@@ -332,155 +332,39 @@ cdef class SquaredEpsilonInsensitive(Regression):
332332 return SquaredEpsilonInsensitive, (self .epsilon,)
333333
334334
335- def plain_sgd (np.ndarray[double , ndim = 1 , mode = ' c'
336- double intercept ,
337- LossFunction loss ,
338- int penalty_type ,
339- double alpha , double C ,
340- double l1_ratio ,
341- SequentialDataset dataset ,
342- np.ndarray[unsigned char , ndim = 1 , mode = ' c'
343- bint early_stopping , validation_score_cb ,
344- int n_iter_no_change ,
345- int max_iter , double tol , int fit_intercept ,
346- int verbose , bint shuffle , np.uint32_t seed ,
347- double weight_pos , double weight_neg ,
348- int learning_rate , double eta0 ,
349- double power_t ,
350- double t = 1.0 ,
351- double intercept_decay = 1.0 ):
352- """ Plain SGD for generic loss functions and penalties.
353-
354- Parameters
355- ----------
356- weights : ndarray[double, ndim=1]
357- The allocated coef_ vector.
358- intercept : double
359- The initial intercept.
360- loss : LossFunction
361- A concrete ``LossFunction`` object.
362- penalty_type : int
363- The penalty 2 for L2, 1 for L1, and 3 for Elastic-Net.
364- alpha : float
365- The regularization parameter.
366- C : float
367- Maximum step size for passive aggressive.
368- l1_ratio : float
369- The Elastic Net mixing parameter, with 0 <= l1_ratio <= 1.
370- l1_ratio=0 corresponds to L2 penalty, l1_ratio=1 to L1.
371- dataset : SequentialDataset
372- A concrete ``SequentialDataset`` object.
373- validation_mask : ndarray[unsigned char, ndim=1]
374- Equal to True on the validation set.
375- early_stopping : boolean
376- Whether to use a stopping criterion based on the validation set.
377- validation_score_cb : callable
378- A callable to compute a validation score given the current
379- coefficients and intercept values.
380- Used only if early_stopping is True.
381- n_iter_no_change : int
382- Number of iteration with no improvement to wait before stopping.
383- max_iter : int
384- The maximum number of iterations (epochs).
385- tol: double
386- The tolerance for the stopping criterion.
387- fit_intercept : int
388- Whether or not to fit the intercept (1 or 0).
389- verbose : int
390- Print verbose output; 0 for quite.
391- shuffle : boolean
392- Whether to shuffle the training data before each epoch.
393- weight_pos : float
394- The weight of the positive class.
395- weight_neg : float
396- The weight of the negative class.
397- seed : np.uint32_t
398- Seed of the pseudorandom number generator used to shuffle the data.
399- learning_rate : int
400- The learning rate:
401- (1) constant, eta = eta0
402- (2) optimal, eta = 1.0/(alpha * t).
403- (3) inverse scaling, eta = eta0 / pow(t, power_t)
404- (4) adaptive decrease
405- (5) Passive Aggressive-I, eta = min(alpha, loss/norm(x))
406- (6) Passive Aggressive-II, eta = 1.0 / (norm(x) + 0.5*alpha)
407- eta0 : double
408- The initial learning rate.
409- power_t : double
410- The exponent for inverse scaling learning rate.
411- t : double
412- Initial state of the learning rate. This value is equal to the
413- iteration count except when the learning rate is set to `optimal`.
414- Default: 1.0.
415- intercept_decay : double
416- The decay ratio of intercept, used in updating intercept.
417-
418- Returns
419- -------
420- weights : array, shape=[n_features]
421- The fitted weight vector.
422- intercept : float
423- The fitted intercept term.
424- n_iter_ : int
425- The actual number of iter (epochs).
426- """
427- standard_weights, standard_intercept,\
428- _, _, n_iter_ = _plain_sgd(weights,
429- intercept,
430- None ,
431- 0 ,
432- loss,
433- penalty_type,
434- alpha, C,
435- l1_ratio,
436- dataset,
437- validation_mask,
438- early_stopping,
439- validation_score_cb,
440- n_iter_no_change,
441- max_iter, tol, fit_intercept,
442- verbose, shuffle, seed,
443- weight_pos, weight_neg,
444- learning_rate, eta0,
445- power_t,
446- t,
447- intercept_decay,
448- 0 )
449- return standard_weights, standard_intercept, n_iter_
450-
451-
452- def average_sgd (np.ndarray[double , ndim = 1 , mode = ' c'
453- double intercept ,
454- np.ndarray[double , ndim = 1 , mode = ' c'
455- double average_intercept ,
456- LossFunction loss ,
457- int penalty_type ,
458- double alpha , double C ,
459- double l1_ratio ,
460- SequentialDataset dataset ,
461- np.ndarray[unsigned char , ndim = 1 , mode = ' c'
462- bint early_stopping , validation_score_cb ,
463- int n_iter_no_change ,
464- int max_iter , double tol , int fit_intercept ,
465- int verbose , bint shuffle , np.uint32_t seed ,
466- double weight_pos , double weight_neg ,
467- int learning_rate , double eta0 ,
468- double power_t ,
469- double t = 1.0 ,
470- double intercept_decay = 1.0 ,
471- int average = 1 ):
472- """ Average SGD for generic loss functions and penalties.
335+ def _plain_sgd (np.ndarray[double , ndim = 1 , mode = ' c'
336+ double intercept ,
337+ np.ndarray[double , ndim = 1 , mode = ' c'
338+ double average_intercept ,
339+ LossFunction loss ,
340+ int penalty_type ,
341+ double alpha , double C ,
342+ double l1_ratio ,
343+ SequentialDataset dataset ,
344+ np.ndarray[unsigned char , ndim = 1 , mode = ' c'
345+ bint early_stopping , validation_score_cb ,
346+ int n_iter_no_change ,
347+ int max_iter , double tol , int fit_intercept ,
348+ int verbose , bint shuffle , np.uint32_t seed ,
349+ double weight_pos , double weight_neg ,
350+ int learning_rate , double eta0 ,
351+ double power_t ,
352+ double t = 1.0 ,
353+ double intercept_decay = 1.0 ,
354+ int average = 0 ):
355+ """ SGD for generic loss functions and penalties with optional averaging
473356
474357 Parameters
475358 ----------
476359 weights : ndarray[double, ndim=1]
477- The allocated coef_ vector.
360+ The allocated vector of weights .
478361 intercept : double
479362 The initial intercept.
480363 average_weights : ndarray[double, ndim=1]
481- The average weights as computed for ASGD
364+ The average weights as computed for ASGD. Should be None if average
365+ is 0.
482366 average_intercept : double
483- The average intercept for ASGD
367+ The average intercept for ASGD. Should be 0 if average is 0.
484368 loss : LossFunction
485369 A concrete ``LossFunction`` object.
486370 penalty_type : int
@@ -549,55 +433,14 @@ def average_sgd(np.ndarray[double, ndim=1, mode='c'] weights,
549433 intercept : float
550434 The fitted intercept term.
551435 average_weights : array shape=[n_features]
552- The averaged weights across iterations
436+ The averaged weights across iterations. Values are valid only if
437+ average > 0.
553438 average_intercept : float
554- The averaged intercept across iterations
439+ The averaged intercept across iterations.
440+ Values are valid only if average > 0.
555441 n_iter_ : int
556442 The actual number of iter (epochs).
557443 """
558- return _plain_sgd(weights,
559- intercept,
560- average_weights,
561- average_intercept,
562- loss,
563- penalty_type,
564- alpha, C,
565- l1_ratio,
566- dataset,
567- validation_mask,
568- early_stopping,
569- validation_score_cb,
570- n_iter_no_change,
571- max_iter, tol, fit_intercept,
572- verbose, shuffle, seed,
573- weight_pos, weight_neg,
574- learning_rate, eta0,
575- power_t,
576- t,
577- intercept_decay,
578- average)
579-
580-
581- def _plain_sgd (np.ndarray[double , ndim = 1 , mode = ' c'
582- double intercept ,
583- np.ndarray[double , ndim = 1 , mode = ' c'
584- double average_intercept ,
585- LossFunction loss ,
586- int penalty_type ,
587- double alpha , double C ,
588- double l1_ratio ,
589- SequentialDataset dataset ,
590- np.ndarray[unsigned char , ndim = 1 , mode = ' c'
591- bint early_stopping , validation_score_cb ,
592- int n_iter_no_change ,
593- int max_iter , double tol , int fit_intercept ,
594- int verbose , bint shuffle , np.uint32_t seed ,
595- double weight_pos , double weight_neg ,
596- int learning_rate , double eta0 ,
597- double power_t ,
598- double t = 1.0 ,
599- double intercept_decay = 1.0 ,
600- int average = 0 ):
601444
602445 # get the data information into easy vars
603446 cdef Py_ssize_t n_samples = dataset.n_samples
0 commit comments