipping) should implement ZeroStats()
/// also.
virtual void ZeroStats() { }
/// \brief This function only does something interesting for non-simple Components.
/// For a given index at the output of the component, tells us what indexes
/// are required at its input (note: "required" encompasses also optionally-required
/// things; it will enumerate all things that we'd like to have). See also
/// IsComputable().
/// \param [in] misc_info This argument is supplied to handle things that the
/// framework can't very easily supply: information like which time
/// indexes are needed for AggregateComponent, which time-indexes are
/// available at the input of a recurrent network, and so on. We will
/// add members to misc_info as needed.
/// \param [in] output_index The Index at the output of the component, for
/// which we are requesting the list of indexes at the component's input.
/// \param [out] desired_indexes A list of indexes that are desired at the
/// input. are to be written to here. By "desired" we mean required or
/// optionally-required.
///
/// The default implementation of this function is suitable for any
/// SimpleComponent; it just copies the output_index to a single identical
/// element in input_indexes.
virtual void GetInputIndexes(const MiscComputationInfo &misc_info,
const Index &output_index,
std::vector<Index> *desired_indexes) const;
/// \brief This function only does something interesting for non-simple
/// Components, and it exists to make it possible to manage
/// optionally-required inputs. It tells the user whether a given output
/// index is computable from a given set of input indexes, and if so,
/// says which input indexes will be used in the computation.
///
/// Implementations of this function are required to have the property that
/// adding an element to "input_index_set" can only ever change IsComputable
/// from false to true, never vice versa.
///
/// @param [in] misc_info Some information specific to the computation, such as
/// minimum and maximum times for certain components to do adaptation on;
/// it's a place to put things that don't easily fit in the framework.
/// @param [in] output_index The index that is to be computed at the output
/// of this Component.
/// @param [in] input_index_set The set of indexes that is available at the
/// input of this Component.
/// @param [out] used_inputs If this is non-NULL and the output is
/// computable this will be set to the list of input indexes that will
/// actually be used in the computation.
/// @return Returns true iff this output is computable from the provided
/// inputs.
///
/// The default implementation of this function is suitable for any
/// SimpleComponent: it just returns true if output_index is in
/// input_index_set, and if so sets used_inputs to vector containing that
/// one Index.
virtual bool IsComputable(const MiscComputationInfo &misc_info,
const Index &output_index,
const IndexSet &input_index_set,
std::vector<Index> *used_inputs) const;
/// \brief This function only does something interesting for non-simple
/// Components. It provides an opportunity for a Component to reorder the or
/// pad the indexes at its input and output. This might be useful, for
/// instance, if a component requires a particular ordering of the indexes
/// that doesn't correspond to their natural ordering. Components that might
/// modify the indexes are required to return the kReordersIndexes flag in
/// their Properties().
/// The ReorderIndexes() function is now allowed to insert blanks
/// into the indexes. The 'blanks' must be of the form (n,kNoTime,x),
/// where the marker kNoTime (a very negative number) is there where
/// the 't' indexes normally live. The reason we don't just have, say,
/// (-1,-1,-1), relates to the need to preserve a regular pattern over
/// the 'n' indexes so that 'shortcut compilation' (c.f. ExpandComputation())
/// can work correctly
///
///
/// \param [in,out] Indexes at the input of the Component.
/// \param [in,out] Indexes at the output of the Component
virtual void ReorderIndexes(std::vector<Index> *input_indexes,
std::vector<Index> *output_indexes) const {}
/// \brief This function must return NULL for simple Components. Returns a
/// pointer to a class that may contain some precomputed
/// component-specific and computation-specific indexes to be in used in
/// the Propagate and Backprop functions.
///
/// \param [in] misc_info This argument is supplied to handle things that the
/// framework can't very easily supply: information like which time
/// indexes are needed for AggregateComponent, which time-indexes are
/// available at the input of a recurrent network, and so on. misc_info
/// may not even ever be used here. We will add members to misc_info as
/// needed.
/// \param [in] input_indexes A vector of indexes that explains
/// what time-indexes (and other indexes) each row of the
/// in/in_value/in_deriv matrices given to Propagate and Backprop will
/// mean.
/// \param [in] output_indexes A vector of indexes that explains
/// what time-indexes (and other indexes) each row of the
/// out/out_value/out_deriv matrices given to Propagate and Backprop will
/// mean.
/// \param [in] need_backprop True if we might need to do backprop
/// with this component, so that if any different indexes are needed
/// for backprop then those should be computed too.
/// \return Returns a child-class of class ComponentPrecomputedIndexes, or
/// NULL if this component for does not need to precompute any indexes
/// (e.g. if it is a simple component and does not care about indexes).
virtual ComponentPrecomputedIndexes* PrecomputeIndexes(
const MiscComputationInfo &misc_info,
const std::vector<Index> &input_indexes,
const std::vector<Index> &output_indexes,
bool need_backprop) const { return NULL; }
/// \brief Returns a string such as "SigmoidComponent", describing
/// the type of the object.
virtual std::string Type() const = 0;
/// \brief Initialize, from a ConfigLine object.
/// \param [in] cfl A ConfigLine containing any parameters that
/// are needed for initialization. For example:
/// "dim=100 param-stddev=0.1"
virtual void InitFromConfig(ConfigLine *cfl) = 0;
/// \brief Returns input-dimension of this component.
virtual int32 InputDim() const = 0;
/// \brief Returns output-dimension of this component.
virtual int32 OutputDim() const = 0;
/// \brief Return bitmask of the component's properties.
/// These properties depend only on the component's type.
/// See enum ComponentProperties.
virtual int32 Properties() const = 0;
/// \brief Read component from stream (works out its type). Dies on error.
static Component* ReadNew(std::istream &is, bool binary);
/// \brief Copies component (deep copy).
virtual Component* Copy() const = 0;
/// \brief Returns a new Component of the given type e.g. "SoftmaxComponent",
/// or NULL if no such component type exists.
static Component *NewComponentOfType(const std::string &type);
/// \brief Read function (used after we know the type of the Component);
/// accepts input that is missing the token that describes the component
/// type, in case it has already been consumed.
virtual void Read(std::istream &is, bool binary) = 0;
/// \brief Write component to stream
virtual void Write(std::ostream &os, bool binary) const = 0;
/// \brief Returns some text-form information about this component, for diagnostics.
/// Starts with the type of the component. E.g. "SigmoidComponent dim=900",
/// although most components will have much more info.
virtual std::string Info() const;
/// This virtual function when called on
/// -- an UpdatableComponent scales the parameters
/// by "scale" when called by an UpdatableComponent.
/// -- a Nonlinear component (or another component that
/// stores stats, like BatchNormComponent)-- it relates
/// to scaling activation stats, not parameters.
/// Otherwise it will normally do nothing.
virtual void Scale(BaseFloat scale) {};
/// This virtual function when called by
/// -- an UpdatableComponent adds the parameters of
/// another updatable component, times some constant, to the current
/// parameters.
/// -- a NonlinearComponent (or another component that stores
/// stats, like BatchNormComponent)-- it relates to adding
/// stats.
/// Otherwise it will normally do nothing.
virtual void Add(BaseFloat alpha, const Component &other) {};
/// This virtual function only needs to be overwritten by Components that
/// return a non-NULL memo from their Propagate() function. It's called by
/// NnetComputer in cases where Propagate returns a memo but there will be no
/// backprop to consume it.
virtual void DeleteMemo(void *memo) const { KALDI_ASSERT(memo == NULL); }
/// This virtual function relates to memory management, and avoiding
/// fragmentation. It is called only once per model, after we do the first
/// minibatch of training. The default implementation does nothing, but it
/// can be overridden by child classes, where it may re-initialize certain
/// quantities that may possibly have been allocated during the forward pass
/// (e.g. certain statistics; OnlineNaturalGradient objects). We use our own
/// CPU-based allocator (see cu-allocator.h) and since it can't do paging
/// since we're not in control of the GPU page table, fragmentation can be a
/// problem. The allocator always tries to put things in 'low-address memory'
/// (i.e. at smaller memory addresses) near the beginning of the block it
/// allocated, to avoid fragmentation; but if permanent things (belonging to
/// the model) are allocated in the forward pass, they can permanently stay in
/// high memory. This function helps to prevent that, by re-allocating those
/// things into low-address memory (It's important that it's called after all the
/// temporary buffers for the forward-backward have been freed, so that there
/// is low-address memory available)).
virtual void ConsolidateMemory() { }
Component() { }
virtual ~Component() { }
private:
KALDI_DISALLOW_COPY_AND_ASSIGN(Component);
};
class RandomComponent: public Component {
public:
// This function is required in testing code and in other places we need
// consistency in the random number generation (e.g. when optimizing
// validation-set performance), but check where else we call srand(). You'll
// need to call srand prior to making this call.
void ResetGenerator() { random_generator_.SeedGpu(); }
// Call this with 'true' to set 'test mode' where the behavior is different
// from normal mode.
void SetTestMode(bool test_mode) { test_mode_ = test_mode; }
RandomComponent(): test_mode_(false) { }
RandomComponent(const RandomComponent &other):
test_mode_(other.test_mode_) {}
protected:
CuRand<BaseFloat> random_generator_;
// This is true if we want a different behavior for inference from that for
// training.
bool test_mode_;
};
/**
Class UpdatableComponent is a Component which has trainable parameters; it
extends the interface of Component. This is a base-class for Components with
parameters. See comment by declaration of kUpdatableComponent.
The functions in this interface must only be called if the component returns
the kUpdatable flag.
Child classes support the following config-line parameters in addition
to more specific ones:
learning-rate e.g. learning-rate=1.0e-05. default=0.001
It's not normally necessary or desirable to set this
in the config line, as it typically gets set
in the training scripts.
learning-rate-factor e.g. learning-rate-factor=0.5, can be used to
conveniently control per-layer learning rates (it's
multiplied by the learning rates given to the
--learning-rate option to nnet3-copy or any
'set-learning-rate' directives to the
--edits-config option of nnet3-copy. default=1.0.
max-change e.g. max-change=0.75. Maximum allowed parameter change
for the parameters of this component, in Euclidean norm,
per update step. If zero, no limit is applied at this
level (the global --max-param-change option will still
apply). default=0.0.
*/
class UpdatableComponent: public Component {
public:
UpdatableComponent(const UpdatableComponent &other);
// If these defaults are changed, the defaults in
// InitLearningRatesFromConfig() should be changed too.
UpdatableComponent(): learning_rate_(0.001), learning_rate_factor_(1.0),
l2_regularize_(0.0), is_gradient_(false),
max_change_(0.0) { }
virtual ~UpdatableComponent() { }
/// \brief Computes dot-product between parameters of two instances of a
/// Component. Can be used for computing parameter-norm of an
/// UpdatableComponent.
virtual BaseFloat DotProduct(const UpdatableComponent &other) const = 0;
/// This function is to be used in testing. It adds unit noise times "stddev"
/// to the parameters of the component.
virtual void PerturbParams(BaseFloat stddev) = 0;
/// Sets the learning rate of gradient descent- gets multiplied by
/// learning_rate_factor_.
virtual void SetUnderlyingLearningRate(BaseFloat lrate) {
learning_rate_ = lrate * learning_rate_factor_;
}
/// Sets the learning rate directly, bypassing learning_rate_factor_.
virtual void SetActualLearningRate(BaseFloat lrate) { learning_rate_ = lrate; }
/// \brief Sets is_gradient_ to true and sets learning_rate_ to 1, ignoring
/// learning_rate_factor_.
virtual void SetAsGradient() { learning_rate_ = 1.0; is_gradient_ = true; }
virtual BaseFloat LearningRateFactor() { return learning_rate_factor_; }
// Sets the learning rate factors to lrate_factor.
virtual void SetLearningRateFactor(BaseFloat lrate_factor) {
learning_rate_factor_ = lrate_factor;
}
// Copies the learning-rate, learning-rate-factor, l2-regularize, is-gradient
// and max-change values from 'other'.
void SetUpdatableConfigs(const UpdatableComponent &other);
/// freezes/unfreezes NaturalGradient updates, if applicable (to be overriden
/// by components that use Natural Gradient).
virtual void FreezeNaturalGradient(bool freeze) { }
/// Gets the learning rate to be used in gradient descent.
BaseFloat LearningRate() const { return learning_rate_; }
/// Returns the per-component max-change value, which is interpreted as the
/// maximum change (in l2 norm) in parameters that is allowed per minibatch
/// for this component. The components themselves do not enforce the
/// per-component max-change; it's enforced in class NnetTrainer by querying
/// the max-changes for each component. See
/// NnetTrainer::UpdateParamsWithMaxChange() in nnet-utils.h.
BaseFloat MaxChange() const { return max_change_; }
void SetMaxChange(BaseFloat max_change) { max_change_ = max_change; }
/// Returns the l2 regularization constant, which may be set in any updatable
/// component (usually from the config file). This value is not interrogated
/// in the component-level code. Instead it is read by the function
/// ApplyL2Regularization(), declared in nnet-utils.h, which is used as part
/// of the training workflow.
BaseFloat L2Regularization() const { return l2_regularize_; }
void SetL2Regularization(BaseFloat a) { l2_regularize_ = a; }
virtual std::string Info() const;
/// The following new virtual function returns the total dimension of
/// the parameters in this class.
virtual int32 NumParameters() const { KALDI_ASSERT(0); return 0; }
/// Turns the parameters into vector form. We put the vector form on the CPU,
/// because in the kinds of situations where we do this, we'll tend to use
/// too much memory for the GPU.
virtual void Vectorize(VectorBase<BaseFloat> *params) const { KALDI_ASSERT(0); }
/// Converts the parameters from vector form.
virtual void UnVectorize(const VectorBase<BaseFloat> ¶ms) {
KALDI_ASSERT(0);
}
protected:
// to be called from child classes, extracts any learning rate information
// from the config line and sets them appropriately.
void InitLearningRatesFromConfig(ConfigLine *cfl);
// To be used in child-class Read() functions, this function reads the opening
// tag <ThisComponentType> and the learning-rate factor and the learning-rate.
//
// Its return value may not always be needed to be inspected by calling code;
// if there was a token that it read but could not process it returns it, else
// it returns "".
std::string ReadUpdatableCommon(std::istream &is, bool binary);
// To be used in child-class Write() functions, writes the opening
// <ThisComponentType> tag and the learning-rate factor (if not 1.0) and the
// learning rate;
void WriteUpdatableCommon(std::ostream &is, bool binary) const;
BaseFloat learning_rate_; ///< learning rate (typically 0.0..0.01)
BaseFloat learning_rate_factor_; ///< learning rate factor (normally 1.0, but
///< can be set to another < value so that
///when < you call SetLearningRate(), that
///value will be scaled by this factor.
BaseFloat l2_regularize_; ///< L2 regularization constant. See comment for
///< the L2Regularization() for details.
bool is_gradient_; ///< True if this component is to be treated as a gradient rather
///< than as parameters. Its main effect is that we disable
///< any natural-gradient update and just compute the standard
///< gradient.
BaseFloat max_change_; ///< configuration value for imposing max-change
private:
const UpdatableComponent &operator = (const UpdatableComponent &other); // Disallow.
};
/* NonlinearComponent is a base-class for things like sigmoid, softmax and
ReLU: nonlinearities that don't change the dimension. This base-class
takes care of storing statistics on the average activations and derivatives
encountered during training, and model initialization and I/O.
Supported parameters on the config line:
dim Dimension of the input and output of the component.
(Caution: for NormalizeComponent, there is a member
"add-log-stddev" which if true, will increase the output
dim by one, so it will be "dim" plus one.
self-repair-scale=0.0 A scale for the self-repair mechanism (which nudges
the activation values towards the 'good' regions when a particular
dimension of the activations seem to be oversaturated or otherwise
unbalanced. This is typically set from the script level to values
like 1.0e-04 to 1.0e-05.
self-repair-lower-threshold=-1000 A lower threshold for the self-repair mechanism;
it will be interpreted in a component-specific way, typically a lower
limit on the average derivative or activation below which the
self-repair mechanism is activated. -1000 is a special value which
will cause a component-specific default to be used.
self-repair-upper-threshold=-1000 An upper threshold for the self-repair mechanism;
it will be interpreted in a component-specific way, typically an upper
limit on the average derivative or activation above which the
self-repair mechanism is activated. -1000 is a special value which
will cause a component-specific default to be used.
block-dim Defaults to dim, but may be any divisor of dim. It affects the
self-repair, which will be done while treating the input/output as
repeating blocks of size 'block-dim' (e.g. blocks of filters). It allows
us to do self-repair on the filter level in CNNs.
Currently this only makes a difference for RectifiedLinearComponent.
*/
class NonlinearComponent: public Component {
public:
NonlinearComponent();
explicit NonlinearComponent(const NonlinearComponent &other);
virtual int32 InputDim() const { return dim_; }
virtual int32 OutputDim() const { return dim_; }
// We implement InitFromConfig at this level and this version is sufficient
// for most of the child classes. Note: it's overridden by class
// NormalizeComponent.
virtual void InitFromConfig(ConfigLine *cfl);
/// We implement Read at this level as it just needs the Type().
virtual void Read(std::istream &is, bool binary);
virtual void ZeroStats();
virtual std::string Info() const;
/// Write component to stream.
virtual void Write(std::ostream &os, bool binary) const;
virtual void Scale(BaseFloat scale);
virtual void Add(BaseFloat alpha, const Component &other);
virtual void ConsolidateMemory();
// The following functions are unique to NonlinearComponent.
// They mostly relate to diagnostics.
const CuVector<double> &ValueSum() const { return value_sum_; }
const CuVector<double> &DerivSum() const { return deriv_sum_; }
double Count() const { return count_; }
protected:
enum { kUnsetThreshold = -1000 };
friend class SigmoidComponent;
friend class TanhComponent;
friend class SoftmaxComponent;
friend class LogSoftmaxComponent;
friend class RectifiedLinearComponent;
// This function updates the stats "value_sum_", "deriv_sum_", and
// count_. (If deriv == NULL, it won't update "deriv_sum_").
// It will be called from the Backprop function of child classes.
void StoreStatsInternal(const CuMatrixBase<BaseFloat> &out_value,
const CuMatrixBase<BaseFloat> *deriv = NULL);
// This function may be called from child class members during backprop. It
// stores the 'oderiv_sumsq_' stats.
void StoreBackpropStats(const CuMatrixBase<BaseFloat> &out_deriv);
const NonlinearComponent &operator = (const NonlinearComponent &other); // Disallow.
// dim_ is the input dimension (and almost always the output dimension) of the
// component.
int32 dim_;
// block_dim_ will normally be the same as dim_, but it may be any nonzero
// divisor of dim_; if so, each vector is treated as a number of blocks
// appended together, and this affects the stats accumulation and self-repair.
// Currently this is only supported for RectifiedLinearComponent.
int32 block_dim_;
CuVector<double> value_sum_; // stats at the output.
CuVector<double> deriv_sum_; // stats of the derivative of the nonlinearity
// (only applicable to element-by-element
// nonlinearities, not Softmax.
// Count corresponding to the stats in 'value_sum_' and 'deriv_sum_'
double count_;
CuVector<double> oderiv_sumsq_; // Sum-square of the derivative of the
// objective function, that we're propagating
// back. Accumulated during the backprop;
// used for diagnostics.
// Count corresponding to the stats in 'oderiv_sumsq_'.
double oderiv_count_;
// some stats for self-repairing nonlinearities.
double num_dims_self_repaired_;
double num_dims_processed_;
// some configuration values relating to self-repairing nonlinearities.
BaseFloat self_repair_lower_threshold_;
BaseFloat self_repair_upper_threshold_;
BaseFloat self_repair_scale_;
};
} // namespace nnet3
} // namespace kaldi
#endif