Mercurial > pylearn

annotate doc/v2_planning/optimization.txt @ 1057:baf1988db557

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
v2planning optimization - added API
author James Bergstra <bergstrj@iro.umontreal.ca>
date Thu, 09 Sep 2010 11:32:42 -0400
parents 89e76e6e074f
children a41cc29cee26
rev   line source
1009
dc5185cca21e Added files for Coding Style and Optimization committees
Olivier Delalleau <delallea@iro>
parents:
diff changeset
1 Discussion of Optimization-Related Issues
dc5185cca21e Added files for Coding Style and Optimization committees
Olivier Delalleau <delallea@iro>
parents:
diff changeset
2 =========================================
dc5185cca21e Added files for Coding Style and Optimization committees
Olivier Delalleau <delallea@iro>
parents:
diff changeset
3
1036
89e76e6e074f XG added to optimization team
Xavier Glorot <glorotxa@iro.umontreal.ca>
parents: 1027
diff changeset
4 Members: JB, PL, OD, XG
1013
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
5
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
6 Representative: JB
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
7
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
8
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
9 Previous work - scikits, openopt, scipy provide function optimization
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
10 algorithms. These are not currently GPU-enabled but may be in the future.
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
11
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
12
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
13 IS PREVIOUS WORK SUFFICIENT?
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
14 --------------------------------
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
15
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
16 In many cases it is (I used it for sparse coding, and it was ok).
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
17
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
18 These packages provide batch optimization, whereas we typically need online
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
19 optimization.
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
20
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
21 It can be faster (to run) and more convenient (to implement) to have
1016
618b9fdbfda5 optimization: Minor typo fixes
Olivier Delalleau <delallea@iro>
parents: 1013
diff changeset
22 optimization algorithms as Theano update expressions.
1013
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
23
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
24
1016
618b9fdbfda5 optimization: Minor typo fixes
Olivier Delalleau <delallea@iro>
parents: 1013
diff changeset
25 What optimization algorithms do we want/need?
1013
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
26 ---------------------------------------------
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
27
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
28 - sgd
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
29 - sgd + momentum
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
30 - sgd with annealing schedule
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
31 - TONGA
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
32 - James Marten's Hessian-free
1027
a1b6ccd5b6dc few comments added
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 1016
diff changeset
33 - Conjugate gradients, batch and (large) mini-batch [that is also what Marten's thing does]
1013
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
34
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
35 Do we need anything to make batch algos work better with Pylearn things?
1027
a1b6ccd5b6dc few comments added
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 1016
diff changeset
36 - conjugate methods? yes
a1b6ccd5b6dc few comments added
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 1016
diff changeset
37 - L-BFGS? maybe, when needed
1013
5e9a3d9bc0b4 optimization - added some text
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1009
diff changeset
38
1027
a1b6ccd5b6dc few comments added
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 1016
diff changeset
39
a1b6ccd5b6dc few comments added
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 1016
diff changeset
40
a1b6ccd5b6dc few comments added
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 1016
diff changeset
41
1057
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
42
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
43 Proposal for API
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
44 ================
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
45
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
46 Stick to the same style of API that we've used for SGD so far. I think it has
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
47 worked well. It takes theano expressions as inputs and returns theano
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
48 expressions as results. The caller is responsible for building those
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
49 expressions into a callable function that does the minimization (and other
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
50 things too maybe).
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
51
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
52
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
53 def stochastic_gradientbased_optimization_updates(parameters, cost=None, grads=None, **kwargs):
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
54 """
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
55 :param parameters: list or tuple of Theano variables (typically shared vars)
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
56 that we want to optimize iteratively algorithm.
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
57
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
58 :param cost: scalar-valued Theano variable that computes noisy estimate of
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
59 cost (what are the conditions on the noise?). The cost is ignored if
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
60 grads are given.
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
61
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
62 :param grads: list or tuple of Theano variables representing the gradients on
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
63 the corresponding parameters. These default to tensor.grad(cost,
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
64 parameters).
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
65
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
66 :param kwargs: algorithm-dependent arguments
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
67
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
68 :returns: a list of pairs (v, new_v) that indicate the value (new_v) each
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
69 variable (v) should take in order to carry out the optimization procedure.
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
70
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
71 The first section of the return value list corresponds to the terms in
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
72 `parameters`, and the optimization algorithm can return additional update
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
73 expression afterward. This list of pairs can be passed directly to the
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
74 dict() constructor to create a dictionary such that dct[v] == new_v.
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
75 """
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
76
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
77
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
78 Why not a class interface with an __init__ that takes the kwargs, and an
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
79 updates() that returns the updates? It would be wrong for auxiliary shared
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
80 variables to be involved in two updates, so the interface should not encourage
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
81 separate methods for those two steps.
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
82
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
83
baf1988db557 v2planning optimization - added API
James Bergstra <bergstrj@iro.umontreal.ca>
parents: 1036
diff changeset
84