-
Notifications
You must be signed in to change notification settings - Fork 44
/
Copy pathSVRGFunction.py
280 lines (211 loc) · 15.5 KB
/
SVRGFunction.py
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
# Copyright 2024 United Kingdom Research and Innovation
# Copyright 2024 The University of Manchester
#
# 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.
#
# Authors:
# - CIL Developers, listed at: https://github.com/TomographicImaging/CIL/blob/master/NOTICE.txt
# - Daniel Deidda (National Physical Laboratory, UK)
# - Claire Delplancke (Electricite de France, Research and Development)
# - Ashley Gillman (Australian e-Health Res. Ctr., CSIRO, Brisbane, Queensland, Australia)
# - Zeljko Kereta (Department of Computer Science, University College London, UK)
# - Evgueni Ovtchinnikov (STFC - UKRI)
# - Georg Schramm (Department of Imaging and Pathology, Division of Nuclear Medicine, KU Leuven, Leuven, Belgium)
from .ApproximateGradientSumFunction import ApproximateGradientSumFunction
import numpy as np
import numbers
class SVRGFunction(ApproximateGradientSumFunction):
r"""
The Stochastic Variance Reduced Gradient (SVRG) function calculates the approximate gradient of :math:`\sum_{i=1}^{n-1}f_i`. For this approximation, every `snapshot_update_interval` number of iterations, a full gradient calculation is made at this "snapshot" point. Intermediate gradient calculations update this snapshot by taking a index :math:`i_k` and calculating the gradient of :math:`f_{i_k}`'s at the current iterate and the snapshot, updating the approximate gradient to be:
.. math ::
n*\nabla f_{i_k}(x_k) - n*\nabla f_{i_k}(\tilde{x}) + \nabla \sum_{i=0}^{n-1}f_i(\tilde{x}),
where :math:`\tilde{x}` is the latest "snapshot" point and :math:`x_k` is the value at the current iteration.
Note
-----
Compared with the literature, we multiply by :math:`n`, the number of functions, so that we return an approximate gradient of the whole sum function and not an average gradient.
Note
----
In the case where `store_gradients=False` the memory requirements are 4 times the image size (1 stored full gradient at the "snapshot", one stored "snapshot" point and two lots of intermediary calculations). Alternatively, if `store_gradients=True` the memory requirement is `n+4` (`n` gradients at the snapshot for each function in the sum, one stored full gradient at the "snapshot", one stored "snapshot" point and two lots of intermediary calculations).
Reference
---------
Johnson, R. and Zhang, T., 2013. Accelerating stochastic gradient descent using predictive variance reduction. Advances in neural information processing systems, 26.https://proceedings.neurips.cc/paper_files/paper/2013/file/ac1dd209cbcc5e5d1c6e28598e8cbbe8-Paper.pdf
Parameters
----------
functions : `list` of functions
A list of functions: :code:`[f_{0}, f_{1}, ..., f_{n-1}]`. Each function is assumed to be smooth with an implemented :func:`~Function.gradient` method. All functions must have the same domain. The number of functions must be strictly greater than 1.
sampler: An instance of a CIL Sampler class ( :meth:`~optimisation.utilities.sampler`) or of another class which has a `next` function implemented to output integers in {0, 1, ..., n-1}.
This sampler is called each time gradient is called and sets the internal `function_num` passed to the `approximate_gradient` function. Default is `Sampler.random_with_replacement(len(functions))`.
snapshot_update_interval : positive int or None, optional
The interval for updating the full gradient (taking a snapshot). The default is 2*len(functions) so a "snapshot" is taken every 2*len(functions) iterations. If the user passes `0` then no full gradient snapshots will be taken.
store_gradients : bool, default: `False`
Flag indicating whether to store an update a list of gradients for each function :math:`f_i` or just to store the snapshot point :math:`\tilde{x}` and its gradient :math:`\nabla \sum_{i=0}^{n-1}f_i(\tilde{x})`.
"""
def __init__(self, functions, sampler=None, snapshot_update_interval=None, store_gradients=False):
super(SVRGFunction, self).__init__(functions, sampler)
# snapshot_update_interval for SVRG
self.snapshot_update_interval = snapshot_update_interval
if snapshot_update_interval is None:
self.snapshot_update_interval = 2*self.num_functions
self.store_gradients = store_gradients
self._svrg_iter_number = 0
self._full_gradient_at_snapshot = None
self._list_stored_gradients = None
self.stoch_grad_at_iterate = None
self._stochastic_grad_difference = None
self.snapshot = None
def gradient(self, x, out=None):
r""" Selects a random function using the `sampler` and then calls the approximate gradient at :code:`x` or calculates a full gradient depending on the update frequency
Parameters
----------
x : DataContainer (e.g. ImageData object)
out: return DataContainer, if `None` a new DataContainer is returned, default `None`.
Returns
--------
DataContainer (e.g. ImageData object)
the value of the approximate gradient of the sum function at :code:`x`
"""
# For SVRG, every `snapshot_update_interval` a full gradient step is calculated, else an approximate gradient is taken.
if ( (self.snapshot_update_interval != 0) and (self._svrg_iter_number % (self.snapshot_update_interval)) == 0):
return self._update_full_gradient_and_return(x, out=out)
else:
self.function_num = self.sampler.next()
if not isinstance(self.function_num, numbers.Number):
raise ValueError("Batch gradient is not yet implemented")
if self.function_num >= self.num_functions or self.function_num < 0:
raise IndexError(
f"The sampler has produced the index {self.function_num} which does not match the expected range of available functions to sample from. Please ensure your sampler only selects from [0,1,...,len(functions)-1] ")
return self.approximate_gradient(x, self.function_num, out=out)
def approximate_gradient(self, x, function_num, out=None):
r""" Calculates the stochastic gradient at the point :math:`x` by using the gradient of the selected function, indexed by :math:`i_k`, the `function_number` in {0,...,len(functions)-1}, and the full gradient at the snapshot :math:`\tilde{x}`
.. math ::
n*\nabla f_{i_k}(x_k) - n*\nabla f_{i_k}(\tilde{x}) + \nabla \sum_{i=0}^{n-1}f_i(\tilde{x})
Note
-----
Compared with the literature, we multiply by :math:`n`, the number of functions, so that we return an approximate gradient of the whole sum function and not an average gradient.
Parameters
----------
x : DataContainer (e.g. ImageData object)
out: return DataContainer, if `None` a new DataContainer is returned, default `None`.
function_num: `int`
Between 0 and n-1, where n is the number of functions in the list
Returns
--------
DataContainer (e.g. ImageData object)
the value of the approximate gradient of the sum function at :code:`x` given a `function_number` in {0,...,len(functions)-1}
"""
self._svrg_iter_number += 1
self.stoch_grad_at_iterate = self.functions[function_num].gradient(x, out=self.stoch_grad_at_iterate)
if self.store_gradients is True:
self._stochastic_grad_difference = self.stoch_grad_at_iterate.sapyb(
1., self._list_stored_gradients[function_num], -1., out=self._stochastic_grad_difference)
else:
self._stochastic_grad_difference = self.stoch_grad_at_iterate.sapyb(
1., self.functions[function_num].gradient(self.snapshot), -1., out=self._stochastic_grad_difference)
self._update_data_passes_indices([function_num])
out = self._stochastic_grad_difference.sapyb(
self.num_functions, self._full_gradient_at_snapshot, 1., out=out)
return out
def _update_full_gradient_and_return(self, x, out=None):
"""
Takes a "snapshot" at the point :math:`x`, saving both the point :math:` \tilde{x}=x` and its gradient :math:`\sum_{i=0}^{n-1}f_i{\tilde{x}}`. The function returns :math:`\sum_{i=0}^{n-1}f_i{\tilde{x}}` as the gradient calculation. If :code:`store_gradients==True`, the gradient of all the :math:`f_i`s is computed and stored at the "snapshot"..
Parameters
----------
Takes a "snapshot" at the point :math:`x`. The function returns :math:`\sum_{i=0}^{n-1}f_i{\tilde{x}}` as the gradient calculation. If :code:`store_gradients==True`, the gradient of all the :math:`f_i`s is stored, otherwise only the sum of the gradients and the snapshot point :math:` \tilde{x}=x` are stored.
out: return DataContainer, if `None` a new DataContainer is returned, default `None`.
Returns
--------
DataContainer (e.g. ImageData object)
the value of the approximate gradient of the sum function at :code:`x` given a `function_number` in {0,...,len(functions)-1}
"""
self._svrg_iter_number += 1
if self.store_gradients is True:
if self._list_stored_gradients is None:
# Save the gradient of each individual f_i and the gradient of the full sum at the point x.
self._list_stored_gradients = [
fi.gradient(x) for fi in self.functions]
self._full_gradient_at_snapshot = sum(
self._list_stored_gradients, start=0*x)
else:
for i, fi in enumerate(self.functions):
fi.gradient(x, out=self._list_stored_gradients[i])
self._full_gradient_at_snapshot.fill(
sum(self._list_stored_gradients, start=0*x))
self._full_gradient_at_snapshot *= 0
for i, el in enumerate(self._list_stored_gradients):
self._full_gradient_at_snapshot += el
else:
# Save the snapshot point and the gradient of the full sum at the point x.
self._full_gradient_at_snapshot = self.full_gradient(
x, out=self._full_gradient_at_snapshot)
if self.snapshot is None:
self.snapshot = x.copy()
self.snapshot.fill(x)
# In this iteration all functions in the sum were used to update the gradient
self._update_data_passes_indices(list(range(self.num_functions)))
# Return the gradient of the full sum at the snapshot.
if out is None:
out = self._full_gradient_at_snapshot
else:
out.fill(self._full_gradient_at_snapshot)
return out
class LSVRGFunction(SVRGFunction):
r"""
A class representing a function for Loopless Stochastic Variance Reduced Gradient (SVRG) approximation. This is similar to SVRG, except the full gradient at a "snapshot" is calculated at random intervals rather than at fixed numbers of iterations.
Parameters
----------
functions : `list` of functions
A list of functions: :code:`[f_{0}, f_{1}, ..., f_{n-1}]`. Each function is assumed to be smooth with an implemented :func:`~Function.gradient` method. All functions must have the same domain. The number of functions `n` must be strictly greater than 1.
sampler: An instance of a CIL Sampler class ( :meth:`~optimisation.utilities.sampler`) or of another class which has a `next` function implemented to output integers in {0,...,n-1}.
This sampler is called each time gradient is called and sets the internal `function_num` passed to the `approximate_gradient` function. Default is `Sampler.random_with_replacement(len(functions))`.
snapshot_update_probability: positive float, default: 1/n
The probability of updating the full gradient (taking a snapshot) at each iteration. The default is :math:`1./n` so, in expectation, a snapshot will be taken every :math:`n` iterations.
store_gradients : bool, default: `False`
Flag indicating whether to store an update a list of gradients for each function :math:`f_i` or just to store the snapshot point :math:`\tilde{x}` and its gradient :math:`\nabla \sum_{i=0}^{n-1}f_i(\tilde{x})`.
Reference
---------
Kovalev, D., Horváth, S. &; Richtárik, P.. (2020). Don’t Jump Through Hoops and Remove Those Loops: SVRG and Katyusha are Better Without the Outer Loop. Proceedings of the 31st International Conference on Algorithmic Learning Theory, in Proceedings of Machine Learning Research 117:451-467 Available from https://proceedings.mlr.press/v117/kovalev20a.html.
Note
----
In the case where `store_gradients=False` the memory requirements are 4 times the image size (1 stored full gradient at the "snapshot", one stored "snapshot" point and two lots of intermediary calculations). Alternatively, if `store_gradients=True` the memory requirement is `n+4` (`n` gradients at the snapshot for each function in the sum, one stored full gradient at the "snapshot", one stored "snapshot" point and two lots of intermediary calculations).
"""
def __init__(self, functions, sampler=None, snapshot_update_probability=None, store_gradients=False, seed=None):
super(LSVRGFunction, self).__init__(
functions, sampler=sampler, store_gradients=store_gradients)
# Update frequency based on probability.
self.snapshot_update_probability = snapshot_update_probability
# Default snapshot_update_probability for Loopless SVRG
if self.snapshot_update_probability is None:
self.snapshot_update_probability = 1./self.num_functions
# The random generator used to decide if the gradient calculation is a full gradient or an approximate gradient
self.generator = np.random.default_rng(seed=seed)
def gradient(self, x, out=None):
""" Selects a random function using the `sampler` and then calls the approximate gradient at :code:`x` or calculates a full gradient depending on the update probability.
Parameters
----------
x : DataContainer (e.g. ImageData objects)
out: return DataContainer, if `None` a new DataContainer is returned, default `None`.
Returns
--------
DataContainer (e.g. ImageData object)
the value of the approximate gradient of the sum function at :code:`x`
"""
if self._svrg_iter_number == 0 or self.generator.uniform() < self.snapshot_update_probability:
return self._update_full_gradient_and_return(x, out=out)
else:
self.function_num = self.sampler.next()
if not isinstance(self.function_num, numbers.Number):
raise ValueError("Batch gradient is not yet implemented")
if self.function_num >= self.num_functions or self.function_num < 0:
raise IndexError(
f"The sampler has produced the index {self.function_num} which does not match the expected range of available functions to sample from. Please ensure your sampler only selects from [0,1,...,len(functions)-1] ")
return self.approximate_gradient(x, self.function_num, out=out)