Hot-keys on this page
r m x p toggle line displays
j k next/prev highlighted chunk
0 (zero) top of page
1 (one) first highlighted chunk
1"""
2Implements simple stochastic gradient optimisation.
3It is inspired from `_stochastic_optimizers.py
4<https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/
5neural_network/_stochastic_optimizers.py>`_.
6"""
7import numpy
8from numpy.core._exceptions import UFuncTypeError
11class BaseOptimizer:
12 """
13 Base stochastic gradient descent optimizer.
15 :param coef: array, initial coefficient
16 :param learning_rate_init: float
17 The initial learning rate used. It controls the step-size
18 in updating the weights.
19 :param min_threshold: coefficients must be higher than *min_thresold*
20 :param max_threshold: coefficients must be below than *max_thresold*
22 The class holds the following attributes:
24 * *learning_rate*: float, the current learning rate
25 """
27 def __init__(self, coef, learning_rate_init=0.1,
28 min_threshold=None, max_threshold=None):
29 if not isinstance(coef, numpy.ndarray):
30 raise TypeError("coef must be an array.")
31 self.coef = coef
32 self.learning_rate_init = learning_rate_init
33 self.learning_rate = float(learning_rate_init)
34 self.min_threshold = min_threshold
35 self.max_threshold = max_threshold
37 def _get_updates(self, grad):
38 raise NotImplementedError("Must be overwritten.") # pragma no cover
40 def update_coef(self, grad):
41 """
42 Updates coefficients with given gradient.
44 :param grad: array, gradient
45 """
46 if self.coef.shape != grad.shape:
47 raise ValueError("coef and grad must have the same shape.")
48 update = self._get_updates(grad)
49 self.coef += update
50 if self.min_threshold is not None:
51 try:
52 self.coef = numpy.maximum(self.coef, self.min_threshold)
53 except UFuncTypeError: # pragma: no cover
54 raise RuntimeError(
55 "Unable to compute an upper bound with coef={} "
56 "max_threshold={}".format(self.coef, self.min_threshold))
57 if self.max_threshold is not None:
58 try:
59 self.coef = numpy.minimum(self.coef, self.max_threshold)
60 except UFuncTypeError: # pragma: no cover
61 raise RuntimeError(
62 "Unable to compute a lower bound with coef={} "
63 "max_threshold={}".format(self.coef, self.max_threshold))
65 def iteration_ends(self, time_step):
66 """
67 Performs update to learning rate and potentially other states at the
68 end of an iteration.
69 """
70 pass # pragma: no cover
72 def train(self, X, y, fct_loss, fct_grad, max_iter=100,
73 early_th=None, verbose=False):
74 """
75 Optimizes the coefficients.
77 :param X: datasets (array)
78 :param y: expected target
79 :param fct_loss: loss function, signature: `f(coef, X, y) -> float`
80 :param fct_grad: gradient function,
81 signature: `g(coef, x, y, i) -> array`
82 :param max_iter: number maximum of iteration
83 :param early_th: stops the training if the error goes below
84 this threshold
85 :param verbose: display information
86 :return: loss
87 """
88 if not isinstance(X, numpy.ndarray):
89 raise TypeError("X must be an array.")
90 if not isinstance(y, numpy.ndarray):
91 raise TypeError("y must be an array.")
92 if X.shape[0] != y.shape[0]:
93 raise ValueError("X and y must have the same number of rows.")
94 if any(numpy.isnan(X.ravel())):
95 raise ValueError("X contains nan value.")
96 if any(numpy.isnan(y.ravel())):
97 raise ValueError("y contains nan value.")
99 loss = fct_loss(self.coef, X, y)
100 losses = [loss]
101 if verbose:
102 self._display_progress(0, max_iter, loss)
103 n_samples = 0
104 for it in range(max_iter):
105 irows = numpy.random.choice(X.shape[0], X.shape[0])
106 for irow in irows:
107 grad = fct_grad(self.coef, X[irow, :], y[irow], irow)
108 if isinstance(verbose, int) and verbose >= 10:
109 self._display_progress( # pragma: no cover
110 0, max_iter, loss, grad, 'grad')
111 if numpy.isnan(grad).sum() > 0:
112 raise RuntimeError( # pragma: no cover
113 "The gradient has nan values.")
114 self.update_coef(grad)
115 n_samples += 1
117 self.iteration_ends(n_samples)
118 loss = fct_loss(self.coef, X, y)
119 if verbose:
120 self._display_progress(it + 1, max_iter, loss)
121 self.iter_ = it + 1
122 losses.append(loss)
123 if self._evaluate_early_stopping(
124 it, max_iter, losses, early_th, verbose=verbose):
125 break
126 return loss
128 def _evaluate_early_stopping(
129 self,
130 it,
131 max_iter,
132 losses,
133 early_th,
134 verbose=False):
135 if len(losses) < 5 or early_th is None:
136 return False
137 if numpy.isnan(losses[-5]):
138 if numpy.isnan(losses[-1]): # pragma: no cover
139 if verbose:
140 self._display_progress(it + 1, max_iter, losses[-1],
141 losses=losses[-5:])
142 return True
143 return False # pragma: no cover
144 if numpy.isnan(losses[-1]):
145 if verbose: # pragma: no cover
146 self._display_progress(it + 1, max_iter, losses[-1],
147 losses=losses[-5:])
148 return True # pragma: no cover
149 if abs(losses[-1] - losses[-5]) <= early_th:
150 if verbose: # pragma: no cover
151 self._display_progress(it + 1, max_iter, losses[-1],
152 losses=losses[-5:])
153 return True
154 return False
156 def _display_progress(self, it, max_iter, loss, losses=None):
157 'Displays training progress.'
158 if losses is None: # pragma: no cover
159 print('{}/{}: loss: {:1.4g}'.format(it, max_iter, loss))
160 else:
161 print( # pragma: no cover
162 '{}/{}: loss: {:1.4g} losses: {}'.format(
163 it, max_iter, loss, losses))
166class SGDOptimizer(BaseOptimizer):
167 """
168 Stochastic gradient descent optimizer with momentum.
170 :param coef: array, initial coefficient
171 :param learning_rate_init: float
172 The initial learning rate used. It controls the step-size
173 in updating the weights,
174 :param lr_schedule: `{'constant', 'adaptive', 'invscaling'}`,
175 learning rate schedule for weight updates,
176 `'constant'` for a constant learning rate given by
177 *learning_rate_init*. `'invscaling'` gradually decreases
178 the learning rate *learning_rate_* at each time step *t*
179 using an inverse scaling exponent of *power_t*.
180 `learning_rate_ = learning_rate_init / pow(t, power_t)`,
181 `'adaptive'`, keeps the learning rate constant to
182 *learning_rate_init* as long as the training keeps decreasing.
183 Each time 2 consecutive epochs fail to decrease the training loss by
184 tol, or fail to increase validation score by tol if 'early_stopping'
185 is on, the current learning rate is divided by 5.
186 :param momentum: float
187 Value of momentum used, must be larger than or equal to 0
188 :param power_t: double
189 The exponent for inverse scaling learning rate.
190 :param early_th: stops if the error goes below that threshold
191 :param min_threshold: lower bound for parameters (can be None)
192 :param max_threshold: upper bound for parameters (can be None)
194 The class holds the following attributes:
196 * *learning_rate*: float, the current learning rate
197 * velocity*: array, velocity that are used to update params
199 .. exref::
200 :title: Stochastic Gradient Descent applied to linear regression
202 The following example how to optimize a simple linear regression.
204 .. runpython::
205 :showcode:
207 import numpy
208 from aftercovid.optim import SGDOptimizer
211 def fct_loss(c, X, y):
212 return numpy.linalg.norm(X @ c - y) ** 2
215 def fct_grad(c, x, y, i=0):
216 return x * (x @ c - y) * 0.1
219 coef = numpy.array([0.5, 0.6, -0.7])
220 X = numpy.random.randn(10, 3)
221 y = X @ coef
223 sgd = SGDOptimizer(numpy.random.randn(3))
224 sgd.train(X, y, fct_loss, fct_grad, max_iter=15, verbose=True)
225 print('optimized coefficients:', sgd.coef)
226 """
228 def __init__(self, coef, learning_rate_init=0.1, lr_schedule='constant',
229 momentum=0.9, power_t=0.5, early_th=None,
230 min_threshold=None, max_threshold=None):
231 super().__init__(coef, learning_rate_init,
232 min_threshold=min_threshold,
233 max_threshold=max_threshold)
234 self.lr_schedule = lr_schedule
235 self.momentum = momentum
236 self.power_t = power_t
237 self.early_th = early_th
238 self.velocity = numpy.zeros_like(coef)
240 def iteration_ends(self, time_step):
241 """
242 Performs updates to learning rate and potential other states at the
243 end of an iteration.
245 :param time_step: int
246 number of training samples trained on so far, used to update
247 learning rate for 'invscaling'
248 """
249 if self.lr_schedule == 'invscaling':
250 self.learning_rate = (float(self.learning_rate_init) /
251 (time_step + 1) ** self.power_t)
253 def _get_updates(self, grad):
254 """
255 Gets the values used to update params with given gradients.
257 :param grad: array, gradient
258 :return: updates, array, the values to add to params
259 """
260 update = self.momentum * self.velocity - self.learning_rate * grad
261 self.velocity = update
262 return update
264 def _display_progress(self, it, max_iter, loss, losses=None, msg='loss'):
265 'Displays training progress.'
266 if losses is None:
267 print('{}/{}: {}: {:1.4g} lr={:1.3g}'.format(
268 it, max_iter, msg, loss, self.learning_rate))
269 else:
270 print( # pragma: no cover
271 '{}/{}: {}: {:1.4g} lr={:1.3g} {}es: {}'.format(
272 it, max_iter, msg, loss, self.learning_rate, msg, losses))