├── src
├── PFANodeMDPRefImp.py
├── MergeLayer.py
├── cache_node.py
├── PFANodeMDP.py
└── PFACoreUtil.py
├── README
└── LICENSE
/src/PFANodeMDPRefImp.py:
--------------------------------------------------------------------------------
1 | #
2 | #
3 | # Copyright of GNUPFA:
4 | # Copyright (c) 2013, 2014 Institut fuer Neuroinformatik,
5 | # Ruhr-Universitaet Bochum, Germany. All rights reserved.
6 | #
7 | #
8 | # This file is part of GNUPFA.
9 | #
10 | # GNUPFA is free software: you can redistribute it and/or modify
11 | # it under the terms of the GNU General Public License as published by
12 | # the Free Software Foundation, either version 3 of the License, or
13 | # (at your option) any later version.
14 | #
15 | # GNUPFA is distributed in the hope that it will be useful,
16 | # but WITHOUT ANY WARRANTY; without even the implied warranty of
17 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 | # GNU General Public License for more details.
19 | #
20 | # You should have received a copy of the GNU General Public License
21 | # along with GNUPFA. If not, see .
22 | #
23 | #
24 | # Linking this library statically or dynamically with other modules is
25 | # making a combined work based on this library. Thus, the terms and
26 | # conditions of the GNU General Public License cover the whole
27 | # combination.
28 | #
29 |
30 |
31 | '''
32 | Created on 13.08.2013
33 |
34 | @author: Stefan Richthofer
35 | '''
36 | import mdp
37 | import numpy as np
38 | import PFACoreUtil as pfa
39 |
40 | class PFANode(mdp.Node):
41 | '''redifined _init_ Method with p, k as arguments '''
42 | def __init__(self, p = 2, k = 0, affine = True, input_dim = None, output_dim = None, dtype = None):
43 | super(PFANode, self).__init__(input_dim = input_dim, output_dim = output_dim, dtype = dtype)
44 | self.p = p
45 | self.k = k
46 | self.data = None
47 | self.affine = affine
48 |
49 | ''' Node is trainable '''
50 | def is_trainable(self):
51 | return True
52 |
53 | '''In this reference implementation, it simply collects the data to process it in _stop_training.'''
54 | def _train(self, x):
55 | n = self.get_input_dim()
56 | x2 = x
57 | if not n is None:
58 | x2 = x.T[:n].T
59 | if self.data is None:
60 | self.data = x2
61 | else:
62 | self.data = np.vstack([self.data, x2])
63 |
64 | def _stop_training(self):
65 | if self.data is None:
66 | raise TrainingException("train was never called")
67 | r = self.get_output_dim()
68 | if r is None:
69 | r = len(self.data.T)
70 | meanRef, SRef, zRef = pfa.calcSpheringParametersAndDataRefImp(self.data)#, threshold = 0.0000001, offset = 0, length = -1, besselsCorrection = 0)
71 | self.mean = meanRef
72 | if not self.affine:
73 | WRef = pfa.calcRegressionCoeffRefImp(zRef, self.p)
74 | XRef = pfa.calcErrorCoeffConstLenRefImp(zRef, WRef, self.k)
75 | self.Ar = pfa.calcExtractionForErrorCovRefImp(XRef, r)
76 | reduced = np.dot(self.data, self.Ar.T)
77 | self.W = pfa.calcRegressionCoeffRefImp(reduced, self.p)
78 | else:
79 | WRef = pfa.calcRegressionCoeffAffineRefImp(zRef, self.p)
80 | XRef = pfa.calcErrorCoeffConstLenAffineRefImp(zRef, WRef, self.k)
81 | Ar = pfa.calcExtractionForErrorCovRefImp(XRef, r)
82 | reduced = np.dot(zRef, Ar)
83 | self.W = pfa.calcRegressionCoeffAffineRefImp(reduced, self.p)
84 | self.Ar = np.dot(SRef, Ar)
85 |
86 | def _execute(self, x):
87 | z0 = x-np.outer(np.ones(len(x)), self.mean)
88 | return np.dot(z0, self.Ar.T)
89 |
90 |
--------------------------------------------------------------------------------
/README:
--------------------------------------------------------------------------------
1 | GNUPFA is an experimental Python-implementation of the PFA
2 | algorithm as described in http://arxiv.org/abs/1311.2503.
3 | PFA is implemented as an MDP-node (http://mdp-toolkit.sourceforge.net).
4 | When this project reaches sufficient stability, an integration
5 | into MDP is planned.
6 |
7 |
8 | Files
9 | =====
10 |
11 | PFANodeMDPRefImp.py
12 | -------------------
13 |
14 | This is a naive one-to-one implementation of the PFA algorithm.
15 | It caches all supplied data into memory and has *no* real support
16 | for chunking (as the intention of chunking is to read only a
17 | small subset of the data to memory, process it, free memory,
18 | process the next subset and so on).
19 | Further it must process the data several times; one time for
20 | each work step.
21 | Its advantage is that it is an easy to read implementation and
22 | is almost certainly bug free. So it can serve as a controlpoint
23 | for more advanced implementations.
24 |
25 |
26 | PFANodeMDP.py
27 | -------------
28 |
29 | Contains a smarter PFA implementation that supports real chunking.
30 | It only saves the mean, second moment matrix and several auto
31 | correlation matrices. Each chunk is just used to update these,
32 | eleminating any need for keeping data in memory.
33 | In contrast to PFANodeMDPRefImp, this implementation is rather
34 | complicated as the used equations are largely expanded in order
35 | to perform PFA just on top of auto correlation matrices.
36 | However, its results were compared to those of the reference
37 | implementation for various test data and the dicrepancy is on
38 | floating point numerical level.
39 |
40 | The Layer-aware node uses MergeLayer, an experimental notion to
41 | perform clone layer functionality in a more precise way. This is
42 | only relevant if one applies PFA hirarchically (c.f. mdp.hinet).
43 |
44 |
45 | PFACoreUtil.py
46 | --------------
47 |
48 | Contains various utility functions, on which PFANodeMDPRefImp and
49 | PFANodeMDP are built on. Additionally features some methods to
50 | evaluate the prediction error of extracted components empirically on
51 | given data. Currently it contains a lot of debugging outputs and
52 | requires clean-up.
53 |
54 |
55 | MergeLayer.py
56 | -------------
57 |
58 | Introduces an MDP-Layer with an additional merging-phase that merges all
59 | nodes in the layer after training. Merging is done by a given merger,
60 | which is itself a node. After merging, the merger will be used for
61 | execution in a CloneLayer-like fashion.
62 |
63 | The idea behind MergeLayer is a hybrid of ordinary layer and CloneLayer.
64 | The goal in this design is to use separate nodes in the train-phase,
65 | while using only a single node for execution. The difference to CloneLayer
66 | is that in MergeLayer, a different algorithm can be used for combining
67 | horizontally parallel data chunks than for combining time-sequent data
68 | chunks. The latter ones are combined by the nodes in the usual train-phase.
69 | In Contrast to CloneLayer, MergeLayer allows to control how horizontal merging
70 | of the data works. While CloneLayer would push this data into the very same
71 | train method like the time-sequent chunks, MergeLayer uses a merger to combine
72 | horizontal data.
73 |
74 | Note that this implementation is highly experimental.
75 |
76 |
77 | cache_node.py
78 | -------------
79 |
80 | Introduces caching functionality for MDP flows. Each flow step is saved to disc
81 | after it has been trained. This way, early nodes in the flow don't need to process
82 | the data repeatedly.
83 | And additional feature is that data can be re-ordered after caching. Most MDP nodes
84 | are agnostic to the order of their data anyway, but in layer-case, the order can matter,
85 | if image data from parallel areas is provided sequentially to the node (clone layer).
86 | MergeLayer can solve this, but a reordering cache can solve it too with even lower
87 | memory consumption, as it does not need to have parallel working memory for all areas
88 | in the layer. Note that memory consumption is crucial for potential GPU-based
89 | PFA-implementations.
90 |
91 | This implementation is highly experimental.
92 |
93 |
94 |
95 | License
96 | =======
97 |
98 | Until it gets integrated into MDP, GNUPFA will be released under GPL, v.3.
99 | See the file "LICENSE" for a copy of this GPL version.
100 |
101 |
--------------------------------------------------------------------------------
/src/MergeLayer.py:
--------------------------------------------------------------------------------
1 | #
2 | #
3 | # Copyright of GNUPFA:
4 | # Copyright (c) 2013, 2014 Institut fuer Neuroinformatik,
5 | # Ruhr-Universitaet Bochum, Germany. All rights reserved.
6 | #
7 | #
8 | # This file is part of GNUPFA.
9 | #
10 | # GNUPFA is free software: you can redistribute it and/or modify
11 | # it under the terms of the GNU General Public License as published by
12 | # the Free Software Foundation, either version 3 of the License, or
13 | # (at your option) any later version.
14 | #
15 | # GNUPFA is distributed in the hope that it will be useful,
16 | # but WITHOUT ANY WARRANTY; without even the implied warranty of
17 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 | # GNU General Public License for more details.
19 | #
20 | # You should have received a copy of the GNU General Public License
21 | # along with GNUPFA. If not, see .
22 | #
23 | #
24 | # Linking this library statically or dynamically with other modules is
25 | # making a combined work based on this library. Thus, the terms and
26 | # conditions of the GNU General Public License cover the whole
27 | # combination.
28 | #
29 |
30 |
31 | '''
32 | Created on Oct 16, 2013
33 |
34 | @author: Stefan Richthofer
35 | '''
36 |
37 | from mdp.hinet.layer import Layer
38 | from mdp import Node
39 | from mdp.hinet import FlowNode
40 |
41 | class Merger(Node):
42 | def is_trainable(self):
43 | """Per default, a merger is not trainable, since it would not be trained except,
44 | it appears multiply in the network.
45 | """
46 | return False
47 |
48 | # def needs_stop_training_before_merge(self):
49 | # return False
50 |
51 | ### Methods to be implemented by the user
52 |
53 | # this are the methods the user has to overwrite
54 |
55 | def _merge(self, node):
56 | pass
57 |
58 | def _stop_merging(self):
59 | pass
60 |
61 | def freeMem(self):
62 | pass
63 |
64 | ### User interface to the overwritten methods
65 |
66 | def merge(self, node):#, *args, **kwargs):
67 | """Update the internal structures according to the input node `node`.
68 |
69 | `node` is an mdp-node that has been trained as part of a layer. The
70 | merger-subclass should exactly know about the node-type and the node's
71 | internal structure to retrieve the relevant data from it.
72 |
73 | By default, subclasses should overwrite `_merge` to implement their
74 | merging phase. The docstring of the `_train` method overwrites this
75 | docstring.
76 | """
77 |
78 | self._merge(node)
79 |
80 | def stop_merging(self):#, *args, **kwargs):
81 | """Stop the merging phase.
82 |
83 | By default, subclasses should overwrite `_stop_merging` to implement
84 | this functionality. The docstring of the `_stop_merging` method
85 | overwrites this docstring.
86 | """
87 | self._stop_merging()#self, *args, **kwargs)
88 | #self._train_phase_started = True
89 | #self._training = False
90 |
91 | class FlowMerger(Merger, FlowNode):
92 | """
93 | Note that FlowMerger can be trainable though it is a merger.
94 | Edit: It should better not be trainable since this causes problems
95 | with train_phase and is_training.
96 |
97 | This is because the flow might contain trainable nodes as
98 | intermediate steps.
99 | You should use MergableFlowNodes instead of ordinary flows
100 | as merge sources. They ensure that the merger is used to
101 | execute the already trained/merged components.
102 | """
103 | def __init__(self, merge_indices, flow, input_dim=None, output_dim=None, dtype=None):
104 | super(FlowMerger, self).__init__(flow, input_dim, output_dim, dtype)
105 | self.merge_indices = merge_indices
106 | self.current_merge = 0
107 | #self._train_phase = len(self._train_seq)
108 |
109 | def is_training(self):
110 | return False
111 |
112 | # def needs_stop_training_before_merge(self):
113 | # i = self.merge_indices[self.current_merge]
114 | # if self.get_current_train_phase() == i:
115 | # return self._flow.flow[i].needs_stop_training_before_merge()
116 | # else:
117 | # return True
118 |
119 | def _merge(self, node):
120 | i = self.merge_indices[self.current_merge]
121 | self._flow.flow[i].merge(node._flow.flow[i])
122 | node._flow.flow[i].freeMem()
123 |
124 | def _stop_merging(self):
125 | i = self.merge_indices[self.current_merge]
126 | self._flow.flow[i]._stop_merging()
127 | self.current_merge += 1
128 | #self.self._train_phase_started = True
129 | #self._training = False
130 |
131 | # def force_stop_training(self, *args, **kwargs):
132 | # """Stop the training phase.
133 | #
134 | # By default, subclasses should overwrite `_stop_training` to implement
135 | # this functionality. The docstring of the `_stop_training` method
136 | # overwrites this docstring.
137 | # """
138 | # #if self.is_training() and self._train_phase_started == False:
139 | # # raise TrainingException("The node has not been trained.")
140 | #
141 | # #if not self.is_training():
142 | # # err_str = "The training phase has already finished."
143 | # # raise TrainingFinishedException(err_str)
144 | #
145 | # # close the current phase.
146 | # self._train_seq[self._train_phase][1](*args, **kwargs)
147 | # self._train_phase += 1
148 | # self._train_phase_started = False
149 | # # check if we have some training phase left
150 | # if self.get_remaining_train_phase() == 0:
151 | # self._training = False
152 |
153 | class MergableFlowNode(FlowNode):
154 |
155 | #def __init__(self, flow, input_dim=None, output_dim=None, dtype=None):
156 | def __init__(self, flow, merger, input_dim, output_dim, dtype):
157 | super(MergableFlowNode, self).__init__(flow, input_dim, output_dim, dtype)
158 | self.merger = merger #None
159 | #self._train_seq_cache = None
160 |
161 | #def set_merger(self, merger):
162 | # self.merger = merger
163 |
164 | # def _get_train_seq(self):
165 | # if self._train_seq_cache is None:
166 | # self._train_seq_cache = self._build_train_seq()
167 | # return self._train_seq_cache
168 | #
169 | # def _build_train_seq(self):
170 | def _get_train_seq(self):
171 | """
172 | Return a training sequence containing all training phases.
173 | In contrast to the original FlowNode, MergableFlowNode uses
174 | a given merger (must be provided by a call to set_merger)
175 | to process the data through the already trained part.
176 | """
177 | def get_train_function(_i_node, _node):
178 | # This internal function is needed to channel the data through
179 | # the nodes in front of the current nodes.
180 | # using nested scopes here instead of default args, see pep-0227
181 | def _train(x, *args, **kwargs):
182 | if i_node > 0:
183 | #_node.train(self._flow.execute(x, nodenr=_i_node-1), *args, **kwargs)
184 | #print "delegate exec in train to merger "
185 | #print str(self.merger._flow.flow[0].Ar)
186 | _node.train(self.merger._flow.execute(x, nodenr=_i_node-1), *args, **kwargs)
187 | #_node.train(self.merger.execute(x), *args, **kwargs)
188 | else:
189 | _node.train(x, *args, **kwargs)
190 | #self.merger._train_phase_started = True
191 | return _train
192 |
193 | train_seq = []
194 | for i_node, node in enumerate(self._flow):
195 | if node.is_trainable():
196 | remaining_len = (len(node._get_train_seq())
197 | - self._pretrained_phase[i_node])
198 | train_seq += ([(get_train_function(i_node, node),
199 | node.stop_training)] * remaining_len)
200 |
201 | # try fix the dimension of the internal nodes and the FlowNode
202 | # after the last node has been trained
203 | def _get_stop_training_wrapper(self, node, func):
204 | def _stop_training_wrapper(*args, **kwargs):
205 | func(*args, **kwargs)
206 | self._fix_nodes_dimensions()
207 | return _stop_training_wrapper
208 |
209 | if train_seq:
210 | train_seq[-1] = (train_seq[-1][0],
211 | _get_stop_training_wrapper(self, self._flow[-1], train_seq[-1][1]))
212 | return train_seq
213 |
214 |
215 | def train(self, x, *args, **kwargs):
216 | print "MergableFlowNode train "+str(self._train_phase)
217 | super(MergableFlowNode, self).train(x, *args, **kwargs)
218 |
219 |
220 | def skip_training(self):
221 | """
222 | Close the current phase without actually performing it.
223 | This is useful if we know that it was already performed for the
224 | current inner node from a different reference.
225 | Only thing left to do is inform the surrounding flow.
226 | This method does it.
227 | """
228 | #node._train_seq[self._train_phase][1](*args, **kwargs)
229 | self._train_phase += 1
230 | self._train_phase_started = False
231 | # check if we have some training phase left
232 | if self.get_remaining_train_phase() == 0:
233 | self._training = False
234 |
235 | class MergeLayer(Layer):
236 | """Layer with an additional merging-phase that merges all nodes in the layer
237 | after training. Merging is done by a given merger, which is itself a node.
238 | After merging, the merger will be used for execution in a CloneLayer-like
239 | fashion.
240 |
241 | The idea behind MergeLayer is a hybrid of ordinary layer and CloneLayer.
242 | The goal in this design is to use separate nodes in the train-phase,
243 | while using only a single node for execution. The difference to CloneLayer
244 | is that in MergeLayer, a different algorithm can be used for combining
245 | horizontally parallel data chunks than for combining time-sequent data
246 | chunks. The latter ones are combined by the nodes in the usual train-phase.
247 | In Contrast to CloneLayer, MergeLayer allows to control how horizontal merging
248 | of the data works. While CloneLayer would push this data into the very same
249 | train method like the time-sequent chunks, MergeLayer uses a merger to combine
250 | horizontal data.
251 | """
252 |
253 | def __init__(self, merger, nodes, call_stop_training = False, dtype=None):
254 | """Setup the layer with the given list of nodes.
255 |
256 | Keyword arguments:
257 | merger -- Merger to be used.
258 | nodes -- List of the nodes to be used.
259 | """
260 | super(MergeLayer, self).__init__(nodes, dtype=dtype)
261 | self.merger = merger
262 | self.call_stop_training = call_stop_training
263 |
264 | def _stop_training(self, *args, **kwargs):
265 | """Stop training of the internal node."""
266 | if self.call_stop_training:
267 | super(MergeLayer, self)._stop_training()
268 | for node in self.nodes:
269 | self.merger.merge(node)
270 | self.merger.stop_merging()
271 | self.trained_nodes = self.nodes
272 | self.nodes = (self.merger,) * len(self.trained_nodes)
273 | if self.output_dim is None:
274 | self.output_dim = self._get_output_dim_from_nodes()
275 |
276 | class FlowMergeLayer(Layer):
277 | """Like MergeLayer, but more aware of multiple training phases.
278 | The inner nodes must be MergableFlows.
279 | Given that these need multiple training phases, AdvancedMergeLayer
280 | can treat some phases like MergeLayer, others like CloneLayer or ordinary Layer.
281 | If the flows contain non-merge steps with a reused
282 | Node-reference, this reference should not have stop_training called
283 | several times.
284 |
285 | Warning: Nesting of FlowMergers is not supported!
286 |
287 | If the index of a training phase appears in merge_indices, it calls:
288 |
289 | super(FlowMergeLayer).stop_training
290 | for node in self.nodes:
291 | self.merger.merge(node)
292 | self.merger.stop_merging()
293 |
294 | If the index appears in no list, it calls:
295 |
296 | super(FlowMergeLayer).stop_training()
297 |
298 | If the index appears in clone_indices, it calls:
299 |
300 | merger.stop_training()
301 | for node in self.nodes:
302 | #close training phase without calling stop_training:
303 | node.skip_training()
304 |
305 | It expects merger to contain the same node reference as the usual inner
306 | nodes. (CloneLayer-like fashion but taking merger as the reference handle).
307 | Expecting the inner nodes of the layer to be different flows containing the
308 | same node reference, the flow's training phase is closed without actually
309 | performing it (it was already performed by merger.stop_training).
310 |
311 | The lists are expected to contain the indices in strictly ascending order.
312 |
313 | Note that unlike merge layer, the call of layer-wide stop_training can't be
314 | avoided, since the training phase index must increase. So AdvancedMergeLayer
315 | has no parameter call_stop_training.
316 |
317 | todo: Maybe find better solution for merge/train sequence
318 | Use mode list instead of index list: [merge, clone, train, train, clone, merge,...]
319 | """
320 |
321 | def __init__(self, merger, nodes, merge_indices, clone_indices, dtype=None):
322 | """Setup the layer with the given list of nodes.
323 |
324 | Keyword arguments:
325 | merger -- Merger to be used. Must be a FlowMerger.
326 | nodes -- List of the nodes to be used. These must be MergableFlows.
327 | """
328 | #for node in nodes:
329 | #node.set_merger(merger)
330 | super(FlowMergeLayer, self).__init__(nodes, dtype)
331 | self.merger = merger
332 | self.merge_indices = merge_indices
333 | self.clone_indices = clone_indices
334 | #offsets:
335 | self.current_merge = 0
336 | self.current_clone = 0
337 | self.current_train = 0
338 |
339 | def _stop_training(self, *args, **kwargs):
340 | """Stop training of the internal node."""
341 | #print "FlowMerger stop training"
342 | if self.current_merge < len(self.merge_indices) and self.current_train == self.merge_indices[self.current_merge]:
343 | #print "FlowMerger.super stop training..."
344 | #super(FlowMergeLayer, self).stop_training()
345 | for node in self.nodes:
346 | node.stop_training()
347 | #print "FlowMerger.super stop training done"
348 | #self._train_phase
349 | for node in self.nodes:
350 | self.merger.merge(node)
351 | self.merger.stop_merging()
352 | #print "stop merging done...."
353 | self.current_merge += 1
354 | elif self.current_clone < len(self.clone_indices) and self.current_train == self.clone_indices[self.current_clone]:
355 | #self.merger.force_stop_training()
356 | #self.merger.skip_training()
357 | self.nodes[0].stop_training()
358 | for i in range(1, len(self.nodes)):
359 | self.nodes[i].skip_training()
360 | self.current_clone += 1
361 | else:
362 | #I currently see no usecase for this. Maybe remove it...
363 | super(FlowMergeLayer, self).stop_training()
364 | self.current_train += 1
365 | #print "Close Layer training? "+str(self.get_remaining_train_phase())
366 | if self.get_remaining_train_phase() == 1: #This already indicates the last phase, because stop_training decrements it after the call to _stop_training
367 | self.trained_nodes = self.nodes
368 | self.nodes = (self.merger,) * len(self.trained_nodes)
369 | #print "Inserted merger as nodes"
370 | if self.output_dim is None:
371 | self.output_dim = self._get_output_dim_from_nodes()
372 |
--------------------------------------------------------------------------------
/src/cache_node.py:
--------------------------------------------------------------------------------
1 | #
2 | #
3 | # Copyright of GNUPFA:
4 | # Copyright (c) 2013, 2014 Institut fuer Neuroinformatik,
5 | # Ruhr-Universitaet Bochum, Germany. All rights reserved.
6 | #
7 | #
8 | # This file is part of GNUPFA.
9 | #
10 | # GNUPFA is free software: you can redistribute it and/or modify
11 | # it under the terms of the GNU General Public License as published by
12 | # the Free Software Foundation, either version 3 of the License, or
13 | # (at your option) any later version.
14 | #
15 | # GNUPFA is distributed in the hope that it will be useful,
16 | # but WITHOUT ANY WARRANTY; without even the implied warranty of
17 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 | # GNU General Public License for more details.
19 | #
20 | # You should have received a copy of the GNU General Public License
21 | # along with GNUPFA. If not, see .
22 | #
23 | #
24 | # Linking this library statically or dynamically with other modules is
25 | # making a combined work based on this library. Thus, the terms and
26 | # conditions of the GNU General Public License cover the whole
27 | # combination.
28 | #
29 |
30 |
31 | '''
32 | Created on Dec 21, 2013
33 |
34 | @author: Stefan Richthofer
35 | '''
36 | from mdp import Node
37 | from mdp import TrainingException
38 | from mdp.hinet import FlowNode
39 | from mdp.hinet.layer import CloneLayer
40 | import numpy as np
41 |
42 | class CacheNode(Node):
43 |
44 | def __init__(self, filename, cacheSize = -1, input_dim=None, output_dim=None, dtype=None):
45 | super(CacheNode, self).__init__(input_dim, output_dim, dtype)
46 | self.cacheName = filename
47 | self.cacheSize = cacheSize
48 | self.cache = None #np.memmap(filename, dtype='float32', mode='w+', shape=(3,4))
49 | self.cachePos = 0
50 | self.defaultOutputLength = 0
51 | self.cacheLength = 0
52 |
53 | def reshape(self, shape):
54 | if self.cache is None:
55 | return
56 | s = self.cache.dtype.itemsize
57 | for n in shape:
58 | s *= n
59 | self.cache._mmap.resize(s)
60 | del self.cache
61 | self.cache = np.memmap(self.cacheName, dtype=self.dtype, mode='readwrite', shape=shape)
62 |
63 | #Used to fill the cache
64 | def _train(self, x):
65 | # print self.dtype
66 | if len(x) > self.defaultOutputLength:
67 | self.defaultOutputLength = len(x)
68 | self.cacheLength += len(x)
69 | if self.cache is None:
70 | if self.cacheSize == -1:
71 | #self.cache = np.memmap(self.cacheName, dtype='float32', mode='w+', shape = x.shape)
72 | self.cache = np.memmap(self.cacheName, dtype=self.dtype, mode='w+', shape = x.shape)
73 | else:
74 | #self.cache = np.memmap(self.cacheName, dtype='float32', mode='w+', shape = (self.cacheSize, len(x[0])))
75 | self.cache = np.memmap(self.cacheName, dtype=self.dtype, mode='w+', shape = (self.cacheSize, len(x[0])))
76 | elif self.cacheSize == -1:
77 | self.reshape((self.cache.shape[0]+len(x), len(x[0])))
78 | # print x[0][0].dtype.itemsize
79 | # print self.cache._mmap.size()
80 | # #self.cache._mmap.resize( (self.cache.shape[0]+len(x), len(x[0])) )
81 | # print self.cache.shape
82 | # newShape = (self.cache.shape[0]+len(x), len(x[0]))
83 | # memmap_resize( newShape, self.cache )
84 | # del self.cache
85 | # self.cache = np.memmap(self.cacheName, dtype=self.dtype, mode='w+', shape = newShape)
86 | # print "new size: "+str(self.cache._mmap.size())
87 | # print self.cache.reshape(newShape)
88 | self.cache[self.cachePos:self.cachePos+len(x)] = x
89 | # print self.cache._mmap.size()
90 | # print self.cache[0][0]
91 | # print self.cache[0][0].dtype.itemsize
92 | # print "---"
93 | self.cachePos += len(x)
94 |
95 | def _stop_training(self):
96 | self.cachePos = 0
97 |
98 | def read(self, off, read_len = -1):
99 | if self.cache is None:
100 | raise TrainingException("CacheNode was not filled (i.e. trained).")
101 | if off >= self.cacheLength:
102 | return None
103 | l = read_len
104 | if l == -1:
105 | l = self.defaultOutputLength
106 | if off+l > self.cacheLength:
107 | l = self.cacheLength-off
108 | return self.cache[off:off+l]
109 |
110 | def _execute(self, x):
111 | er = self.read(self.cachePos, len(x))
112 | self.cachePos += len(x)
113 | return er
114 |
115 | class ReorderingCacheNode(CacheNode):
116 |
117 | def __init__(self, fieldsize, filename, cacheSize = -1, input_dim=None, output_dim=None, dtype=None):
118 | super(ReorderingCacheNode, self).__init__(filename, cacheSize, input_dim, fieldsize, dtype)
119 | self.fieldsize = fieldsize
120 |
121 | def _train(self, x):
122 | self.cacheLength += len(x)*len(x[0])/self.fieldsize
123 | if len(x)*len(x[0])/self.fieldsize > self.defaultOutputLength:
124 | self.defaultOutputLength = len(x)*len(x[0])/self.fieldsize
125 | if self.cache is None:
126 | if self.cacheSize == -1:
127 | self.cache = np.memmap(self.cacheName, dtype=self.dtype, mode='w+', shape = x.shape)
128 | else:
129 | self.cache = np.memmap(self.cacheName, dtype=self.dtype, mode='w+', shape = (self.cacheSize, len(x[0])))
130 | elif self.cacheSize == -1:
131 | self.reshape( (self.cache.shape[0]+len(x), len(x[0])) )
132 | self.cache[self.cachePos:self.cachePos+len(x)] = x
133 | self.cachePos += len(x)
134 |
135 | def read(self, off, read_len = -1):
136 | if self.cache is None:
137 | raise TrainingException("CacheNode was not filled (i.e. trained).")
138 | if off >= self.cacheLength:
139 | return None
140 | l = read_len
141 | if l == -1:
142 | l = self.defaultOutputLength
143 | if off+l > self.cacheLength:
144 | l = self.cacheLength-off
145 | off1 = off % len(self.cache)
146 | if off1+l > len(self.cache):
147 | l = len(self.cache)-off1
148 | off2 = (off / len(self.cache)) * self.fieldsize
149 | return self.cache[off1:off1+l, off2:off2+self.fieldsize]
150 |
151 | def read_in_order(self, off, read_len = -1):
152 | return super(ReorderingCacheNode, self).read(off, read_len)
153 |
154 | def _execute(self, x):
155 | er = self.read(self.cachePos, len(x)*len(x[0])/self.fieldsize)
156 | self.cachePos += len(x)
157 | return er
158 |
159 | def contains(array, value):
160 | for i in range(len(array)):
161 | if array[i] == value:
162 | return True
163 | return False
164 |
165 | def buildCaches(cacheIndices, length, basename, cacheSize = -1):
166 | caches = []
167 | for i in range(length):
168 | if (cacheIndices is None) or contains(cacheIndices, i):
169 | caches.append(CacheNode(basename+"_"+str(i), cacheSize))
170 | else:
171 | caches.append(None)
172 | return caches
173 |
174 | class CachingFlowNode(FlowNode):
175 |
176 | def __init__(self, flow, caches = [], cacheSize = -1, input_dim=None, output_dim=None, dtype=None):
177 | super(CachingFlowNode, self).__init__(flow, input_dim, output_dim, dtype)
178 | self._train_seq_cache = None
179 | self.caches = caches
180 |
181 | def _check_nodes_consistency(self, flow = None):
182 | """Check the dimension consistency of a list of nodes."""
183 | if flow is None:
184 | flow = self._flow.flow
185 | for i in range(1, len(flow)):
186 | out = flow[i-1].output_dim
187 | if not self.caches[i] is None:
188 | out = self.caches[i].output_dim
189 | inp = flow[i].input_dim
190 | self._flow._check_dimension_consistency(out, inp)
191 |
192 | def _fix_nodes_dimensions(self):
193 | """Try to fix the dimensions of the internal nodes."""
194 | if len(self._flow) > 1:
195 | prev_node = self._flow[0]
196 | for node in self._flow[1:]:
197 | if node.input_dim is None:
198 | node.input_dim = prev_node.output_dim
199 | prev_node = node
200 | self._check_nodes_consistency()
201 | if self._flow[-1].output_dim is not None:
202 | # additional checks are performed here
203 | self.output_dim = self._flow[-1].output_dim
204 |
205 | def find_first_cache(self):#, _i_node):
206 | #print "n: "+str(_i_node)
207 | for i in range(len(self.caches)):#_i_node+1):
208 | if not (self.caches[i] is None):
209 | return i
210 | return -1
211 |
212 | def _get_train_seq(self):
213 | """Return a training sequence containing all training phases."""
214 | def get_train_function(_i_node, _node):
215 | # This internal function is needed to channel the data through
216 | # the nodes in front of the current nodes.
217 | # using nested scopes here instead of default args, see pep-0227
218 | #cachePos = self.find_nearest_cache(_i_node)
219 | def _train(x, *args, **kwargs):
220 | if i_node > 0:
221 | _node.train(self._flow.execute(x, nodenr=_i_node-1), *args, **kwargs)
222 | # if cachePos == _i_node:
223 | # _node.train(self.caches[cachePos].execute(x), *args, **kwargs)
224 | # elif cachePos == -1:
225 | # _node.train(self._flow.execute(x, nodenr=_i_node-1), *args, **kwargs)
226 | # else:
227 | # y = self.caches[cachePos].execute(x)
228 | # for i in range(cachePos, _i_node-1):
229 | # y = self._flow.flow[i].execute(y)
230 | # _node.train(y, *args, **kwargs)
231 | else:
232 | _node.train(x, *args, **kwargs)
233 | return _train
234 |
235 | train_seq = []
236 | startCache = self.find_first_cache()
237 | if startCache == -1:
238 | startCache = len(self._flow.flow)
239 |
240 | #for i_node, node in enumerate(self._flow):
241 | for i_node in range(startCache):
242 | node = self._flow[i_node]
243 | if node.is_trainable():
244 | remaining_len = (len(node._get_train_seq()) - self._pretrained_phase[i_node])
245 | train_seq += ([(get_train_function(i_node, node), node.stop_training)] * remaining_len)
246 |
247 | if not self.caches[startCache] is None:
248 | def _train_tail(x, *args, **kwargs):
249 | if startCache > 0:
250 | self.caches[startCache].train(self._flow.execute(x, nodenr=startCache-1), *args, **kwargs)
251 | else:
252 | self.caches[startCache].train(x)
253 |
254 | def _stop_train_tail(*args, **kwargs):
255 | startCache = self.find_first_cache()
256 | if startCache == -1:
257 | startCache = len(self._flow.flow)
258 | currentTrainNode = startCache
259 | while currentTrainNode < len(self._flow.flow):
260 | #trainNode = self._flow.flow[currentTrainNode]
261 | if self._flow.flow[currentTrainNode].is_trainable():
262 | remaining_len = (len(self._flow.flow[currentTrainNode]._get_train_seq()) - self._pretrained_phase[i_node])
263 | for i in range(remaining_len):
264 | readOff = 0
265 | trainData = self.caches[startCache].read(readOff)
266 | while not trainData is None:
267 | for j in range(startCache, currentTrainNode):
268 | trainData = self._flow.flow[j].execute(trainData)
269 | self._flow.flow[currentTrainNode].train(trainData, *args, **kwargs)
270 | readOff += len(trainData)
271 | trainData = self.caches[startCache].read(readOff)
272 | self._flow.flow[currentTrainNode].stop_training(*args, **kwargs)
273 | currentTrainNode += 1
274 | if len(self.caches) < currentTrainNode and not self.caches[currentTrainNode] is None:
275 | readOff = 0
276 | trainData = self.caches[startCache].read(readOff)
277 | while not trainData is None:
278 | for j in range(startCache, currentTrainNode):
279 | trainData = self._flow.flow[j].execute(trainData)
280 | self.caches[currentTrainNode].train(trainData, *args, **kwargs)
281 | readOff += len(trainData)
282 | trainData = self.caches[startCache].read(readOff)
283 | self.caches[currentTrainNode].stop_training(*args, **kwargs)
284 | startCache = currentTrainNode
285 |
286 | self._fix_nodes_dimensions()
287 |
288 | train_seq.append((_train_tail, _stop_train_tail))
289 |
290 | else:
291 |
292 | # try fix the dimension of the internal nodes and the FlowNode
293 | # after the last node has been trained
294 | def _get_stop_training_wrapper(self, node, func):
295 | def _stop_training_wrapper(*args, **kwargs):
296 | func(*args, **kwargs)
297 | self._fix_nodes_dimensions()
298 | return _stop_training_wrapper
299 |
300 | if train_seq:
301 | train_seq[-1] = (train_seq[-1][0], _get_stop_training_wrapper(self, self._flow[-1], train_seq[-1][1]))
302 |
303 | return train_seq
304 |
305 | class CacheCloneLayer(CloneLayer):#, CacheNode):
306 | #multi inheritance fails because mro would delgate __init__ call in Layer to CacheNode instead of Node.
307 |
308 | """A CloneLayer variant that caches all training data to disc and
309 | then trains the backing node with reordered data. The data is reordered
310 | such that each field is completely trained (i.e. all chunks, if multiple
311 | chunks are used) before the next field starts.
312 | """
313 |
314 | def __init__(self, cacheName, node, n_nodes=1, dtype=None):
315 | """Setup the layer with the given list of nodes.
316 |
317 | Keyword arguments:
318 | node -- Node to be cloned.
319 | n_nodes -- Number of repetitions/clones of the given node.
320 | """
321 |
322 | super(CacheCloneLayer, self).__init__(node=node, n_nodes=n_nodes, dtype=dtype)
323 | self.node = node # attribute for convenience
324 | self.cache = ReorderingCacheNode(node.input_dim, cacheName)
325 | self.cacheLength = 0
326 |
327 | def _get_train_seq(self):
328 | return [(self._train, self._stop_training)]
329 |
330 | def read(self, off, read_len = -1):
331 | return self.cache.read_in_order(off, read_len)
332 |
333 | # def _get_train_seq(self):
334 | # """Return the train sequence.
335 | #
336 | # The length is set by the node with maximum length.
337 | # """
338 | # max_train_length = 0
339 | # for node in self.nodes:
340 | # node_length = len(node._get_train_seq())
341 | # if node_length > max_train_length:
342 | # max_train_length = node_length
343 | # return ([[self._train, self._stop_training]] * max_train_length)
344 |
345 | def _train(self, x, *args, **kwargs):
346 | self.cacheLength += len(x)
347 | self.cache.train(x)
348 | # """Perform single training step by training the internal nodes."""
349 | # start_index = 0
350 | # stop_index = 0
351 | # for node in self.nodes:
352 | # start_index = stop_index
353 | # stop_index += node.input_dim
354 | # if node.is_training():
355 | # node.train(x[:, start_index : stop_index], *args, **kwargs)
356 |
357 | # def _stop_training(self, *args, **kwargs):
358 | # """Stop training of the internal nodes."""
359 | # for node in self.nodes:
360 | # if node.is_training():
361 | # node.stop_training(*args, **kwargs)
362 | # if self.output_dim is None:
363 | # self.output_dim = self._get_output_dim_from_nodes()
364 |
365 | def _stop_training(self, *args, **kwargs):
366 | """Stop training of the internal node."""
367 | self.cache.stop_training()
368 | phases = self.node.get_remaining_train_phase()
369 | for i in range(phases):
370 | pos = 0
371 | while pos < self.cache.cacheLength:
372 | x = self.cache.read(pos)
373 | pos += len(x)
374 | self.node.train(x)
375 | self.node.stop_training(*args, **kwargs)
376 | #if self.node.is_training():
377 | # self.node.stop_training(*args, **kwargs)
378 | if self.output_dim is None:
379 | self.output_dim = self._get_output_dim_from_nodes()
380 |
--------------------------------------------------------------------------------
/src/PFANodeMDP.py:
--------------------------------------------------------------------------------
1 | #
2 | #
3 | # Copyright of GNUPFA:
4 | # Copyright (c) 2013, 2014 Institut fuer Neuroinformatik,
5 | # Ruhr-Universitaet Bochum, Germany. All rights reserved.
6 | #
7 | #
8 | # This file is part of GNUPFA.
9 | #
10 | # GNUPFA is free software: you can redistribute it and/or modify
11 | # it under the terms of the GNU General Public License as published by
12 | # the Free Software Foundation, either version 3 of the License, or
13 | # (at your option) any later version.
14 | #
15 | # GNUPFA is distributed in the hope that it will be useful,
16 | # but WITHOUT ANY WARRANTY; without even the implied warranty of
17 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 | # GNU General Public License for more details.
19 | #
20 | # You should have received a copy of the GNU General Public License
21 | # along with GNUPFA. If not, see .
22 | #
23 | #
24 | # Linking this library statically or dynamically with other modules is
25 | # making a combined work based on this library. Thus, the terms and
26 | # conditions of the GNU General Public License cover the whole
27 | # combination.
28 | #
29 |
30 |
31 | '''
32 | Created on Sep 9, 2013
33 |
34 | @author: Stefan Richthofer
35 | '''
36 | import mdp
37 | import numpy as np
38 | import PFACoreUtil as pfa
39 | import PFANodeMDPRefImp as ref
40 | from MergeLayer import Merger
41 |
42 | class PFANode(mdp.Node):
43 | '''redefined _init_ Method with p, k as arguments'''
44 | def __init__(self, p = 2, k = 0, affine = True, input_dim = None, output_dim = None, dtype = None):
45 | super(PFANode, self).__init__(input_dim = input_dim, output_dim = output_dim, dtype = dtype)
46 | self.p = p
47 | self.k = k
48 | self.l = None
49 | self.affine = affine
50 | self.evThreshold = 0.0000000001
51 | self.sindex = -1
52 | self.maxindex = -1
53 | self.layerIndex = -1
54 | self.layerMax = -1
55 |
56 | '''Node is trainable'''
57 | def is_trainable(self):
58 | return True
59 |
60 | '''Saves relevant information of the data for further processing in stop_training.'''
61 | def _train(self, x):
62 | #print "PFA train "+str(self.sindex)+" of "+str(self.maxindex)+" layer "+str(self.layerIndex)+" of "+str(self.layerMax)
63 | n = self.get_input_dim()
64 | x2 = x
65 | if not n is None:
66 | x2 = x.T[:n].T
67 | if self.l is None:
68 | # self.data = x2
69 | self.startData = x2[:self.p+self.k]
70 | self.l = 1.0*len(x2)
71 | self.mean = x2.mean(0)
72 | #self.secondMoment = np.dot(x2.T, x2)/(1.0*len(x2))
73 | data_pk = x2[self.p+self.k:]
74 | self.corList = [np.dot(data_pk.T, data_pk)]
75 | for i in range(1, self.k+self.p+1):
76 | self.corList.append(np.dot(data_pk.T, x2[self.p+self.k-i:-i]))
77 | else:
78 | #Keeping always the averaged version instead of just accumulating here and finally dividing through
79 | #the whole length in _stop_training may be better for large data, since the magnitude of the matrix
80 | #entries is better kept in a sane range
81 | # self.data = np.vstack([self.data, x2])
82 | self.mean = (self.l/(self.l+len(x2))) * self.mean + (len(x2)/(self.l+len(x2))) * x2.mean(0)
83 | #self.secondMoment = (self.l/(self.l+len(x2))) * self.secondMoment + (1.0/(self.l+len(x2))) * np.dot(x2.T, x2)
84 | #self.secondMoment = (self.l*self.secondMoment+np.dot(x2.T, x2))/(self.l+len(x2))
85 | self.corList[0] += np.dot(x2.T, x2)
86 | #self.corList[1] += np.dot(x2[1:].T, x2[:-1])+np.dot(x2[:1].T, self.endData[-1:])
87 | for i in range(1, self.k+self.p+1):
88 | self.corList[i] += np.dot(x2[i:].T, x2[:-i])+np.dot(x2[:i].T, self.endData[-i:])
89 | self.l += 1.0*len(x2)
90 | self.endData = x2[-self.p-self.k:]
91 |
92 | def _prepare_start_end_cor(self):
93 | start_cor = []
94 | end_cor = []
95 | for i in range(len(self.startData)):
96 | startLine = []
97 | endLine = []
98 | for j in range(len(self.startData)):
99 | if j < i:
100 | startLine.append(start_cor[j][i].T)
101 | endLine.append(end_cor[j][i].T)
102 | else:
103 | startLine.append(np.outer(self.startData[i], self.startData[j]))
104 | endLine.append(np.outer(self.endData[i], self.endData[j]))
105 | start_cor.append(startLine)
106 | end_cor.append(endLine)
107 | self.start_cor = start_cor
108 | self.end_cor = end_cor
109 |
110 | #This is seperate from _prepare_start_end_cor because merging would be done between these
111 | #Note that it requires that x_cor[i][j] is a reference to x_cor[j][i].T
112 | #This fact must be preserved during merging!
113 | def _start_end_cor_clear_mean(self):
114 | M = np.outer(self.mean, self.mean)
115 | for i in range(len(self.start_cor)):
116 | for j in range(i, len(self.start_cor[i])):
117 | self.start_cor[i][j] += -np.outer(self.startData[i], self.mean)-np.outer(self.mean, self.startData[j])+M
118 | self.end_cor[i][j] += -np.outer(self.endData[i], self.mean)-np.outer(self.mean, self.endData[j])+M
119 |
120 | def _stop_training(self):
121 | #print "PFA stop_training "+str(self.sindex)+" of "+str(self.maxindex)+" layer "+str(self.layerIndex)+" of "+str(self.layerMax)
122 | if self.endData is None:
123 | raise mdp.TrainingException("train was never called")
124 | self._prepare_start_end_cor()
125 | self._start_end_cor_clear_mean()
126 | self.calc_PFA()
127 |
128 | def calc_PFA(self):
129 | #print "PFA calc_PFA "+str(self.sindex)+" of "+str(self.maxindex)+" layer "+str(self.layerIndex)+" of "+str(self.layerMax)
130 | r = self.get_output_dim()
131 | if r is None:
132 | r = len(self.startData.T)
133 |
134 | #meanRef, SRef, zRef = pfa.calcSpheringParametersAndDataRefImp(self.data)#, threshold = 0.0000001, offset = 0, length = -1, besselsCorrection = 0)
135 |
136 | # print "secondMomentTest:"
137 | # print self.secondMoment*self.l
138 | # print (self.corList[0]+np.dot(self.startData.T, self.startData))
139 | # print "----------------"
140 | # print "chunkTest:"
141 | # print self.corList
142 |
143 | S = pfa.calcSpheringMatrixFromMeanAndSecondMoment(self.mean, (self.corList[0]+np.dot(self.startData.T, self.startData))/self.l, self.l, threshold = self.evThreshold)#, besselsCorrection = 0)
144 | self.S = S
145 | if S.shape[1] < r:
146 | r = S.shape[1]
147 | #this creates an inconsistent output_dim.
148 | #However, it does not matter, if one sets output_dim to an appropriate value
149 | #from the beginning.
150 | #self.output_dim = r
151 | # print self.corList[0]
152 | # print np.dot(self.startData.T, self.startData)
153 | # corList00 = self.start_cor[0][0]
154 | # for r2 in range(1, len(self.startData)):
155 | # corList00 += self.start_cor[r2][r2]
156 | # print corList00
157 | # print "S"
158 | # print S.shape
159 |
160 | meanList0 = [self.mean*self.l-self.startData.mean(0)*(self.p+self.k)]
161 | for i in range(1, self.k+self.p+1):
162 | meanList0.append(meanList0[-1]-self.endData[-i]+self.startData[-i])
163 | M = np.outer(self.mean*(self.l-self.p-self.k), self.mean)-np.outer(meanList0[0], self.mean)
164 | mns = np.outer(np.ones(self.p+self.k), self.mean)
165 | startZ = np.dot(self.startData-mns, S)
166 | endZ = np.dot(self.endData-mns, S)
167 | #startZ0 = self.startData-mns
168 | #endZ0 = self.endData-mns
169 |
170 | #print np.outer(startZ0[1], startZ0[2])
171 | # print np.outer(self.endData[0]-self.mean, self.endData[1]-self.mean)
172 | # print np.outer(self.endData[0], self.endData[1])-np.outer(self.mean, self.endData[1])-np.outer(self.endData[0], self.mean)+np.outer(self.mean, self.mean)
173 | # print np.outer(self.endData[0], self.endData[1])
174 | #print self.start_cor[1][2]
175 |
176 | # zRef = np.dot(self.data-np.outer(np.ones(len(self.data)), self.mean), S)
177 | # z_pk = zRef[self.p+self.k:]
178 | # z_p = zRef[self.p:]
179 | corList = []#np.dot(z_pk.T, z_pk)]
180 | corList0 = []
181 | for i in range(0, self.k+self.p+1):
182 | corList.append(np.dot(S.T, np.dot(self.corList[i] - np.outer(self.mean, meanList0[i]) + M, S)))
183 | corList0.append(self.corList[i] - np.outer(self.mean, meanList0[i]) + M)
184 |
185 | zetaList = []
186 | lastLine = []
187 | for i in range(self.p):
188 | zetaLine = []
189 | for j in range(self.p):
190 | if j < i:
191 | zetaLine.append(zetaList[j][i].T)
192 | else:
193 | if i == 0:
194 | #zetaLine.append(corList[j]-np.outer(endZ[-1], endZ[-1-j])+np.outer(startZ[self.p+self.k-1], startZ[self.p+self.k-1-j]))
195 | #zetaLine.append(np.dot(S.T, np.dot(corList0[j]-np.outer(endZ0[-1], endZ0[-1-j])+np.outer(startZ0[self.p+self.k-1], startZ0[self.p+self.k-1-j]), S)))
196 | #zetaLine.append(np.dot(S.T, np.dot(corList0[j]-np.outer(endZ0[-1], endZ0[-1-j])+np.outer(startZ0[-1], startZ0[-1-j]), S)))
197 | zetaLine.append(np.dot(S.T, np.dot(corList0[j]-self.end_cor[-1][-1-j]+self.start_cor[-1][-1-j], S)))
198 | else:
199 | #zetaLine.append(lastLine[j-1]-np.outer(endZ[-1-i], endZ[-1-j])+np.outer(startZ[self.p+self.k-1-i], startZ[self.p+self.k-1-j]))
200 | #zetaLine.append(lastLine[j-1]+np.dot(S.T, np.dot(-np.outer(endZ0[-1-i], endZ0[-1-j])+np.outer(startZ0[self.p+self.k-1-i], startZ0[self.p+self.k-1-j]), S)))
201 | #zetaLine.append(lastLine[j-1]+np.dot(S.T, np.dot(-np.outer(endZ0[-1-i], endZ0[-1-j])+np.outer(startZ0[-1-i], startZ0[-1-j]), S)))
202 | zetaLine.append(lastLine[j-1]+np.dot(S.T, np.dot(-self.end_cor[-1-i][-1-j]+self.start_cor[-1-i][-1-j], S)))
203 | zetaList.append(zetaLine)
204 | lastLine = zetaLine
205 | for i in range(self.p):
206 | zetaList[i] = np.hstack(zetaList[i])
207 |
208 | corList_p = None
209 | zetaList_p = None
210 | lastLine_p = None
211 | if self.k > 0:
212 | corList_p = [corList[0]+np.dot(startZ[self.p:].T, startZ[self.p:])]#[np.dot(z_p.T, z_p)]
213 | #corList_p0 = [corList0[0]+np.dot(startZ0[self.p:].T, startZ0[self.p:])]
214 | # R = np.array(self.start_cor[self.p][self.p])
215 | # for r2 in range(self.p+1, len(self.startData)):
216 | # R += self.start_cor[r2][r2]
217 | # corList_p0 = [corList0[0]+R]
218 | corList_p0 = [corList0[0]+self.start_cor[self.p][self.p]]
219 | for r2 in range(self.p+1, len(self.startData)):
220 | corList_p0[0] += self.start_cor[r2][r2]
221 |
222 | # corList_p00 = [np.dot(z_p.T, z_p)]
223 | for i in range(1, self.p+1):
224 | # corList_p00.append(np.dot(z_p.T, zRef[self.p-i:-i]))
225 | corList_p.append(corList[i]+np.dot(startZ[self.p:].T, startZ[self.p-i:-i]))
226 | #corList_p0.append(corList0[i]+np.dot(startZ0[self.p:].T, startZ0[self.p-i:-i]))
227 | # R = np.array(self.start_cor[self.p][self.p-i])
228 | # for r2 in range(self.p+1, len(self.startData)):
229 | # R += self.start_cor[r2][r2-i]
230 | # corList_p0.append(corList0[i]+R)
231 | corList_p0.append(corList0[i]+self.start_cor[self.p][self.p-i])
232 | for r2 in range(self.p+1, len(self.startData)):
233 | corList_p0[i] += self.start_cor[r2][r2-i]
234 |
235 | # for r in range(self.p+1, len(startZ0)):
236 | # corList_p0[i] += self.start_cor[r][r-i]
237 | # print "--------------"
238 | # print np.dot(startZ0[self.p:].T, startZ0[self.p-i:-i])
239 | # R = np.outer(startZ0[self.p], startZ0[self.p-i])
240 | # for ir in range(self.p+1, len(startZ0)):
241 | # R += np.outer(startZ0[ir], startZ0[ir-i])
242 | # print R
243 | zetaList_p = []
244 | lastLine_p = []
245 | for i in range(self.p):
246 | zetaLine_p = []
247 | for j in range(self.p):
248 | if j < i:
249 | zetaLine_p.append(zetaList_p[j][i].T)
250 | else:
251 | if i == 0:
252 | #zetaLine_p.append(corList_p[j]-np.outer(endZ[-1], endZ[-1-j])+np.outer(startZ[self.p-1], startZ[self.p-1-j]))
253 | #zetaLine_p.append(np.dot(S.T, np.dot(corList_p0[j]-np.outer(endZ0[-1], endZ0[-1-j])+np.outer(startZ0[self.p-1], startZ0[self.p-1-j]), S)))
254 | zetaLine_p.append(np.dot(S.T, np.dot(corList_p0[j]-self.end_cor[-1][-1-j]+self.start_cor[self.p-1][self.p-1-j], S)))
255 | else:
256 | #zetaLine_p.append(lastLine_p[j-1]-np.outer(endZ[-1-i], endZ[-1-j])+np.outer(startZ[self.p-1-i], startZ[self.p-1-j]))
257 | #zetaLine_p.append(lastLine_p[j-1]+np.dot(S.T, np.dot(-np.outer(endZ0[-1-i], endZ0[-1-j])+np.outer(startZ0[self.p-1-i], startZ0[self.p-1-j]), S)))
258 | zetaLine_p.append(lastLine_p[j-1]+np.dot(S.T, np.dot(-self.end_cor[-1-i][-1-j]+self.start_cor[self.p-1-i][self.p-1-j], S)))
259 | zetaList_p.append(zetaLine_p)
260 | lastLine_p = zetaLine_p
261 | for i in range(self.p):
262 | zetaList_p[i] = np.hstack(zetaList_p[i])
263 |
264 | if not self.affine:
265 | zZ = np.hstack(corList[1:self.p+1])
266 | ZZ = np.vstack(zetaList)
267 | W = None
268 | if self.k == 0:
269 | ZZI = pfa.invertByProjectionRefImp(ZZ, self.evThreshold)
270 | W = np.dot(zZ, ZZI)
271 | else:
272 | zZ_p = np.hstack(corList_p[1:self.p+1])
273 | ZZ_p = np.vstack(zetaList_p)
274 | ZZ_pI = pfa.invertByProjectionRefImp(ZZ_p, self.evThreshold)
275 | W = np.dot(zZ_p, ZZ_pI)
276 | self.W0 = W
277 | # WRef = pfa.calcRegressionCoeffRefImp(zRef, self.p, self.evThreshold)
278 | # XRef = pfa.calcErrorCoeffConstLenRefImp(zRef, WRef, self.k)
279 | #X = pfa.calcErrorCoeffConstLenFromCorrelations2(W, zZ, ZZ, lastLine, corList, startZ, endZ, S, self.start_cor, self.end_cor, self.k)
280 | X = pfa.calcErrorCoeffConstLenFromCorrelations2(W, zZ, ZZ, lastLine, corList, S, self.start_cor, self.end_cor, self.k)
281 | #print "xxxx"
282 | #X = pfa.calcErrorCoeffConstLenFromCorrelations2(W, zZ, ZZ, lastLine, corList, S, self.start_cor, self.end_cor, self.k)
283 |
284 | # print "WCompare:"
285 | # print WRef
286 | # print W
287 | # print "XCompare:"
288 | # print XRef
289 | # print X
290 | # print"------"
291 | self.X = X #needed only for debugging
292 | Ar = pfa.calcExtractionForErrorCovRefImp(X, r)
293 | #reduced = np.dot(self.data, self.Ar.T)
294 | # reduced = np.dot(zRef, Ar)
295 | # WRef = pfa.calcRegressionCoeffRefImp(reduced, self.p)
296 | # print "Reduction compare"
297 | A_ = np.kron(np.identity(self.p), Ar)
298 | if (self.k == 0):
299 | zZ = np.dot(Ar.T, np.dot(zZ, A_))
300 | ZZ = np.dot(A_.T, np.dot(ZZ, A_))
301 | ZZI = pfa.invertByProjectionRefImp(ZZ, self.evThreshold)
302 | W = np.dot(zZ, ZZI)
303 | else:
304 | zZ_p = np.dot(Ar.T, np.dot(zZ_p, A_))
305 | ZZ_p = np.dot(A_.T, np.dot(ZZ_p, A_))
306 | ZZ_pI = pfa.invertByProjectionRefImp(ZZ_p, self.evThreshold)
307 | W = np.dot(zZ_p, ZZ_pI)
308 | # print WRef
309 | # print W
310 | self.W = W
311 | # print "--------------------------"
312 | self.Ar = np.dot(S, Ar)
313 | else:
314 | meanList = [startZ.mean(0)*(-self.p-self.k)]
315 | for i in range(1, self.k+self.p+1):
316 | meanList.append(meanList[-1]-endZ[-i]+startZ[self.p+self.k-i])
317 | ml = np.hstack(meanList[1:self.p+1])
318 | zZ_c = np.vstack([np.hstack(corList[1:self.p+1]).T, meanList[0]]).T
319 | ZZ_c = np.vstack([np.vstack(zetaList), ml])
320 | ml1 = np.hstack([ml, [self.l-self.p-self.k]])
321 | ZZ_c = np.vstack([ZZ_c.T, ml1]).T
322 |
323 | W_c = None
324 | if self.k == 0:
325 | ZZ_cI = pfa.invertByProjectionRefImp(ZZ_c, self.evThreshold)
326 | W_c = np.dot(zZ_c, ZZ_cI)
327 | else:
328 | meanList_p = [startZ[:self.p].mean(0)*(-self.p)]
329 | for i in range(1, self.p+1):
330 | meanList_p.append(meanList_p[-1]-endZ[-i]+startZ[self.p-i])
331 | ml_p = np.hstack(meanList_p[1:self.p+1])
332 | zZ_pc = np.vstack([np.hstack(corList_p[1:self.p+1]).T, meanList_p[0]]).T
333 | ZZ_pc = np.vstack([np.vstack(zetaList_p), ml_p])
334 | ml1_p = np.hstack([ml_p, [self.l-self.p]])
335 | ZZ_pc = np.vstack([ZZ_pc.T, ml1_p]).T
336 |
337 | ZZ_pcI = pfa.invertByProjectionRefImp(ZZ_pc, self.evThreshold)
338 | W_c = np.dot(zZ_pc, ZZ_pcI)
339 | self.W0 = W_c
340 | #print self.start_cor
341 | # WRef = pfa.calcRegressionCoeffAffineRefImp(zRef, self.p, self.evThreshold)
342 | # XRef = pfa.calcErrorCoeffConstLenAffineRefImp(zRef, WRef, self.k)
343 | #X = pfa.calcErrorCoeffConstLenAffineFromCorrelations2(W_c, zZ_c, ZZ_c, lastLine, corList, meanList, self.l, startZ, endZ, self.k)
344 | X = pfa.calcErrorCoeffConstLenAffineFromCorrelations2(W_c, zZ_c, ZZ_c, lastLine, corList, meanList, self.l, S, self.start_cor, self.end_cor, self.k)
345 | # print "WCompare (Affine):"
346 | # print WRef
347 | # print W_c
348 | # print "XCompare (Affine):"
349 | # print XRef
350 | #print X
351 | # print"------"
352 | self.X = X #needed only for debugging
353 | Ar = pfa.calcExtractionForErrorCovRefImp(X, r)
354 | # reduced = np.dot(zRef, Ar)
355 | # WRef = pfa.calcRegressionCoeffAffineRefImp(reduced, self.p)
356 | # print "Reduction compare (Affine)"
357 | A_ = pfa.kronAffine(Ar, self.p)
358 | if self.k == 0:
359 | zZ_c = np.dot(Ar.T, np.dot(zZ_c, A_))
360 | ZZ_c = np.dot(A_.T, np.dot(ZZ_c, A_))
361 | ZZ_cI = pfa.invertByProjectionRefImp(ZZ_c, self.evThreshold)
362 | W_c = np.dot(zZ_c, ZZ_cI)
363 | else:
364 | zZ_pc = np.dot(Ar.T, np.dot(zZ_pc, A_))
365 | ZZ_pc = np.dot(A_.T, np.dot(ZZ_pc, A_))
366 | ZZ_pcI = pfa.invertByProjectionRefImp(ZZ_pc, self.evThreshold)
367 | W_c = np.dot(zZ_pc, ZZ_pcI)
368 | # print WRef
369 | # print W_c
370 | self.W = W_c
371 | # print "--------------------------"
372 | self.Ar = np.dot(S, Ar)
373 | #print "PFA_calc done "+str(self.sindex)+" of "+str(self.maxindex)+" layer "+str(self.layerIndex)+" of "+str(self.layerMax)
374 |
375 | def _execute(self, x):
376 | #print "PFA execute "+str(self.sindex)+" of "+str(self.maxindex)+" layer "+str(self.layerIndex)+" of "+str(self.layerMax)
377 | z0 = x-np.outer(np.ones(len(x)), self.mean)
378 | return np.dot(z0, self.Ar)
379 |
380 | class PFANodeLayerAware(PFANode):
381 | '''redefined _init_ Method with p, k as arguments '''
382 | def __init__(self, train_length, p = 2, k = 0, affine = True, input_dim = None, output_dim = None, dtype = None):
383 | super(PFANodeLayerAware, self).__init__(p, k, affine, input_dim, output_dim, dtype)
384 | self.train_length = train_length
385 | self.field_index = 0
386 | self.current_train_index = 0
387 | self.l_acc = 0
388 |
389 | def _merge_init(self):
390 | self.mean_acc = np.array(self.mean)
391 | self.l_acc += self.l
392 | self.corList_acc = []
393 | for i in range(len(self.corList)):
394 | self.corList_acc.append(np.array(self.corList[i]))
395 |
396 | self.startData_acc = np.array(self.startData)
397 | self.endData_acc = np.array(self.endData)
398 |
399 | self.start_cor_acc = []
400 | self.end_cor_acc = []
401 | for i in range(len(self.startData)):
402 | start_line = []
403 | end_line = []
404 | for j in range(len(self.startData)):
405 | if j >= i:
406 | start_line.append(np.array(self.start_cor[i][j]))
407 | end_line.append(np.array(self.end_cor[i][j]))
408 | else:
409 | start_line.append(self.start_cor_acc[j][i].T)
410 | end_line.append(self.end_cor_acc[j][i].T)
411 | self.start_cor_acc.append(start_line)
412 | self.end_cor_acc.append(end_line)
413 |
414 | def _merge(self):
415 | #print "PFA inline-merge "+str(node.sindex)+" of "+str(node.maxindex)+" layer "+str(node.layerIndex)+" of "+str(node.layerMax)
416 | self.mean_acc = (self.l_acc/(self.l_acc+self.l)) * self.mean_acc + (self.l/(self.l_acc+self.l)) * self.mean
417 | for i in range(len(self.corList_acc)):
418 | self.corList_acc[i] += self.corList[i]
419 |
420 | self.startData_acc += self.startData
421 | self.endData_acc += self.endData
422 |
423 | for i in range(len(self.startData_acc)):
424 | for j in range(len(self.startData_acc)):
425 | if j >= i:
426 | self.start_cor_acc[i][j] += self.start_cor[i][j]
427 | self.end_cor_acc[i][j] += self.end_cor[i][j]
428 |
429 | self.l_acc += self.l
430 |
431 | def _merge_scale(self):
432 | sc = 1.0*self.field_index
433 | self.startData_acc /= sc
434 | self.endData_acc /= sc
435 |
436 | self.l_acc /= sc
437 | for i in range(len(self.corList_acc)):
438 | self.corList_acc[i] /= sc
439 |
440 | for i in range(len(self.startData_acc)):
441 | for j in range(len(self.startData_acc)):
442 | if j >= i:
443 | self.start_cor_acc[i][j] /= sc
444 | self.end_cor_acc[i][j] /= sc
445 |
446 | def _merge_cor(self):
447 | if self.field_index == 1:
448 | self._merge_init()
449 | else:
450 | self._merge()
451 | self.l = None #causes _train to re-init everything but start_cor, end_cor
452 |
453 | def _train(self, x):
454 | self.current_train_index += len(x)
455 | super(PFANodeLayerAware, self)._train(x)
456 | if self.current_train_index == self.train_length:
457 | self.current_train_index = 0
458 | self.field_index += 1
459 | self._prepare_start_end_cor()
460 | self._merge_cor()
461 |
462 | def _insert_acc(self):
463 | self.mean = self.mean_acc
464 | self.l = self.l_acc
465 | self.corList = self.corList_acc
466 | self.startData = self.startData_acc
467 | self.endData = self.endData_acc
468 | self.start_cor = self.start_cor_acc
469 | self.end_cor = self.end_cor_acc
470 |
471 | def _stop_training(self):
472 | #print "PFA stop_training "+str(self.sindex)+" of "+str(self.maxindex)+" layer "+str(self.layerIndex)+" of "+str(self.layerMax)
473 | if self.endData_acc is None:
474 | raise mdp.TrainingException("train was never called")
475 |
476 | self._merge_scale()
477 | self._insert_acc()
478 | self._start_end_cor_clear_mean()
479 | self.calc_PFA()
480 |
481 |
482 | class PFAMerger(PFANode, Merger):
483 | def __init__(self, p = 2, k = 0, affine = True, input_dim = None, output_dim = None, dtype = None):
484 | #super(PFAMerger, self).__init__(p, k, affine, input_dim, output_dim, dtype)
485 | super(PFAMerger, self).__init__(p, k, affine, input_dim, output_dim, dtype)
486 | self.mergeCount = 0
487 | self.execCount = 0
488 | self.Ar = None
489 |
490 | # def _execute(self, x):
491 | # print "PFA merger execute"# ("+str(self.execCount)+") "+str(self.sindex)+" of "+str(self.maxindex)+" layer "+str(self.layerIndex)+" of "+str(self.layerMax)
492 | # print "mean a1 "+str(self.mean)
493 | # self.execCount += 1
494 | # z0 = x-np.outer(np.ones(len(x)), self.mean)
495 | # er = np.dot(z0, self.Ar)
496 | # print "mean a2 "+str(self.mean)
497 | # return er
498 |
499 | def _merge_init(self, node):
500 | if node.start_cor is None:
501 | node._prepare_start_end_cor()
502 |
503 | self.mean = np.array(node.mean)
504 | self.l = node.l
505 | self.corList = []
506 | for i in range(len(node.corList)):
507 | self.corList.append(np.array(node.corList[i]))
508 |
509 | self.startData = np.array(node.startData)
510 | self.endData = np.array(node.endData)
511 |
512 | self.start_cor = []
513 | self.end_cor = []
514 | for i in range(len(node.startData)):
515 | start_line = []
516 | end_line = []
517 | for j in range(len(node.startData)):
518 | if j >= i:
519 | start_line.append(np.array(node.start_cor[i][j]))
520 | end_line.append(np.array(node.end_cor[i][j]))
521 | else:
522 | start_line.append(self.start_cor[j][i].T)
523 | end_line.append(self.end_cor[j][i].T)
524 | self.start_cor.append(start_line)
525 | self.end_cor.append(end_line)
526 | self.mergeCount = 1
527 |
528 | # def freeMem(self):
529 | # #This hopefully allows the gc to free a lot of memory:
530 | # self.start_cor = None
531 | # self.end_cor = None
532 | # self.mean = None
533 | # self.corList = None
534 | # self.startData = None
535 | # self.endData = None
536 | # self.l = None
537 | # self.mergeCount += 1
538 |
539 | def _merge(self, node):
540 | #print "PFA merge "+str(node.sindex)+" of "+str(node.maxindex)+" layer "+str(node.layerIndex)+" of "+str(node.layerMax)
541 | if self.mergeCount == 0:
542 | self._merge_init(node)
543 | else:
544 | if self.start_cor is None:
545 | #super(PFAMerger, self).__init__(p, k, affine, input_dim, output_dim, dtype)
546 | self._prepare_start_end_cor()
547 | if node.start_cor is None:
548 | node._prepare_start_end_cor()
549 | self.mean = (self.l/(self.l+node.l)) * self.mean + (node.l/(self.l+node.l)) * node.mean
550 | for i in range(len(self.corList)):
551 | self.corList[i] += node.corList[i]
552 |
553 | self.startData += node.startData
554 | self.endData += node.endData
555 |
556 | for i in range(len(self.startData)):
557 | for j in range(len(self.startData)):
558 | if j >= i:
559 | self.start_cor[i][j] += node.start_cor[i][j]
560 | self.end_cor[i][j] += node.end_cor[i][j]
561 |
562 | self.l += node.l
563 | self.mergeCount += 1
564 |
565 | def _merge_scale(self):
566 | #print "mergeCount: "+str(self.mergeCount)
567 | sc = 1.0*self.mergeCount
568 | self.startData /= sc
569 | self.endData /= sc
570 |
571 | self.l /= sc
572 | for i in range(len(self.corList)):
573 | self.corList[i] /= sc
574 |
575 | for i in range(len(self.startData)):
576 | for j in range(len(self.startData)):
577 | if j >= i:
578 | self.start_cor[i][j] /= sc
579 | self.end_cor[i][j] /= sc
580 |
581 | def _stop_training(self):
582 | #print "PFA merger stop training "+str(self.sindex)+" of "+str(self.maxindex)+" layer "+str(self.layerIndex)+" of "+str(self.layerMax)
583 | self._prepare_start_end_cor()
584 |
585 | def _stop_merging(self):
586 | #print "PFA stop merging "+str(self.sindex)+" of "+str(self.maxindex)+" layer "+str(self.layerIndex)+" of "+str(self.layerMax)
587 | self._merge_scale()
588 | self._start_end_cor_clear_mean()
589 | self.calc_PFA()
590 |
591 | def execute(self, x, *args, **kwargs):
592 | """
593 | We skip _pre_execution_checks here, because train-flags
594 | might be in inconsistent states since the merger doesn't
595 | use train and stop training.
596 | Users of MergeLayer must know what they do.
597 | """
598 | #self._pre_execution_checks(x)
599 | return self._execute(self._refcast(x), *args, **kwargs)
600 |
601 |
--------------------------------------------------------------------------------
/LICENSE:
--------------------------------------------------------------------------------
1 | GNU GENERAL PUBLIC LICENSE
2 | Version 3, 29 June 2007
3 |
4 | Copyright (C) 2007 Free Software Foundation, Inc.
5 | Everyone is permitted to copy and distribute verbatim copies
6 | of this license document, but changing it is not allowed.
7 |
8 | Preamble
9 |
10 | The GNU General Public License is a free, copyleft license for
11 | software and other kinds of works.
12 |
13 | The licenses for most software and other practical works are designed
14 | to take away your freedom to share and change the works. By contrast,
15 | the GNU General Public License is intended to guarantee your freedom to
16 | share and change all versions of a program--to make sure it remains free
17 | software for all its users. We, the Free Software Foundation, use the
18 | GNU General Public License for most of our software; it applies also to
19 | any other work released this way by its authors. You can apply it to
20 | your programs, too.
21 |
22 | When we speak of free software, we are referring to freedom, not
23 | price. Our General Public Licenses are designed to make sure that you
24 | have the freedom to distribute copies of free software (and charge for
25 | them if you wish), that you receive source code or can get it if you
26 | want it, that you can change the software or use pieces of it in new
27 | free programs, and that you know you can do these things.
28 |
29 | To protect your rights, we need to prevent others from denying you
30 | these rights or asking you to surrender the rights. Therefore, you have
31 | certain responsibilities if you distribute copies of the software, or if
32 | you modify it: responsibilities to respect the freedom of others.
33 |
34 | For example, if you distribute copies of such a program, whether
35 | gratis or for a fee, you must pass on to the recipients the same
36 | freedoms that you received. You must make sure that they, too, receive
37 | or can get the source code. And you must show them these terms so they
38 | know their rights.
39 |
40 | Developers that use the GNU GPL protect your rights with two steps:
41 | (1) assert copyright on the software, and (2) offer you this License
42 | giving you legal permission to copy, distribute and/or modify it.
43 |
44 | For the developers' and authors' protection, the GPL clearly explains
45 | that there is no warranty for this free software. For both users' and
46 | authors' sake, the GPL requires that modified versions be marked as
47 | changed, so that their problems will not be attributed erroneously to
48 | authors of previous versions.
49 |
50 | Some devices are designed to deny users access to install or run
51 | modified versions of the software inside them, although the manufacturer
52 | can do so. This is fundamentally incompatible with the aim of
53 | protecting users' freedom to change the software. The systematic
54 | pattern of such abuse occurs in the area of products for individuals to
55 | use, which is precisely where it is most unacceptable. Therefore, we
56 | have designed this version of the GPL to prohibit the practice for those
57 | products. If such problems arise substantially in other domains, we
58 | stand ready to extend this provision to those domains in future versions
59 | of the GPL, as needed to protect the freedom of users.
60 |
61 | Finally, every program is threatened constantly by software patents.
62 | States should not allow patents to restrict development and use of
63 | software on general-purpose computers, but in those that do, we wish to
64 | avoid the special danger that patents applied to a free program could
65 | make it effectively proprietary. To prevent this, the GPL assures that
66 | patents cannot be used to render the program non-free.
67 |
68 | The precise terms and conditions for copying, distribution and
69 | modification follow.
70 |
71 | TERMS AND CONDITIONS
72 |
73 | 0. Definitions.
74 |
75 | "This License" refers to version 3 of the GNU General Public License.
76 |
77 | "Copyright" also means copyright-like laws that apply to other kinds of
78 | works, such as semiconductor masks.
79 |
80 | "The Program" refers to any copyrightable work licensed under this
81 | License. Each licensee is addressed as "you". "Licensees" and
82 | "recipients" may be individuals or organizations.
83 |
84 | To "modify" a work means to copy from or adapt all or part of the work
85 | in a fashion requiring copyright permission, other than the making of an
86 | exact copy. The resulting work is called a "modified version" of the
87 | earlier work or a work "based on" the earlier work.
88 |
89 | A "covered work" means either the unmodified Program or a work based
90 | on the Program.
91 |
92 | To "propagate" a work means to do anything with it that, without
93 | permission, would make you directly or secondarily liable for
94 | infringement under applicable copyright law, except executing it on a
95 | computer or modifying a private copy. Propagation includes copying,
96 | distribution (with or without modification), making available to the
97 | public, and in some countries other activities as well.
98 |
99 | To "convey" a work means any kind of propagation that enables other
100 | parties to make or receive copies. Mere interaction with a user through
101 | a computer network, with no transfer of a copy, is not conveying.
102 |
103 | An interactive user interface displays "Appropriate Legal Notices"
104 | to the extent that it includes a convenient and prominently visible
105 | feature that (1) displays an appropriate copyright notice, and (2)
106 | tells the user that there is no warranty for the work (except to the
107 | extent that warranties are provided), that licensees may convey the
108 | work under this License, and how to view a copy of this License. If
109 | the interface presents a list of user commands or options, such as a
110 | menu, a prominent item in the list meets this criterion.
111 |
112 | 1. Source Code.
113 |
114 | The "source code" for a work means the preferred form of the work
115 | for making modifications to it. "Object code" means any non-source
116 | form of a work.
117 |
118 | A "Standard Interface" means an interface that either is an official
119 | standard defined by a recognized standards body, or, in the case of
120 | interfaces specified for a particular programming language, one that
121 | is widely used among developers working in that language.
122 |
123 | The "System Libraries" of an executable work include anything, other
124 | than the work as a whole, that (a) is included in the normal form of
125 | packaging a Major Component, but which is not part of that Major
126 | Component, and (b) serves only to enable use of the work with that
127 | Major Component, or to implement a Standard Interface for which an
128 | implementation is available to the public in source code form. A
129 | "Major Component", in this context, means a major essential component
130 | (kernel, window system, and so on) of the specific operating system
131 | (if any) on which the executable work runs, or a compiler used to
132 | produce the work, or an object code interpreter used to run it.
133 |
134 | The "Corresponding Source" for a work in object code form means all
135 | the source code needed to generate, install, and (for an executable
136 | work) run the object code and to modify the work, including scripts to
137 | control those activities. However, it does not include the work's
138 | System Libraries, or general-purpose tools or generally available free
139 | programs which are used unmodified in performing those activities but
140 | which are not part of the work. For example, Corresponding Source
141 | includes interface definition files associated with source files for
142 | the work, and the source code for shared libraries and dynamically
143 | linked subprograms that the work is specifically designed to require,
144 | such as by intimate data communication or control flow between those
145 | subprograms and other parts of the work.
146 |
147 | The Corresponding Source need not include anything that users
148 | can regenerate automatically from other parts of the Corresponding
149 | Source.
150 |
151 | The Corresponding Source for a work in source code form is that
152 | same work.
153 |
154 | 2. Basic Permissions.
155 |
156 | All rights granted under this License are granted for the term of
157 | copyright on the Program, and are irrevocable provided the stated
158 | conditions are met. This License explicitly affirms your unlimited
159 | permission to run the unmodified Program. The output from running a
160 | covered work is covered by this License only if the output, given its
161 | content, constitutes a covered work. This License acknowledges your
162 | rights of fair use or other equivalent, as provided by copyright law.
163 |
164 | You may make, run and propagate covered works that you do not
165 | convey, without conditions so long as your license otherwise remains
166 | in force. You may convey covered works to others for the sole purpose
167 | of having them make modifications exclusively for you, or provide you
168 | with facilities for running those works, provided that you comply with
169 | the terms of this License in conveying all material for which you do
170 | not control copyright. Those thus making or running the covered works
171 | for you must do so exclusively on your behalf, under your direction
172 | and control, on terms that prohibit them from making any copies of
173 | your copyrighted material outside their relationship with you.
174 |
175 | Conveying under any other circumstances is permitted solely under
176 | the conditions stated below. Sublicensing is not allowed; section 10
177 | makes it unnecessary.
178 |
179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law.
180 |
181 | No covered work shall be deemed part of an effective technological
182 | measure under any applicable law fulfilling obligations under article
183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or
184 | similar laws prohibiting or restricting circumvention of such
185 | measures.
186 |
187 | When you convey a covered work, you waive any legal power to forbid
188 | circumvention of technological measures to the extent such circumvention
189 | is effected by exercising rights under this License with respect to
190 | the covered work, and you disclaim any intention to limit operation or
191 | modification of the work as a means of enforcing, against the work's
192 | users, your or third parties' legal rights to forbid circumvention of
193 | technological measures.
194 |
195 | 4. Conveying Verbatim Copies.
196 |
197 | You may convey verbatim copies of the Program's source code as you
198 | receive it, in any medium, provided that you conspicuously and
199 | appropriately publish on each copy an appropriate copyright notice;
200 | keep intact all notices stating that this License and any
201 | non-permissive terms added in accord with section 7 apply to the code;
202 | keep intact all notices of the absence of any warranty; and give all
203 | recipients a copy of this License along with the Program.
204 |
205 | You may charge any price or no price for each copy that you convey,
206 | and you may offer support or warranty protection for a fee.
207 |
208 | 5. Conveying Modified Source Versions.
209 |
210 | You may convey a work based on the Program, or the modifications to
211 | produce it from the Program, in the form of source code under the
212 | terms of section 4, provided that you also meet all of these conditions:
213 |
214 | a) The work must carry prominent notices stating that you modified
215 | it, and giving a relevant date.
216 |
217 | b) The work must carry prominent notices stating that it is
218 | released under this License and any conditions added under section
219 | 7. This requirement modifies the requirement in section 4 to
220 | "keep intact all notices".
221 |
222 | c) You must license the entire work, as a whole, under this
223 | License to anyone who comes into possession of a copy. This
224 | License will therefore apply, along with any applicable section 7
225 | additional terms, to the whole of the work, and all its parts,
226 | regardless of how they are packaged. This License gives no
227 | permission to license the work in any other way, but it does not
228 | invalidate such permission if you have separately received it.
229 |
230 | d) If the work has interactive user interfaces, each must display
231 | Appropriate Legal Notices; however, if the Program has interactive
232 | interfaces that do not display Appropriate Legal Notices, your
233 | work need not make them do so.
234 |
235 | A compilation of a covered work with other separate and independent
236 | works, which are not by their nature extensions of the covered work,
237 | and which are not combined with it such as to form a larger program,
238 | in or on a volume of a storage or distribution medium, is called an
239 | "aggregate" if the compilation and its resulting copyright are not
240 | used to limit the access or legal rights of the compilation's users
241 | beyond what the individual works permit. Inclusion of a covered work
242 | in an aggregate does not cause this License to apply to the other
243 | parts of the aggregate.
244 |
245 | 6. Conveying Non-Source Forms.
246 |
247 | You may convey a covered work in object code form under the terms
248 | of sections 4 and 5, provided that you also convey the
249 | machine-readable Corresponding Source under the terms of this License,
250 | in one of these ways:
251 |
252 | a) Convey the object code in, or embodied in, a physical product
253 | (including a physical distribution medium), accompanied by the
254 | Corresponding Source fixed on a durable physical medium
255 | customarily used for software interchange.
256 |
257 | b) Convey the object code in, or embodied in, a physical product
258 | (including a physical distribution medium), accompanied by a
259 | written offer, valid for at least three years and valid for as
260 | long as you offer spare parts or customer support for that product
261 | model, to give anyone who possesses the object code either (1) a
262 | copy of the Corresponding Source for all the software in the
263 | product that is covered by this License, on a durable physical
264 | medium customarily used for software interchange, for a price no
265 | more than your reasonable cost of physically performing this
266 | conveying of source, or (2) access to copy the
267 | Corresponding Source from a network server at no charge.
268 |
269 | c) Convey individual copies of the object code with a copy of the
270 | written offer to provide the Corresponding Source. This
271 | alternative is allowed only occasionally and noncommercially, and
272 | only if you received the object code with such an offer, in accord
273 | with subsection 6b.
274 |
275 | d) Convey the object code by offering access from a designated
276 | place (gratis or for a charge), and offer equivalent access to the
277 | Corresponding Source in the same way through the same place at no
278 | further charge. You need not require recipients to copy the
279 | Corresponding Source along with the object code. If the place to
280 | copy the object code is a network server, the Corresponding Source
281 | may be on a different server (operated by you or a third party)
282 | that supports equivalent copying facilities, provided you maintain
283 | clear directions next to the object code saying where to find the
284 | Corresponding Source. Regardless of what server hosts the
285 | Corresponding Source, you remain obligated to ensure that it is
286 | available for as long as needed to satisfy these requirements.
287 |
288 | e) Convey the object code using peer-to-peer transmission, provided
289 | you inform other peers where the object code and Corresponding
290 | Source of the work are being offered to the general public at no
291 | charge under subsection 6d.
292 |
293 | A separable portion of the object code, whose source code is excluded
294 | from the Corresponding Source as a System Library, need not be
295 | included in conveying the object code work.
296 |
297 | A "User Product" is either (1) a "consumer product", which means any
298 | tangible personal property which is normally used for personal, family,
299 | or household purposes, or (2) anything designed or sold for incorporation
300 | into a dwelling. In determining whether a product is a consumer product,
301 | doubtful cases shall be resolved in favor of coverage. For a particular
302 | product received by a particular user, "normally used" refers to a
303 | typical or common use of that class of product, regardless of the status
304 | of the particular user or of the way in which the particular user
305 | actually uses, or expects or is expected to use, the product. A product
306 | is a consumer product regardless of whether the product has substantial
307 | commercial, industrial or non-consumer uses, unless such uses represent
308 | the only significant mode of use of the product.
309 |
310 | "Installation Information" for a User Product means any methods,
311 | procedures, authorization keys, or other information required to install
312 | and execute modified versions of a covered work in that User Product from
313 | a modified version of its Corresponding Source. The information must
314 | suffice to ensure that the continued functioning of the modified object
315 | code is in no case prevented or interfered with solely because
316 | modification has been made.
317 |
318 | If you convey an object code work under this section in, or with, or
319 | specifically for use in, a User Product, and the conveying occurs as
320 | part of a transaction in which the right of possession and use of the
321 | User Product is transferred to the recipient in perpetuity or for a
322 | fixed term (regardless of how the transaction is characterized), the
323 | Corresponding Source conveyed under this section must be accompanied
324 | by the Installation Information. But this requirement does not apply
325 | if neither you nor any third party retains the ability to install
326 | modified object code on the User Product (for example, the work has
327 | been installed in ROM).
328 |
329 | The requirement to provide Installation Information does not include a
330 | requirement to continue to provide support service, warranty, or updates
331 | for a work that has been modified or installed by the recipient, or for
332 | the User Product in which it has been modified or installed. Access to a
333 | network may be denied when the modification itself materially and
334 | adversely affects the operation of the network or violates the rules and
335 | protocols for communication across the network.
336 |
337 | Corresponding Source conveyed, and Installation Information provided,
338 | in accord with this section must be in a format that is publicly
339 | documented (and with an implementation available to the public in
340 | source code form), and must require no special password or key for
341 | unpacking, reading or copying.
342 |
343 | 7. Additional Terms.
344 |
345 | "Additional permissions" are terms that supplement the terms of this
346 | License by making exceptions from one or more of its conditions.
347 | Additional permissions that are applicable to the entire Program shall
348 | be treated as though they were included in this License, to the extent
349 | that they are valid under applicable law. If additional permissions
350 | apply only to part of the Program, that part may be used separately
351 | under those permissions, but the entire Program remains governed by
352 | this License without regard to the additional permissions.
353 |
354 | When you convey a copy of a covered work, you may at your option
355 | remove any additional permissions from that copy, or from any part of
356 | it. (Additional permissions may be written to require their own
357 | removal in certain cases when you modify the work.) You may place
358 | additional permissions on material, added by you to a covered work,
359 | for which you have or can give appropriate copyright permission.
360 |
361 | Notwithstanding any other provision of this License, for material you
362 | add to a covered work, you may (if authorized by the copyright holders of
363 | that material) supplement the terms of this License with terms:
364 |
365 | a) Disclaiming warranty or limiting liability differently from the
366 | terms of sections 15 and 16 of this License; or
367 |
368 | b) Requiring preservation of specified reasonable legal notices or
369 | author attributions in that material or in the Appropriate Legal
370 | Notices displayed by works containing it; or
371 |
372 | c) Prohibiting misrepresentation of the origin of that material, or
373 | requiring that modified versions of such material be marked in
374 | reasonable ways as different from the original version; or
375 |
376 | d) Limiting the use for publicity purposes of names of licensors or
377 | authors of the material; or
378 |
379 | e) Declining to grant rights under trademark law for use of some
380 | trade names, trademarks, or service marks; or
381 |
382 | f) Requiring indemnification of licensors and authors of that
383 | material by anyone who conveys the material (or modified versions of
384 | it) with contractual assumptions of liability to the recipient, for
385 | any liability that these contractual assumptions directly impose on
386 | those licensors and authors.
387 |
388 | All other non-permissive additional terms are considered "further
389 | restrictions" within the meaning of section 10. If the Program as you
390 | received it, or any part of it, contains a notice stating that it is
391 | governed by this License along with a term that is a further
392 | restriction, you may remove that term. If a license document contains
393 | a further restriction but permits relicensing or conveying under this
394 | License, you may add to a covered work material governed by the terms
395 | of that license document, provided that the further restriction does
396 | not survive such relicensing or conveying.
397 |
398 | If you add terms to a covered work in accord with this section, you
399 | must place, in the relevant source files, a statement of the
400 | additional terms that apply to those files, or a notice indicating
401 | where to find the applicable terms.
402 |
403 | Additional terms, permissive or non-permissive, may be stated in the
404 | form of a separately written license, or stated as exceptions;
405 | the above requirements apply either way.
406 |
407 | 8. Termination.
408 |
409 | You may not propagate or modify a covered work except as expressly
410 | provided under this License. Any attempt otherwise to propagate or
411 | modify it is void, and will automatically terminate your rights under
412 | this License (including any patent licenses granted under the third
413 | paragraph of section 11).
414 |
415 | However, if you cease all violation of this License, then your
416 | license from a particular copyright holder is reinstated (a)
417 | provisionally, unless and until the copyright holder explicitly and
418 | finally terminates your license, and (b) permanently, if the copyright
419 | holder fails to notify you of the violation by some reasonable means
420 | prior to 60 days after the cessation.
421 |
422 | Moreover, your license from a particular copyright holder is
423 | reinstated permanently if the copyright holder notifies you of the
424 | violation by some reasonable means, this is the first time you have
425 | received notice of violation of this License (for any work) from that
426 | copyright holder, and you cure the violation prior to 30 days after
427 | your receipt of the notice.
428 |
429 | Termination of your rights under this section does not terminate the
430 | licenses of parties who have received copies or rights from you under
431 | this License. If your rights have been terminated and not permanently
432 | reinstated, you do not qualify to receive new licenses for the same
433 | material under section 10.
434 |
435 | 9. Acceptance Not Required for Having Copies.
436 |
437 | You are not required to accept this License in order to receive or
438 | run a copy of the Program. Ancillary propagation of a covered work
439 | occurring solely as a consequence of using peer-to-peer transmission
440 | to receive a copy likewise does not require acceptance. However,
441 | nothing other than this License grants you permission to propagate or
442 | modify any covered work. These actions infringe copyright if you do
443 | not accept this License. Therefore, by modifying or propagating a
444 | covered work, you indicate your acceptance of this License to do so.
445 |
446 | 10. Automatic Licensing of Downstream Recipients.
447 |
448 | Each time you convey a covered work, the recipient automatically
449 | receives a license from the original licensors, to run, modify and
450 | propagate that work, subject to this License. You are not responsible
451 | for enforcing compliance by third parties with this License.
452 |
453 | An "entity transaction" is a transaction transferring control of an
454 | organization, or substantially all assets of one, or subdividing an
455 | organization, or merging organizations. If propagation of a covered
456 | work results from an entity transaction, each party to that
457 | transaction who receives a copy of the work also receives whatever
458 | licenses to the work the party's predecessor in interest had or could
459 | give under the previous paragraph, plus a right to possession of the
460 | Corresponding Source of the work from the predecessor in interest, if
461 | the predecessor has it or can get it with reasonable efforts.
462 |
463 | You may not impose any further restrictions on the exercise of the
464 | rights granted or affirmed under this License. For example, you may
465 | not impose a license fee, royalty, or other charge for exercise of
466 | rights granted under this License, and you may not initiate litigation
467 | (including a cross-claim or counterclaim in a lawsuit) alleging that
468 | any patent claim is infringed by making, using, selling, offering for
469 | sale, or importing the Program or any portion of it.
470 |
471 | 11. Patents.
472 |
473 | A "contributor" is a copyright holder who authorizes use under this
474 | License of the Program or a work on which the Program is based. The
475 | work thus licensed is called the contributor's "contributor version".
476 |
477 | A contributor's "essential patent claims" are all patent claims
478 | owned or controlled by the contributor, whether already acquired or
479 | hereafter acquired, that would be infringed by some manner, permitted
480 | by this License, of making, using, or selling its contributor version,
481 | but do not include claims that would be infringed only as a
482 | consequence of further modification of the contributor version. For
483 | purposes of this definition, "control" includes the right to grant
484 | patent sublicenses in a manner consistent with the requirements of
485 | this License.
486 |
487 | Each contributor grants you a non-exclusive, worldwide, royalty-free
488 | patent license under the contributor's essential patent claims, to
489 | make, use, sell, offer for sale, import and otherwise run, modify and
490 | propagate the contents of its contributor version.
491 |
492 | In the following three paragraphs, a "patent license" is any express
493 | agreement or commitment, however denominated, not to enforce a patent
494 | (such as an express permission to practice a patent or covenant not to
495 | sue for patent infringement). To "grant" such a patent license to a
496 | party means to make such an agreement or commitment not to enforce a
497 | patent against the party.
498 |
499 | If you convey a covered work, knowingly relying on a patent license,
500 | and the Corresponding Source of the work is not available for anyone
501 | to copy, free of charge and under the terms of this License, through a
502 | publicly available network server or other readily accessible means,
503 | then you must either (1) cause the Corresponding Source to be so
504 | available, or (2) arrange to deprive yourself of the benefit of the
505 | patent license for this particular work, or (3) arrange, in a manner
506 | consistent with the requirements of this License, to extend the patent
507 | license to downstream recipients. "Knowingly relying" means you have
508 | actual knowledge that, but for the patent license, your conveying the
509 | covered work in a country, or your recipient's use of the covered work
510 | in a country, would infringe one or more identifiable patents in that
511 | country that you have reason to believe are valid.
512 |
513 | If, pursuant to or in connection with a single transaction or
514 | arrangement, you convey, or propagate by procuring conveyance of, a
515 | covered work, and grant a patent license to some of the parties
516 | receiving the covered work authorizing them to use, propagate, modify
517 | or convey a specific copy of the covered work, then the patent license
518 | you grant is automatically extended to all recipients of the covered
519 | work and works based on it.
520 |
521 | A patent license is "discriminatory" if it does not include within
522 | the scope of its coverage, prohibits the exercise of, or is
523 | conditioned on the non-exercise of one or more of the rights that are
524 | specifically granted under this License. You may not convey a covered
525 | work if you are a party to an arrangement with a third party that is
526 | in the business of distributing software, under which you make payment
527 | to the third party based on the extent of your activity of conveying
528 | the work, and under which the third party grants, to any of the
529 | parties who would receive the covered work from you, a discriminatory
530 | patent license (a) in connection with copies of the covered work
531 | conveyed by you (or copies made from those copies), or (b) primarily
532 | for and in connection with specific products or compilations that
533 | contain the covered work, unless you entered into that arrangement,
534 | or that patent license was granted, prior to 28 March 2007.
535 |
536 | Nothing in this License shall be construed as excluding or limiting
537 | any implied license or other defenses to infringement that may
538 | otherwise be available to you under applicable patent law.
539 |
540 | 12. No Surrender of Others' Freedom.
541 |
542 | If conditions are imposed on you (whether by court order, agreement or
543 | otherwise) that contradict the conditions of this License, they do not
544 | excuse you from the conditions of this License. If you cannot convey a
545 | covered work so as to satisfy simultaneously your obligations under this
546 | License and any other pertinent obligations, then as a consequence you may
547 | not convey it at all. For example, if you agree to terms that obligate you
548 | to collect a royalty for further conveying from those to whom you convey
549 | the Program, the only way you could satisfy both those terms and this
550 | License would be to refrain entirely from conveying the Program.
551 |
552 | 13. Use with the GNU Affero General Public License.
553 |
554 | Notwithstanding any other provision of this License, you have
555 | permission to link or combine any covered work with a work licensed
556 | under version 3 of the GNU Affero General Public License into a single
557 | combined work, and to convey the resulting work. The terms of this
558 | License will continue to apply to the part which is the covered work,
559 | but the special requirements of the GNU Affero General Public License,
560 | section 13, concerning interaction through a network will apply to the
561 | combination as such.
562 |
563 | 14. Revised Versions of this License.
564 |
565 | The Free Software Foundation may publish revised and/or new versions of
566 | the GNU General Public License from time to time. Such new versions will
567 | be similar in spirit to the present version, but may differ in detail to
568 | address new problems or concerns.
569 |
570 | Each version is given a distinguishing version number. If the
571 | Program specifies that a certain numbered version of the GNU General
572 | Public License "or any later version" applies to it, you have the
573 | option of following the terms and conditions either of that numbered
574 | version or of any later version published by the Free Software
575 | Foundation. If the Program does not specify a version number of the
576 | GNU General Public License, you may choose any version ever published
577 | by the Free Software Foundation.
578 |
579 | If the Program specifies that a proxy can decide which future
580 | versions of the GNU General Public License can be used, that proxy's
581 | public statement of acceptance of a version permanently authorizes you
582 | to choose that version for the Program.
583 |
584 | Later license versions may give you additional or different
585 | permissions. However, no additional obligations are imposed on any
586 | author or copyright holder as a result of your choosing to follow a
587 | later version.
588 |
589 | 15. Disclaimer of Warranty.
590 |
591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
599 |
600 | 16. Limitation of Liability.
601 |
602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
610 | SUCH DAMAGES.
611 |
612 | 17. Interpretation of Sections 15 and 16.
613 |
614 | If the disclaimer of warranty and limitation of liability provided
615 | above cannot be given local legal effect according to their terms,
616 | reviewing courts shall apply local law that most closely approximates
617 | an absolute waiver of all civil liability in connection with the
618 | Program, unless a warranty or assumption of liability accompanies a
619 | copy of the Program in return for a fee.
620 |
621 | END OF TERMS AND CONDITIONS
622 |
623 | How to Apply These Terms to Your New Programs
624 |
625 | If you develop a new program, and you want it to be of the greatest
626 | possible use to the public, the best way to achieve this is to make it
627 | free software which everyone can redistribute and change under these terms.
628 |
629 | To do so, attach the following notices to the program. It is safest
630 | to attach them to the start of each source file to most effectively
631 | state the exclusion of warranty; and each file should have at least
632 | the "copyright" line and a pointer to where the full notice is found.
633 |
634 |
635 | Copyright (C)
636 |
637 | This program is free software: you can redistribute it and/or modify
638 | it under the terms of the GNU General Public License as published by
639 | the Free Software Foundation, either version 3 of the License, or
640 | (at your option) any later version.
641 |
642 | This program is distributed in the hope that it will be useful,
643 | but WITHOUT ANY WARRANTY; without even the implied warranty of
644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
645 | GNU General Public License for more details.
646 |
647 | You should have received a copy of the GNU General Public License
648 | along with this program. If not, see .
649 |
650 | Also add information on how to contact you by electronic and paper mail.
651 |
652 | If the program does terminal interaction, make it output a short
653 | notice like this when it starts in an interactive mode:
654 |
655 | Copyright (C)
656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
657 | This is free software, and you are welcome to redistribute it
658 | under certain conditions; type `show c' for details.
659 |
660 | The hypothetical commands `show w' and `show c' should show the appropriate
661 | parts of the General Public License. Of course, your program's commands
662 | might be different; for a GUI interface, you would use an "about box".
663 |
664 | You should also get your employer (if you work as a programmer) or school,
665 | if any, to sign a "copyright disclaimer" for the program, if necessary.
666 | For more information on this, and how to apply and follow the GNU GPL, see
667 | .
668 |
669 | The GNU General Public License does not permit incorporating your program
670 | into proprietary programs. If your program is a subroutine library, you
671 | may consider it more useful to permit linking proprietary applications with
672 | the library. If this is what you want to do, use the GNU Lesser General
673 | Public License instead of this License. But first, please read
674 | .
675 |
--------------------------------------------------------------------------------
/src/PFACoreUtil.py:
--------------------------------------------------------------------------------
1 | #
2 | #
3 | # Copyright of GNUPFA:
4 | # Copyright (c) 2013, 2014 Institut fuer Neuroinformatik,
5 | # Ruhr-Universitaet Bochum, Germany. All rights reserved.
6 | #
7 | #
8 | # This file is part of GNUPFA.
9 | #
10 | # GNUPFA is free software: you can redistribute it and/or modify
11 | # it under the terms of the GNU General Public License as published by
12 | # the Free Software Foundation, either version 3 of the License, or
13 | # (at your option) any later version.
14 | #
15 | # GNUPFA is distributed in the hope that it will be useful,
16 | # but WITHOUT ANY WARRANTY; without even the implied warranty of
17 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 | # GNU General Public License for more details.
19 | #
20 | # You should have received a copy of the GNU General Public License
21 | # along with GNUPFA. If not, see .
22 | #
23 | #
24 | # Linking this library statically or dynamically with other modules is
25 | # making a combined work based on this library. Thus, the terms and
26 | # conditions of the GNU General Public License cover the whole
27 | # combination.
28 | #
29 |
30 |
31 | import numpy as np
32 | from numpy import linalg as LA
33 |
34 | def NonSquareIdentity(shape, rowOffset = 0, columnOffset = 0):
35 | er = np.zeros(shape)
36 | l = shape[0]
37 | if shape[1] < l:
38 | l = shape[1]
39 | for i in range(l):
40 | er[rowOffset+i][columnOffset+i] = 1.0
41 | return er
42 |
43 | #W is expected to be a broad matrix, transform would be multiplied from left to it, inverse transform (kron) from right
44 | #Transform is expected to be orthogonal (or orthogonal subbase)
45 | def transformPredictor(W, transform):
46 | X_ = np.kron(np.identity(len(W.T)/len(W)), transform.T)
47 | if len(W.T)%len(W) == 0: #Non-affine
48 | return np.dot(transform, np.dot(W, X_))
49 | else: #affine
50 | return np.hstack([np.dot(transform, np.dot(W.T[:-1].T, X_)), np.dot(transform, W.T[-1:].T)])
51 |
52 | def kronAffine(A, p):
53 | X_ = np.kron(np.identity(p), A)
54 | z = np.zeros([len(X_), 1])
55 | X_ = np.hstack([X_, z])
56 | z = np.zeros([1, len(X_.T)])
57 | z[0][-1] = 1.0
58 | return np.vstack([X_, z])
59 |
60 | #Note that secondMoment is expected without besselsCorrection, i.e. secondMomentSum/length rather than secondMomentSum/(length-1)
61 | def calcSpheringMatrixFromMeanAndSecondMoment(mean, secondMoment, length, threshold = 0.00000000001, besselsCorrection = 0):
62 | """
63 | Returns sphered version of the given data and the sphering parameters as well.
64 | The data is expected in the format (time, dims). The sphering matrix may reduce
65 | the dimensionality of the data by deleting dimensions that have (near) zero
66 | variance. The returned sphering matrix has the shape (high dim, low dim).
67 | The sphering is done with respect to a covariance-matrix including Bessel's correction,
68 | if the besselsCorrection parameter is True (default is False).
69 | """
70 | # if length == -1:
71 | # length = len(data)
72 | #cov = (secondMoment/length-np.outer(mean, mean))
73 | cov = (secondMoment-np.outer(mean, mean))
74 | if besselsCorrection != 0:
75 | cov *= (length/(length-1.0))
76 | eg, ev = LA.eigh(cov)
77 | eg2 = []
78 | ev2 = []
79 | # print "+++sph eg+++"
80 | # print eg
81 | # print "----"
82 | for i in range(0, len(eg)):
83 | if (eg[i] >= threshold):
84 | eg2.append(1.0/np.sqrt(eg[i]))
85 | ev2.append(ev.T[i])
86 | return np.dot(np.transpose(ev2), np.diag(eg2))
87 |
88 | def calcSpheringParametersAndDataRefImp(data, threshold = 0.00001, offset = 0, length = -1, besselsCorrection = 0):
89 | """
90 | Returns sphered version of the given data and the sphering parameters as well.
91 | The data is expected in the format (time, dims). The sphering matrix may reduce
92 | the dimensionality of the data by deleting dimensions that have (near) zero
93 | variance. The returned sphering matrix has the shape (high dim, low dim).
94 | The sphering is done with respect to a covariance-matrix including Bessel's correction,
95 | if the besselsCorrection parameter is True (default is False).
96 | """
97 | if besselsCorrection != 0:
98 | besselsCorrection = 1
99 | if length == -1:
100 | length = len(data)
101 | mean = data[offset:].mean(0)
102 | mean2 = data[offset:].sum(0)/len(data[offset:])
103 | data0 = data-np.outer(np.ones(len(data)), mean)
104 | #With Bessel's correction:
105 | #cov = np.multiply(np.dot(data0[offset:].T, data0[offset:]), 1.0/(len(data)-offset-besselsCorrection))
106 | cov = np.dot(data0[offset:].T, data0[offset:])
107 | eg, ev = LA.eigh(cov)
108 | eg2 = []
109 | ev2 = []
110 | for i in range(0, len(eg)):
111 | if (eg[i] >= threshold):
112 | #eg2.append(1.0/(np.sqrt(eg[i]*len(data))))
113 | eg2.append(1.0/(np.sqrt(eg[i]/(len(data)-offset-besselsCorrection))))
114 | ev2.append(ev.T[i])
115 | S = np.dot(np.transpose(ev2), np.diag(eg2))
116 | # for i in range(0, len(eg)):
117 | # eg[i] = eg[i]**(-0.5)
118 | # S = np.dot(ev.T, np.diag(eg))
119 | return [mean, S, data0.dot(S)]
120 |
121 | #def calcAutoCorrelationList(data, p, besselsCorrection = 0):
122 | # er = []
123 | # data0 = data[p:].T
124 | # for i in range(1, p+1):
125 | # #er.append(np.multiply(np.dot(data0, data[p-i:len(data)-i]), 1.0/(len(data)-p)))
126 | # er.append(np.dot(data0, data[p-i:len(data)-i]))
127 | # return er
128 | #
129 | ## if besselsCorrection != 0:
130 | ## besselsCorrection = 1
131 | ### data0 = data[p:].T
132 | ## for i in range(1, p+1):
133 | ### er.append(np.dot(data0, data[p-i:len(data)-i]))
134 | ## er.append(np.multiply(np.dot(data[i:].T, data[:-i]), (1.0*(len(data)-besselsCorrection))/(len(data)-i-besselsCorrection)))
135 | ## return er
136 | #
137 | #def zZiFromAutoCorrelationsList(auto, i, p):
138 | # return np.hstack(auto[i:i+p])
139 | #
140 | #def ZZFromAutoCorrelationsList(auto, p, data, besselsCorrection = 0, dataSphered = True):
141 | # if besselsCorrection != 0:
142 | # besselsCorrection = 1
143 | # cov = np.multiply(np.identity(len(data[0])), len(data)-besselsCorrection)
144 | # if not dataSphered:
145 | # #cov = np.multiply(np.dot(data.T, data), 1.0/(len(data)-besselsCorrection))
146 | # cov = np.dot(data.T, data)
147 | # lines = []
148 | # #i iterates through lines, j through columns
149 | # for i in range(p):
150 | # line = []
151 | # for j in range(p):
152 | # if i == j:
153 | # line.append(cov)
154 | # elif i > j:
155 | # line.append(lines[j][i].T)
156 | # else:
157 | # line.append(auto[i-1])
158 | # lines.append(line)
159 | # for i in range(p):
160 | # lines[i] = np.hstack(lines[i])
161 | # return np.vstack(lines)
162 |
163 | def invertByProjectionRefImp(M, evThreshold = 0.000001):
164 | eg, ev = LA.eigh(M)
165 | r = evThreshold**2
166 | for i in range(0, len(eg)):
167 | if (eg[i]**2 > r):
168 | eg[i] = 1.0/eg[i]
169 | else:
170 | eg[i] = 0.0
171 | return np.dot(ev, np.dot(np.diag(eg), ev.T))
172 |
173 | def linearRegressionCoeff(srcData, destData):
174 | srcCov = np.dot(srcData.T, srcData)
175 | cor = np.dot(srcData.T, destData)
176 | srcCovI = invertByProjectionRefImp(srcCov)
177 | return np.dot(srcCovI, cor)
178 |
179 | # def affineRegressionCoeff(srcData, destData):
180 | # srcCov = np.dot(srcData.T, srcData)
181 | # cor = np.dot(srcData.T, destData)
182 | # srcCovI = invertByProjectionRefImp(srcCov)
183 | # return np.dot(srcCovI, cor)
184 |
185 | #def fullRegressionMatrixFromRegressionCoeff(W):
186 | # tmp = np.zeros([len(W[0])-len(W), len(W[0])])
187 | # for i in range(0, len(tmp)):
188 | # tmp[i][i] = 1.0
189 | # return np.vstack([W, tmp])
190 |
191 | def calcShiftedDataListRefImp(data, p, p_offset = 1):
192 | dataZeta = []
193 | off = p_offset
194 | if off == 0:
195 | off += 1
196 | dataZeta.append(data[p:])
197 | for i in range(off, p+1):
198 | dataZeta.append(data[p-i: -i])
199 | return dataZeta
200 |
201 | def calcZetaDataRefImp(data, p, delay = 0):
202 | return np.hstack(calcShiftedDataListRefImp(data, p+delay, 1+delay))
203 | # dataZeta = calcShiftedDataList(data, p-1)
204 | # return np.hstack(dataZeta)[0:len(dataZeta[0])-1-delay]
205 |
206 | def calcZetacDataRefImp(data, p, delay = 0):
207 | return np.hstack([calcZetaDataRefImp(data, p, delay), np.ones([len(data)-p-delay, 1])])
208 |
209 | def calcZeta0DataRefImp(data, p):
210 | return np.hstack(calcShiftedDataListRefImp(data[1:], p-1, 0))
211 |
212 | def calcZeta0cDataRefImp(data, p):
213 | return np.hstack([calcZeta0DataRefImp(data, p), np.ones([len(data)-p, 1])])
214 |
215 | def empiricalRawErrorRefImp(data, W, srcData = None):
216 | """
217 | Measures how good following equation is fulfilled in average over time: z = W zeta
218 | Evaluates data from p to len(data), retrieves p via len(W[0])/len(W)
219 | """
220 | sDat = srcData
221 | if sDat is None:
222 | sDat = data
223 | p = len(W[0])/len(W)
224 | pre = np.dot(calcZetaDataRefImp(sDat, p), W.T)
225 | #real = data[0:len(data)-p]
226 | real = data[p:len(data)]
227 | #err = (pre[0]-real[0])**2
228 | #for i in range(1, len(real)):
229 | # err += (pre[i]-real[i])**2
230 | #return np.multiply(err, 1.0/len(real))
231 | return (LA.norm(real-pre)**2)/len(real)
232 |
233 | def empiricalRawErrorAffineRefImp(data, W_c, srcData = None):
234 | """
235 | Measures how good following equation is fulfilled in average over time: z = W zeta + c
236 | Evaluates data from p to len(data), retrieves p via len(W[0])/len(W)
237 | """
238 | sDat = srcData
239 | if sDat is None:
240 | sDat = data
241 | p = (len(W_c[0])-1)/len(W_c)
242 | pre = np.dot(calcZetacDataRefImp(sDat, p), W_c.T)
243 | real = data[p:len(data)]
244 | return (LA.norm(real-pre)**2)/len(real)
245 |
246 | def empiricalRawErrorComponentsRefImp(data, W, srcData = None):
247 | """
248 | Measures how good following equation is fulfilled in average over time: z = W zeta
249 | Evaluates data from p to len(data), retrieves p via len(W[0])/len(W)
250 | """
251 | sDat = srcData
252 | if sDat is None:
253 | sDat = data
254 | p = len(W[0])/len(W)
255 | pre = np.dot(calcZetaDataRefImp(sDat, p), W.T)
256 | real = data[p:len(data)]
257 | res = (pre-real).T
258 | er = [] #np.zeros(len(res))
259 | for i in range(0, len(res)):
260 | er.append(np.inner(res[i], res[i]))
261 | return np.multiply(er, 1.0/len(real))
262 |
263 | def empiricalRawErrorComponentsAffineRefImp(data, W_c, srcData = None):
264 | """
265 | Measures how good following equation is fulfilled in average over time: z = W zeta + c
266 | Evaluates data from p to len(data), retrieves p via len(W[0])/len(W)
267 | """
268 | sDat = srcData
269 | if sDat is None:
270 | sDat = data
271 | p = (len(W_c[0])-1)/len(W_c)
272 | pre = np.dot(calcZetacDataRefImp(sDat, p), W_c.T)
273 | real = data[p:len(data)]
274 | res = (pre-real).T
275 | er = [] #np.zeros(len(res))
276 | for i in range(0, len(res)):
277 | er.append(np.inner(res[i], res[i]))
278 | return np.multiply(er, 1.0/len(real))
279 |
280 | def predictNextFeatures(data, W):
281 | p = len(W[0])/len(W)
282 | return np.dot(np.hstack(data[-p:][::-1]), W.T)
283 |
284 | def predictNextFeaturesAffine(data, W_c):
285 | p = (len(W_c[0])-1)/len(W_c)
286 | return np.dot(np.hstack(data[-p:][::-1]), W_c.T[:-1])+W_c.T[-1]
287 |
288 | def calcRegressionCoeffRefImp(data, p, evThreshold = 0.0001):
289 | zeta = calcZetaDataRefImp(data, p)
290 | z_p = data[p:]
291 | zZT = np.dot(z_p.T, zeta)
292 | ZZT = np.dot(zeta.T, zeta)
293 | ZZTI = invertByProjectionRefImp(ZZT, evThreshold)
294 | return np.dot(zZT, ZZTI)
295 |
296 | def calcRegressionCoeffAffineRefImp(data, p, evThreshold = 0.0001):
297 | zetac = calcZetacDataRefImp(data, p)
298 | z_p = data[p:]
299 | zZT = np.dot(z_p.T, zetac)
300 | ZZT = np.dot(zetac.T, zetac)
301 | ZZTI = invertByProjectionRefImp(ZZT, evThreshold)
302 | return np.dot(zZT, ZZTI)
303 |
304 | def buildV_cRefImp(W_c):
305 | p = (len(W_c[0])-1)/len(W_c)
306 | id = np.identity((p-1)*len(W_c))
307 | zer = np.zeros([len(id), len(W_c)+1])
308 | bot = np.zeros([len(W_c[0])])
309 | bot[len(bot)-1] = 1.0
310 | return np.vstack([W_c, np.hstack([id, zer]), bot])
311 |
312 | def buildVRefImp(W):
313 | p = (len(W[0]))/len(W)
314 | id = np.identity((p-1)*len(W))
315 | zer = np.zeros([len(id), len(W)])
316 | return np.vstack([W, np.hstack([id, zer])])
317 |
318 | def calcExtractionForErrorCovRefImp(X, r):
319 | eg, ev = LA.eigh(X)
320 | sort = np.argsort(eg)
321 | ev2 = []
322 | for i in range(0, r):
323 | ev2.append(ev.T[sort[i]])
324 | return np.transpose(ev2)
325 |
326 | def calcExtractionWithWeightsForErrorCovRefImp(X, r):
327 | eg, ev = LA.eigh(X)
328 | sort = np.argsort(eg)
329 | ev2 = []
330 | eg2 = []
331 | for i in range(0, r):
332 | ev2.append(ev.T[sort[i]])
333 | eg2.append(eg[sort[i]])
334 | return np.transpose(ev2), np.array(eg), np.array(eg2)
335 |
336 | def calcErrorCoeffRefImp(data, W, k = 0):
337 | p = (len(W[0]))/len(W)
338 | V = buildVRefImp(W)
339 | X = np.zeros([len(W), len(W)])
340 | #z_pk = data[p+k:]
341 | WV = W
342 | for i in range(0, k+1):
343 | if i > 0:
344 | WV = np.dot(WV, V)
345 | zeta1_i = calcZetaDataRefImp(data, p, i)
346 | z_i_pre = np.dot(zeta1_i, WV.T)
347 | res_i = data[p+i:]-z_i_pre
348 | X += np.dot(res_i.T, res_i)
349 | return X
350 |
351 | def calcErrorCoeffConstLenRefImp(data, W, k = 0):
352 | p = (len(W[0]))/len(W)
353 | V = buildVRefImp(W)
354 | X = np.zeros([len(W), len(W)])
355 | z_pk = data[p+k:]
356 | WV = W
357 | #print z_pk
358 | for i in range(0, k+1):
359 | if i > 0:
360 | WV = np.dot(WV, V)
361 | zeta1_i = calcZetaDataRefImp(data, p, i)[k-i:]
362 | #print zeta1_i
363 | z_i_pre = np.dot(zeta1_i, WV.T)
364 | res_i = z_pk-z_i_pre
365 | X += np.dot(res_i.T, res_i)
366 | return X
367 |
368 | def calcErrorCoeffAffineRefImp(data, W_c, k = 0):
369 | p = (len(W_c[0])-1)/len(W_c)
370 | V_c = buildV_cRefImp(W_c)
371 | X = np.zeros([len(W_c), len(W_c)])
372 | #z_pk = data[p+k:]
373 | WV_c = W_c
374 | for i in range(0, k+1):
375 | if i > 0:
376 | WV_c = np.dot(WV_c, V_c)
377 | zeta1_i_c = calcZetacDataRefImp(data, p, i)
378 | z_i_pre = np.dot(zeta1_i_c, WV_c.T)
379 | res_i = data[p+i:]-z_i_pre
380 | X += np.dot(res_i.T, res_i)
381 | # if i == 1:
382 | # #print np.dot(zeta1_i_c.T, zeta1_i_c)#/(len(zeta1_i_c))
383 | # print np.dot(data[p+i:].T, zeta1_i_c)
384 | # print zeta1_i_c.mean(0)*(len(zeta1_i_c))
385 | # print data[p-1:-1].mean(0)*(len(data)-p)
386 | return X
387 |
388 | def calcErrorCoeffConstLenAffineRefImp(data, W_c, k = 0):
389 | p = (len(W_c[0])-1)/len(W_c)
390 | V_c = buildV_cRefImp(W_c)
391 | X = np.zeros([len(W_c), len(W_c)])
392 | z_pk = data[p+k:]
393 | WV_c = W_c
394 | #print z_pk
395 | for i in range(0, k+1):
396 | if i > 0:
397 | WV_c = np.dot(WV_c, V_c)
398 | zeta1_i_c = calcZetacDataRefImp(data, p, i)[k-i:]
399 | #print zeta1_i
400 | z_i_pre = np.dot(zeta1_i_c, WV_c.T)
401 | res_i = z_pk-z_i_pre
402 | X += np.dot(res_i.T, res_i)
403 | # if i == 1:
404 | # print np.dot(zeta1_i_c.T, zeta1_i_c)#/(len(zeta1_i_c))
405 | #print np.dot(data[p+i:].T, zeta1_i_c)
406 | # print zeta1_i_c.mean(0)*(len(zeta1_i_c))
407 | # print data[p-1:-1].mean(0)*(len(data)-p)
408 | return X
409 |
410 | def calcErrorCoeff(data, W, k = 0):
411 | p = len(W[0])/len(W)
412 | X = np.zeros([len(W), len(W)])
413 | corList = [np.dot(data[p:].T, data[p:])]
414 |
415 | for i in range(1, p+1):
416 | corList.append(np.dot(data[p:].T, data[p-i:-i]))
417 |
418 | zetaList = []
419 | lastLine = []
420 | zetaBlockList = []
421 | for i in range(p):
422 | zetaLine = []
423 | for j in range(p):
424 | if j < i:
425 | zetaLine.append(zetaList[j][i].T)
426 | else:
427 | if i == 0:
428 | zetaLine.append(corList[j]+np.outer(data[p-1], data[p-1-j])-np.outer(data[-1], data[-1-j]))
429 | else:
430 | zetaLine.append(lastLine[j-1]+np.outer(data[p-1-i], data[p-1-j])-np.outer(data[-1-i], data[-1-j]))
431 |
432 | zetaList.append(zetaLine)
433 | lastLine = zetaLine
434 |
435 | cov_p = np.dot(data[p:].T, data[p:])
436 | WV = W
437 | V = buildVRefImp(W)
438 | for i in range(k+1):
439 | if i > 0:
440 | WV = np.dot(WV, V)
441 | if i > 0:
442 | for j in range(i+1, len(corList)):
443 | corList[j] -= np.outer(data[p+i-1], data[p-j+i-1])
444 | corList.append(np.dot(data[p+i:].T, data[:-p-i]))
445 | zZi = np.hstack(corList[1+i:])
446 | if i == 0:
447 | for l in range(p):
448 | zetaBlockList.append(np.hstack(zetaList[l]))
449 | else:
450 | for l in range(p):
451 | for j in range(p):
452 | if j < l:
453 | zetaList[l][j] = zetaList[j][l].T
454 | else:
455 | zetaList[l][j] -= np.outer(data[-(i+1)-l], data[-(i+1)-j])
456 | for l in range(p):
457 | zetaBlockList[l] = np.hstack(zetaList[l])
458 | ZZi = np.vstack(zetaBlockList)
459 | K = np.dot(zZi, WV.T)
460 | X += cov_p -K-K.T + np.dot(np.dot(WV, ZZi), WV.T)
461 | if i < k:
462 | cov_p -= np.outer(data[p+i], data[p+i])
463 | return X
464 |
465 | def calcErrorCoeffConstLen(data, W, k = 0):
466 | p = len(W[0])/len(W)
467 | z_pk = data[p+k:]
468 | corList = [np.dot(z_pk.T, z_pk)]
469 | for i in range(1, k+p+1):
470 | corList.append(np.dot(z_pk.T, data[p+k-i:-i]))
471 | return calcErrorCoeffConstLenFromAutoCorrelations(W, corList, data[:p+k], data[-p-k:], k)
472 |
473 | def calcErrorCoeffConstLenFromAutoCorrelations(W, corList, startData, endData, k = 0):
474 | p = len(corList)-1-k
475 | zetaList = []
476 | lastLine = []
477 | for i in range(p):
478 | zetaLine = []
479 | for j in range(p):
480 | if j < i:
481 | zetaLine.append(zetaList[j][i].T)
482 | else:
483 | if i == 0:
484 | zetaLine.append(corList[j]-np.outer(endData[-1], endData[-1-j])+np.outer(startData[p+k-1], startData[p+k-1-j]))
485 | else:
486 | zetaLine.append(lastLine[j-1]-np.outer(endData[-1-i], endData[-1-j])+np.outer(startData[p+k-1-i], startData[p+k-1-j]))
487 | zetaList.append(zetaLine)
488 | lastLine = zetaLine
489 | for i in range(p):
490 | zetaList[i] = np.hstack(zetaList[i])
491 | ZZi = np.vstack(zetaList)
492 | return calcErrorCoeffConstLenFromCorrelations(W, np.hstack(corList[1:p+1]), ZZi, lastLine, corList, startData, endData, k)
493 |
494 | def calcErrorCoeffConstLenFromCorrelations(W, zZ, ZZ, lastLine, corList, startData, endData, k = 0):
495 | p = len(corList)-1-k
496 | WV = W
497 | V = buildVRefImp(W)
498 | Y = corList[0]*(k+1)
499 | pnl = len(W)*(p-1)
500 | zZi = zZ
501 | ZZi = ZZ
502 | for i in range(0, k+1):
503 | if i > 0:
504 | WV = np.dot(WV, V)
505 | zZi = np.hstack(corList[1+i:p+i+1])
506 | K = np.dot(zZi, WV.T)
507 | Y += -K-K.T + np.dot(np.dot(WV, ZZi), WV.T)
508 | if i < k:
509 | for j in range(p):
510 | lastLine[j] += -np.outer(endData[-p-i-1], endData[-j-2-i])+np.outer(startData[k-1-i], startData[p-i+k-j-2])
511 | line = np.hstack(lastLine)
512 | ZZi = np.vstack([np.hstack([ ZZi[len(W):].T[len(W):], line.T[:pnl]]), line])
513 | return Y
514 |
515 | #def calcErrorCoeffConstLenFromCorrelations2(W, zZ, ZZ, lastLine, corList, S, start_cor, end_cor, k = 0):
516 | #def calcErrorCoeffConstLenFromCorrelations2(W, zZ, ZZ, lastLine, corList, startData, endData, S, start_cor, end_cor, k = 0):
517 | def calcErrorCoeffConstLenFromCorrelations2(W, zZ, ZZ, lastLine, corList, S, start_cor, end_cor, k = 0):
518 | p = len(corList)-1-k
519 | WV = W
520 | V = buildVRefImp(W)
521 | Y = corList[0]*(k+1)
522 | pnl = len(W)*(p-1)
523 | zZi = zZ
524 | ZZi = ZZ
525 | for i in range(0, k+1):
526 | if i > 0:
527 | WV = np.dot(WV, V)
528 | zZi = np.hstack(corList[1+i:p+i+1])
529 | K = np.dot(zZi, WV.T)
530 | Y += -K-K.T + np.dot(np.dot(WV, ZZi), WV.T)
531 | if i < k:
532 | for j in range(p):
533 | #lastLine[j] += -np.outer(endData[-p-i-1], endData[-j-2-i])+np.outer(startData[-p-1-i], startData[-i-j-2])
534 | lastLine[j] += np.dot(S.T, np.dot(-end_cor[-p-i-1][-j-2-i]+start_cor[-p-1-i][-i-j-2], S))
535 | line = np.hstack(lastLine)
536 | ZZi = np.vstack([np.hstack([ ZZi[len(W):].T[len(W):], line.T[:pnl]]), line])
537 | return Y
538 |
539 | def calcErrorCoeffConstLenRoll(data, W, k = 0):
540 | p = len(W[0])/len(W)
541 | z_pk = data[p+k:]
542 | WV = W
543 | V = buildVRefImp(W)
544 | Y = np.dot(z_pk.T, z_pk)*(k+1)
545 | corList = [np.dot(z_pk.T, z_pk)]
546 | for i in range(1, k+p+1):
547 | corList.append(np.dot(z_pk.T, data[p+k-i:-i]))
548 |
549 | zetaList = []
550 | lastLine = []
551 | for i in range(p):
552 | zetaLine = []
553 | for j in range(p):
554 | if j < i:
555 | zetaLine.append(zetaList[j][i].T)
556 | else:
557 | if i == 0:
558 | zetaLine.append(corList[j]-np.outer(data[-1], data[-1-j])+np.outer(data[p+k-1], data[p+k-1-j]))
559 | else:
560 | zetaLine.append(lastLine[j-1]-np.outer(data[-1-i], data[-1-j])+np.outer(data[p+k-1-i], data[p+k-1-j]))
561 | zetaList.append(zetaLine)
562 | lastLine = zetaLine
563 | for i in range(p):
564 | zetaList[i] = np.hstack(zetaList[i])
565 | ZZi = np.vstack(zetaList)
566 |
567 | pnl = len(W)*(p-1)
568 | for i in range(0, k+1):
569 | if i > 0:
570 | WV = np.dot(WV, V)
571 | zZi = np.hstack(corList[1+i:p+i+1])
572 | K = np.dot(zZi, WV.T)
573 | Y += -K-K.T + np.dot(np.dot(WV, ZZi), WV.T)
574 | if i < k:
575 | ZZi = np.roll(np.roll(ZZi, len(W), 0), len(W), 1)
576 | for j in range(p):
577 | lastLine[j] += -np.outer(data[-p-i-1], data[-j-2-i])+np.outer(data[k-1-i], data[p-i+k-j-2])
578 | ZZi[pnl:, len(W)*j:len(W)*(j+1)] = lastLine[j]
579 | if j != p:
580 | ZZi[len(W)*j:len(W)*(j+1), pnl:] = lastLine[j].T
581 | # line = np.hstack(lastLine)
582 | # for j in range(p):
583 | # lastLine[j] += -np.outer(data[-p-i-1], data[-j-2-i])+np.outer(data[k-1-i], data[p-i+k-j-2])
584 | # line = np.hstack(lastLine)
585 | # ZZi = np.vstack([np.hstack([ ZZi[pnl:].T[pnl:], line.T[:pnl]]), line])
586 | # print "ZZi"
587 | # print ZZi
588 | # print ZZi2
589 |
590 | return Y
591 |
592 | def calcErrorCoeffAffine(data, W_c, k = 0):
593 | p = (len(W_c[0])-1)/len(W_c)
594 | X = np.zeros([len(W_c), len(W_c)])
595 | corList = [np.dot(data[p:].T, data[p:])]
596 |
597 | m = data[p:].mean(0)*(len(data)-p)
598 | meanList = [m]
599 |
600 | for i in range(1, p+1):
601 | corList.append(np.dot(data[p:].T, data[p-i:-i]))
602 | meanList.append(meanList[i-1] - data[-i]+data[p-i])
603 | #meanList[i] -= data[i-1]
604 | #print np.hstack(meanList[1:])
605 |
606 | zetaList = []
607 | lastLine = []
608 | zetaBlockList = []
609 | for i in range(p):
610 | zetaLine = []
611 | for j in range(p):
612 | if j < i:
613 | zetaLine.append(zetaList[j][i].T)
614 | else:
615 | if i == 0:
616 | zetaLine.append(corList[j]+np.outer(data[p-1], data[p-1-j])-np.outer(data[-1], data[-1-j]))
617 | else:
618 | zetaLine.append(lastLine[j-1]+np.outer(data[p-1-i], data[p-1-j])-np.outer(data[-1-i], data[-1-j]))
619 |
620 | zetaList.append(zetaLine)
621 | lastLine = zetaLine
622 |
623 | cov_p = np.dot(data[p:].T, data[p:])
624 | WV = W_c
625 | V_c = buildV_cRefImp(W_c)
626 | for i in range(k+1):
627 | if i > 0:
628 | WV = np.dot(WV, V_c)
629 | if i > 0:
630 | meanList[0] -= data[p+i-1]
631 | for j in range(i+1, len(corList)):
632 | corList[j] -= np.outer(data[p+i-1], data[p-j+i-1])
633 | meanList[j] -= data[p-j+i-1]
634 | corList.append(np.dot(data[p+i:].T, data[:-p-i]))
635 | #corList.append(np.dot(data[p:].T, data[p-i:-i]))
636 | #meanList.append(meanList[-1] - data[-p-i])
637 | meanList.append(data[:-p-i].mean(0)*len(data[:-p-i])) #todo: calc this from meanList[-1]
638 |
639 | zZi_c = np.vstack([np.hstack(corList[1+i:]).T, meanList[0]]).T
640 | if i == 0:
641 | for l in range(p):
642 | zetaBlockList.append(np.hstack(zetaList[l]))
643 | else:
644 | for l in range(p):
645 | for j in range(p):
646 | if j < l:
647 | zetaList[l][j] = zetaList[j][l].T
648 | else:
649 | zetaList[l][j] -= np.outer(data[-(i+1)-l], data[-(i+1)-j])
650 | for l in range(p):
651 | zetaBlockList[l] = np.hstack(zetaList[l])
652 | ZZi = np.vstack(zetaBlockList)
653 | ml = np.hstack(meanList[1+i:])
654 | ZZi_c = np.vstack([ZZi, ml]).T
655 | ml = np.hstack([ml, [len(data)-p-i]])
656 | ZZi_c = np.vstack([ZZi_c, ml])
657 | # if i == 1:
658 | # print zZi_c
659 | K = np.dot(zZi_c, WV.T)
660 | X += cov_p -K-K.T + np.dot(np.dot(WV, ZZi_c), WV.T)
661 | if i < k:
662 | cov_p -= np.outer(data[p+i], data[p+i])
663 | return X
664 |
665 | def calcErrorCoeffConstLenAffine(data, W_c, k = 0):
666 | p = (len(W_c[0])-1)/len(W_c)
667 | z_pk = data[p+k:]
668 | corList = [np.dot(z_pk.T, z_pk)]
669 | meanList = [z_pk.mean(0)*len(z_pk)]
670 | for i in range(1, k+p+1):
671 | corList.append(np.dot(z_pk.T, data[p+k-i:-i]))
672 | meanList.append(meanList[-1]-data[-i]+data[p+k-i])
673 | return calcErrorCoeffConstLenAffineFromAutoCorrelations(W_c, corList, meanList, len(data), data[:p+k], data[-p-k:], k)
674 |
675 | def calcErrorCoeffConstLenAffineFromAutoCorrelations(W_c, corList, meanList, dataLen, startData, endData, k = 0):
676 | zetaList = []
677 | lastLine = []
678 | p = len(corList)-1-k
679 | for i in range(p):
680 | zetaLine = []
681 | for j in range(p):
682 | if j < i:
683 | zetaLine.append(zetaList[j][i].T)
684 | else:
685 | if i == 0:
686 | zetaLine.append(corList[j]-np.outer(endData[-1], endData[-1-j])+np.outer(startData[p+k-1], startData[p+k-1-j]))
687 | else:
688 | zetaLine.append(lastLine[j-1]-np.outer(endData[-1-i], endData[-1-j])+np.outer(startData[p+k-1-i], startData[p+k-1-j]))
689 | zetaList.append(zetaLine)
690 | lastLine = zetaLine
691 | for i in range(p):
692 | zetaList[i] = np.hstack(zetaList[i])
693 | ml = np.hstack(meanList[1:p+1])
694 | print ml
695 | ZZi_c = np.vstack([np.vstack(zetaList), ml])
696 | ml1 = np.hstack([ml, [dataLen-p-k]])
697 | ZZi_c = np.vstack([ZZi_c.T, ml1]).T
698 | return calcErrorCoeffConstLenAffineFromCorrelations(W_c, np.vstack([np.hstack(corList[1:p+1]).T, meanList[0]]).T, ZZi_c, lastLine, corList, meanList, dataLen, startData, endData, k)
699 |
700 | def calcErrorCoeffConstLenAffineFromCorrelations(W_c, zZ_c, ZZ_c, lastLine, corList, meanList, dataLen, startData, endData, k = 0):
701 | #print ZZi_c
702 | p = len(corList)-1-k
703 | WV = W_c
704 | V_c = buildV_cRefImp(W_c)
705 | #Y = np.dot(z_pk.T, z_pk)*(k+1)
706 | Y = corList[0]*(k+1)
707 | pnl = len(W_c)*(p-1)
708 | ZZi_c = ZZ_c
709 | zZi_c = zZ_c
710 | for i in range(0, k+1):
711 | if i > 0:
712 | WV = np.dot(WV, V_c)
713 | zZi_c = np.vstack([np.hstack(corList[1+i:p+i+1]).T, meanList[0]]).T
714 | # if i == 1:
715 | # print ZZi_c
716 | K = np.dot(zZi_c, WV.T)
717 | Y += -K-K.T + np.dot(np.dot(WV, ZZi_c), WV.T)
718 | if i < k:
719 | for j in range(p):
720 | lastLine[j] += -np.outer(endData[-p-i-1], endData[-j-2-i])+np.outer(startData[k-1-i], startData[p-i+k-j-2])
721 | line = np.hstack(lastLine)
722 | ZZi_c = np.vstack([np.hstack([ ZZi_c[len(W_c):-1].T[len(W_c):-1], line.T[:pnl]]), line])
723 | ml = np.hstack(meanList[i+2:p+i+2])
724 | ZZi_c = np.vstack([ZZi_c, ml])
725 | ml1 = np.hstack([ml, [dataLen-p-k]])
726 | ZZi_c = np.vstack([ZZi_c.T, ml1]).T
727 | return Y
728 |
729 | #def calcErrorCoeffConstLenAffineFromCorrelations2(W_c, zZ_c, ZZ_c, lastLine, corList, meanList, dataLen, startData, endData, k = 0):
730 | def calcErrorCoeffConstLenAffineFromCorrelations2(W_c, zZ_c, ZZ_c, lastLine, corList, meanList, dataLen, S, start_cor, end_cor, k = 0):
731 | #print ZZi_c
732 | p = len(corList)-1-k
733 | WV = W_c
734 | V_c = buildV_cRefImp(W_c)
735 | #Y = np.dot(z_pk.T, z_pk)*(k+1)
736 | Y = corList[0]*(k+1)
737 | pnl = len(W_c)*(p-1)
738 | ZZi_c = ZZ_c
739 | zZi_c = zZ_c
740 | for i in range(0, k+1):
741 | if i > 0:
742 | WV = np.dot(WV, V_c)
743 | zZi_c = np.vstack([np.hstack(corList[1+i:p+i+1]).T, meanList[0]]).T
744 | # if i == 1:
745 | # print ZZi_c
746 | K = np.dot(zZi_c, WV.T)
747 | Y += -K-K.T + np.dot(np.dot(WV, ZZi_c), WV.T)
748 | if i < k:
749 | for j in range(p):
750 | #lastLine[j] += -np.outer(endData[-p-i-1], endData[-j-2-i])+np.outer(startData[-p-1-i], startData[-i-j-2])
751 | lastLine[j] += np.dot(S.T, np.dot(-end_cor[-p-i-1][-j-2-i]+start_cor[-p-1-i][-i-j-2], S))
752 | line = np.hstack(lastLine)
753 | #print pnl
754 | #print len(W_c)
755 | #print ZZi_c[len(W_c):-1].T[len(W_c):-1].shape
756 | #print line.T[:pnl].shape
757 | a = np.hstack([ ZZi_c[len(W_c):-1].T[len(W_c):-1], line.T[:pnl]])
758 | ZZi_c = np.vstack([a, line])
759 | ml = np.hstack(meanList[i+2:p+i+2])
760 | ZZi_c = np.vstack([ZZi_c, ml])
761 | ml1 = np.hstack([ml, [dataLen-p-k]])
762 | ZZi_c = np.vstack([ZZi_c.T, ml1]).T
763 | return Y
764 |
765 | def calcErrorCoeffConstLenAffineRoll(data, W_c, k = 0):
766 | p = (len(W_c[0])-1)/len(W_c)
767 | z_pk = data[p+k:]
768 | WV = W_c
769 | V_c = buildV_cRefImp(W_c)
770 | Y = np.dot(z_pk.T, z_pk)*(k+1)
771 | corList = [np.dot(z_pk.T, z_pk)]
772 | meanList = [z_pk.mean(0)*len(z_pk)]
773 | for i in range(1, k+p+1):
774 | corList.append(np.dot(z_pk.T, data[p+k-i:-i]))
775 | meanList.append(meanList[-1]-data[-i]+data[p+k-i])
776 |
777 | zetaList = []
778 | lastLine = []
779 | for i in range(p):
780 | zetaLine = []
781 | for j in range(p):
782 | if j < i:
783 | zetaLine.append(zetaList[j][i].T)
784 | else:
785 | if i == 0:
786 | zetaLine.append(corList[j]-np.outer(data[-1], data[-1-j])+np.outer(data[p+k-1], data[p+k-1-j]))
787 | else:
788 | zetaLine.append(lastLine[j-1]-np.outer(data[-1-i], data[-1-j])+np.outer(data[p+k-1-i], data[p+k-1-j]))
789 | zetaList.append(zetaLine)
790 | lastLine = zetaLine
791 | for i in range(p):
792 | zetaList[i] = np.hstack(zetaList[i])
793 | ml = np.hstack(meanList[1:p+1])
794 | # print ml
795 | ZZi_c = np.vstack([np.vstack(zetaList), ml])
796 | ml1 = np.hstack([ml, [len(z_pk)]])
797 | ZZi_c = np.vstack([ZZi_c.T, ml1]).T
798 |
799 | #print ZZi_c
800 |
801 | pnl = len(W_c)*(p-1)
802 | for i in range(0, k+1):
803 | if i > 0:
804 | WV = np.dot(WV, V_c)
805 | zZi_c = np.vstack([np.hstack(corList[1+i:p+i+1]).T, meanList[0]]).T
806 | # if i == 1:
807 | # print ZZi_c
808 | K = np.dot(zZi_c, WV.T)
809 | Y += -K-K.T + np.dot(np.dot(WV, ZZi_c), WV.T)
810 | if i < k:
811 | ZZi_c = np.roll(np.roll(ZZi_c[:-1, :-1], len(W_c), 0), len(W_c), 1)
812 | for j in range(p):
813 | lastLine[j] += -np.outer(data[-p-i-1], data[-j-2-i])+np.outer(data[k-1-i], data[p-i+k-j-2])
814 | ZZi_c[pnl:, len(W_c)*j:len(W_c)*(j+1)] = lastLine[j]
815 | if j != p:
816 | ZZi_c[len(W_c)*j:len(W_c)*(j+1), pnl:] = lastLine[j].T
817 | # ZZi_c2[-1:, len(W_c)*j:len(W_c)*(j+1)] = meanList[i+2+j]
818 |
819 | # for j in range(p):
820 | # lastLine[j] += -np.outer(data[-p-i-1], data[-j-2-i])+np.outer(data[k-1-i], data[p-i+k-j-2])
821 | # line = np.hstack(lastLine)
822 | # ZZi_c = np.vstack([np.hstack([ ZZi_c[pnl:-1].T[pnl:-1], line.T[:pnl]]), line])
823 | ml = np.hstack(meanList[i+2:p+i+2])
824 | ZZi_c = np.vstack([ZZi_c, ml])
825 | ml1 = np.hstack([ml, [len(z_pk)]])
826 | ZZi_c = np.vstack([ZZi_c.T, ml1]).T
827 |
828 | # print "ZZi_c"
829 | # print ZZi_c
830 | # print ZZi_c2
831 | return Y
832 |
833 | def createTestData():
834 | #data = np.outer(range(0, 9), [1, 1.1, 1.11])
835 | data = np.outer(range(0, 12), [1, 2.0])#, 3.0])
836 | #place some ouliers:
837 | #data[3][1] = 1.0
838 | #data[5][2] = 1.0
839 | data[5][1] = 1.0
840 | data[4][1] = -1.0
841 | data[7][0] = -2.0
842 | data[7][0] = -3.0
843 | data[10][0] = -4.5
844 | data[11][1] = -6.0
845 | return data
846 |
847 | def PFAReferenceImplementationTest():
848 | data = createTestData()
849 | p = 2
850 | print data
851 | print "------------------Sphering------------------"
852 | #Sphering:
853 | mean, S, z = calcSpheringParametersAndDataRefImp(data)#, threshold = 0.0000001, offset = 0, length = -1, besselsCorrection = 0)
854 | #Test Sphering:
855 | print z
856 | print np.multiply(np.dot(z.T, z), 1.0/len(data))
857 |
858 | print "------------------Fitting------------------"
859 | #Fitting:
860 | zeta = calcZetaDataRefImp(z, p)
861 | z_p = z[p:]
862 | W = calcRegressionCoeffRefImp(z, p)
863 | #Test fitting:
864 | #print W
865 | pre = np.dot(zeta, W.T) #np.dot(W, zeta.T).T
866 | res = z_p-pre
867 | print np.trace(np.dot(res.T, res))/len(res)
868 | print empiricalRawErrorRefImp(z, W)
869 | errComp = empiricalRawErrorComponentsRefImp(z, W)
870 | print errComp
871 | print np.sum(errComp)
872 |
873 | print "------------------PCA on Error Covariance------------------"
874 | r = 1
875 | #PCA on error covariance:
876 | X = np.dot(res.T, res)
877 | print X
878 | Ar = calcExtractionForErrorCovRefImp(X, r)
879 | m = np.dot(z, Ar)
880 | Wm = calcRegressionCoeffRefImp(m, p)
881 | print m
882 | errComp_m = empiricalRawErrorComponentsRefImp(m, Wm)
883 | print errComp_m
884 | print np.sum(errComp_m)
885 | print np.dot(m.T, m)/len(m)
886 |
887 | print "--------------------------------------------------"
888 | print "------------------Affine variant------------------"
889 | print "--------------------------------------------------"
890 | print "------------------Sphering------------------"
891 | print "like before"
892 |
893 | print "------------------Fitting------------------"
894 | #Fitting:
895 | zeta_c = calcZetacDataRefImp(z, p)
896 | # print zeta_c
897 |
898 | # z_p = z[p:]
899 | W_c = calcRegressionCoeffAffineRefImp(z, p)
900 | #Test fitting:
901 | #print W
902 | pre_c = np.dot(zeta_c, W_c.T) #np.dot(W, zeta.T).T
903 | res_c = z_p-pre_c
904 | print np.trace(np.dot(res_c.T, res_c))/len(res_c)
905 | print empiricalRawErrorAffineRefImp(z, W_c)
906 | errComp_c = empiricalRawErrorComponentsAffineRefImp(z, W_c)
907 | print errComp_c
908 | print np.sum(errComp_c)
909 | print W_c
910 |
911 | print "------------------Fitting Sinus------------------"
912 | sig_x = lambda t: [np.sin(t)]
913 | ts = range(0, 50)
914 | sin_x = np.array([sig_x(t) for t in ts])
915 | #print sin_x
916 | sin_mean, sin_S, sin_z = calcSpheringParametersAndDataRefImp(sin_x)#, threshold = 0.0000001, offset = 0, length = -1, besselsCorrection = 0)
917 | #Test Sphering:
918 | #print sin_z
919 | print "sin_z cov:"
920 | print np.multiply(np.dot(sin_z.T, sin_z), 1.0/len(sin_z))
921 |
922 | sin_p = 2
923 | sin_zeta_c = calcZetacDataRefImp(sin_z, sin_p)
924 | # print zeta_c
925 |
926 | sin_z_p = sin_z[sin_p:]
927 | sin_W_c = calcRegressionCoeffAffineRefImp(sin_z, sin_p)
928 | #Test fitting:
929 | #print W
930 | sin_pre_c = np.dot(sin_zeta_c, sin_W_c.T) #np.dot(W, zeta.T).T
931 | sin_res_c = sin_z_p-sin_pre_c
932 | print "sin_z, sin_W_c error:"
933 | print np.trace(np.dot(sin_res_c.T, sin_res_c))/len(sin_res_c)
934 | print empiricalRawErrorAffineRefImp(sin_z, sin_W_c)
935 | sin_errComp_c = empiricalRawErrorComponentsAffineRefImp(sin_z, sin_W_c)
936 | print sin_errComp_c
937 | print np.sum(sin_errComp_c)
938 | print "fitted sin W_c:"
939 | print sin_W_c
940 | print "analytic sin W_c:"
941 | cs = np.cos(1.0)
942 | sin_W_c_2 = np.array([[2*cs, -1.0, (-2.0+2.0*cs)*sin_S[0][0]*sin_mean[0]]])
943 | print empiricalRawErrorAffineRefImp(sin_z, sin_W_c_2)
944 | print sin_W_c_2
945 |
946 | print "------------------PCA on Error Covariance------------------"
947 | r_c = 1
948 | #PCA on error covariance:
949 | X_c = np.dot(res_c.T, res_c)
950 | print X_c
951 | Ar_c = calcExtractionForErrorCovRefImp(X_c, r_c)
952 | m_c = np.dot(z, Ar_c)
953 | Wm_c = calcRegressionCoeffAffineRefImp(m_c, p)
954 | print m_c
955 | print empiricalRawErrorAffineRefImp(m_c, Wm_c)
956 | print np.dot(m_c.T, m_c)/len(m_c)
957 |
958 | print "------------------PCA on Error Covariance - Noisy sine------------------"
959 | sig_x2 = lambda t: [np.sin(t), np.random.rand(1)[0]]
960 | #ts = range(0, 250)
961 | nsin_x = np.array([sig_x2(t) for t in ts])
962 | nsin_mean, nsin_S, nsin_z = calcSpheringParametersAndDataRefImp(nsin_x)#, threshold = 0.0000001, offset = 0, length = -1, besselsCorrection = 0)
963 | #Test Sphering:
964 | #print sin_z
965 | print "nsin_z cov:"
966 | print np.multiply(np.dot(nsin_z.T, nsin_z), 1.0/len(nsin_z))
967 |
968 | nsin_p = 2
969 | nsin_zeta_c = calcZetacDataRefImp(nsin_z, nsin_p)
970 | ## print zeta_c
971 |
972 | nsin_z_p = nsin_z[nsin_p:]
973 | nsin_W_c = calcRegressionCoeffAffineRefImp(nsin_z, nsin_p)
974 | #Test fitting:
975 | #print W
976 | nsin_pre_c = np.dot(nsin_zeta_c, nsin_W_c.T) #np.dot(W, zeta.T).T
977 | nsin_res_c = nsin_z_p-nsin_pre_c
978 | print "nsin_z, nsin_W_c error:"
979 | print np.trace(np.dot(nsin_res_c.T, nsin_res_c))/len(nsin_res_c)
980 | print empiricalRawErrorAffineRefImp(nsin_z, nsin_W_c)
981 | nsin_errComp_c = empiricalRawErrorComponentsAffineRefImp(nsin_z, nsin_W_c)
982 | print nsin_errComp_c
983 | print np.sum(nsin_errComp_c)
984 | # print "fitted sin W_c:"
985 | # print nsin_W_c
986 | # print "analytic sin W_c:"
987 | # cs = np.cos(1.0)
988 | # nsin_W_c_2 = np.array([[2*cs, -1.0, (-2.0+2.0*cs)*nsin_S[0][0]*nsin_mean[0]]])
989 | # print empiricalRawErrorAffine(nsin_z, nsin_W_c_2)
990 | # print nsin_W_c_2
991 |
992 | #print "------------------PCA on Error Covariance------------------"
993 | nsin_r_c = 1
994 | #PCA on error covariance:
995 | nsin_X_c = np.dot(nsin_res_c.T, nsin_res_c)
996 | print nsin_X_c
997 | nsin_Ar_c = calcExtractionForErrorCovRefImp(nsin_X_c, nsin_r_c)
998 | nsin_m_c = np.dot(nsin_z, nsin_Ar_c)
999 | nsin_Wm_c = calcRegressionCoeffAffineRefImp(nsin_m_c, nsin_p)
1000 | #print nsin_m_c
1001 | print empiricalRawErrorAffineRefImp(nsin_m_c, nsin_Wm_c)
1002 | print np.dot(nsin_m_c.T, nsin_m_c)/len(nsin_m_c)
1003 | print "fitted nsin Wm_c:"
1004 | print nsin_Wm_c
1005 | print "analytic nsin Wm_c:"
1006 | SI = LA.inv(nsin_S.T)
1007 | factor = LA.norm(np.dot(np.array([1.0, 0]), SI))
1008 | Ar2 = np.multiply(np.dot(np.array([1.0, 0]), SI), 1.0/factor)
1009 | nsin_Wm_c_2 = np.array([[2*cs, -1.0, (-2.0+2.0*cs)*(1.0/factor)*nsin_mean[0]]])
1010 | print empiricalRawErrorAffineRefImp(nsin_m_c, nsin_Wm_c_2)
1011 | print nsin_Wm_c_2
1012 |
1013 | print ""
1014 | print " _____________________________________________"
1015 | print "/ \\"
1016 | print "|---------------iteration stuff---------------|"
1017 | print "\\_____________________________________________/"
1018 |
1019 | print ""
1020 | #print zeta
1021 | zeta0 = calcZeta0DataRefImp(z, p)
1022 | #print zeta0
1023 | ZZ0 = np.dot(zeta0.T, zeta)
1024 | zZ = np.dot(z_p.T, zeta)
1025 | ZZ = np.dot(zeta.T, zeta)
1026 | ZZI = invertByProjectionRefImp(ZZ)
1027 | W0 = np.dot(zZ, ZZI)
1028 | V = np.dot(ZZ0, ZZI)
1029 | print W0
1030 | print W
1031 | print ""
1032 | print V
1033 | print buildVRefImp(W)
1034 |
1035 | print ""
1036 | print np.dot(ZZ, ZZI)
1037 |
1038 | print "calc iterated error coeff:"
1039 | X_k = calcErrorCoeffRefImp(z, W, 1)
1040 | Y_k = calcErrorCoeffConstLenRefImp(z, W, 1)
1041 | print X
1042 | print X_k
1043 | print Y_k
1044 |
1045 | print "-------------affine version---------------"
1046 | print ""
1047 | #print zeta
1048 | zeta0_c = calcZeta0cDataRefImp(z, p)
1049 | # print zeta0_c
1050 | # print zeta_c
1051 | ZZ0_c = np.dot(zeta0_c.T, zeta_c)
1052 | zZ_c = np.dot(z_p.T, zeta_c)
1053 | ZZ_c = np.dot(zeta_c.T, zeta_c)
1054 | ZZI_c = invertByProjectionRefImp(ZZ_c)
1055 | W0_c = np.dot(zZ_c, ZZI_c)
1056 | #The following idea of V_c is cumbersome, because it
1057 | #puts error on the constant-one component of zeta_c,
1058 | #if this reduces the overall error.
1059 | #But the constant component may not be compromized in any case, so
1060 | #this way to compute V_c is misleading:
1061 | # V_c_cumb = np.dot(ZZ0_c, ZZI_c)
1062 | # print V_c_cumb
1063 |
1064 | print W0_c
1065 | print W_c
1066 | print ""
1067 | #Better way:
1068 | V_c = buildV_cRefImp(W_c)
1069 | print V_c
1070 | print "calc iterated error coeff:"
1071 | X_k_c = calcErrorCoeffAffineRefImp(z, W_c, 1)
1072 | Y_k_c = calcErrorCoeffConstLenAffineRefImp(z, W_c, 1)
1073 | print X_k
1074 | print X_k_c
1075 | print Y_k
1076 | print Y_k_c
1077 |
1078 | #todo: build unit-test from this:
1079 | def compareSphering():
1080 | data = createTestData()
1081 | p = 2
1082 | print data
1083 | print "------------------Sphering------------------"
1084 | #Sphering:
1085 | meanRef, SRef, zRef = calcSpheringParametersAndDataRefImp(data, besselsCorrection = 1)#, threshold = 0.0000001, offset = 0, length = -1, besselsCorrection = 0)
1086 | #Test Sphering:
1087 | #print z
1088 | #print np.multiply(np.dot(z.T, z), 1.0/len(data))
1089 | print SRef
1090 | print meanRef
1091 |
1092 | #print data
1093 | mean = data.mean(0)
1094 | secondMoment = np.dot(data.T, data)
1095 | print (secondMoment/len(data)-np.outer(mean, mean))
1096 | S = calcSpheringMatrixFromMeanAndSecondMoment(mean, secondMoment, len(data), threshold = 0.0000001, besselsCorrection = 1)
1097 | print S
1098 | data0 = data-np.outer(np.ones(len(data)), mean)
1099 | z = np.dot(data0, S)
1100 | print zRef
1101 | print z
1102 |
1103 | def compareFitting():
1104 | data = createTestData()
1105 | p = 2
1106 | print data
1107 | print "------------------Sphering------------------"
1108 | #Sphering:
1109 | meanRef, SRef, zRef = calcSpheringParametersAndDataRefImp(data, besselsCorrection = 0)#, threshold = 0.0000001, offset = 0, length = -1, besselsCorrection = 0)
1110 | print "------------------Fitting------------------"
1111 | #Fitting:
1112 | zeta = calcZetaDataRefImp(zRef, p)
1113 | z_pRef = zRef[p:]
1114 | WRef = calcRegressionCoeffRefImp(zRef, p)
1115 | #print WRef
1116 | ZZRef = np.dot(zeta.T, zeta)
1117 | print ZZRef
1118 | Z11Ref = np.dot(zRef[p-1:-1].T, zRef[p-1:-1])
1119 | Z12Ref = np.dot(zRef[p-1:-1].T, zRef[p-2:-2])
1120 | Z22Ref = np.dot(zRef[p-2:-2].T, zRef[p-2:-2])
1121 | print Z11Ref
1122 | #print Z12Ref
1123 | #print Z22Ref
1124 | print ""
1125 | mean = data.mean(0)
1126 | secondMoment = np.dot(data.T, data)
1127 | S = calcSpheringMatrixFromMeanAndSecondMoment(mean, secondMoment, len(data), threshold = 0.0000001, besselsCorrection = 0)
1128 | #x1 = data[p-1:-1]
1129 | #x2 = data[p-2:-2]
1130 | #X11 = np.dot(x1.T, x1)
1131 | X00 = secondMoment-np.dot(data[0:p].T, data[0:p])
1132 | X11 = X00-np.outer(data[-1], data[-1])+np.outer(data[p-1], data[p-1])
1133 | #X12 = np.dot(x1.T, x2)
1134 | X12 = np.dot(data[p-1:-1].T, data[p-2:-2])
1135 | #X22 = np.dot(x2.T, x2)
1136 | X22 = X11-np.outer(data[-2], data[-2])+np.outer(data[p-2], data[p-2])
1137 |
1138 | M = np.outer(mean, mean)*(len(data)-p)
1139 | x0 = (mean*len(data)-data[0:p].mean(0)*p)
1140 | xm1 = (x0 - data[-1]+data[p-1])
1141 | xm2 = (xm1 - data[-2]+data[p-2])
1142 | #xm1 = data[p-1:-1].mean(0)*(len(data)-p)
1143 | #xm2 = data[p-2:-2].mean(0)*(len(data)-p)
1144 |
1145 | #z1_ = x1 - np.outer(np.ones(len(x1)), meanRef)
1146 | #print np.dot(z1_, SRef)
1147 | print Z11Ref
1148 | # print np.dot(np.dot(x1 - np.outer(np.ones(len(x1)), meanRef), SRef).T, np.dot(x1 - np.outer(np.ones(len(x1)), meanRef), SRef))
1149 | # print np.dot(np.dot(SRef.T, (x1 - np.outer(np.ones(len(x1)), meanRef)).T), np.dot(x1 - np.outer(np.ones(len(x1)), meanRef), SRef))
1150 |
1151 | #K = np.dot((x1 - np.outer(np.ones(len(x1)), meanRef)).T, x1 - np.outer(np.ones(len(x1)), meanRef))
1152 | K11 = X11 +( - np.outer(xm1, mean) - np.outer(mean, xm1) + M)
1153 | print np.dot(S.T, np.dot(K11, S))
1154 | print ""
1155 | print Z12Ref
1156 | K12 = X12 +( - np.outer(xm1, mean) - np.outer(mean, xm2) + M)
1157 | print np.dot(S.T, np.dot(K12, S))
1158 | print ""
1159 | print Z22Ref
1160 | K22 = X22 +( - np.outer(xm2, mean) - np.outer(mean, xm2) + M)
1161 | print np.dot(S.T, np.dot(K22, S))
1162 | print ""
1163 |
1164 | # print ZZRef
1165 | # print ZZRef[2:4].T[2:4]
1166 |
1167 | # print np.dot(x1.T, np.outer(np.ones(len(x1)), meanRef))
1168 | # print np.outer(x1.mean(0)*len(x1), meanRef)
1169 |
1170 | # print np.dot(np.outer(np.ones(len(x1)), meanRef).T, x1)
1171 | # print np.outer(meanRef, x1.mean(0)*len(x1))
1172 |
1173 | # print np.dot(np.outer(np.ones(len(x1)), meanRef).T, np.outer(np.ones(len(x1)), meanRef))
1174 | # print np.outer(meanRef, meanRef)*len(x1)
1175 |
1176 | # x2 = data[p-2:-2]
1177 | # X11 = np.dot(x1.T, x1)
1178 | # X12 = np.dot(x1.T, x2)
1179 | # X22 = np.dot(x2.T, x2)
1180 | # xm1 = x1.mean(0)
1181 | # xm2 = x2.mean(0)
1182 | # K11 = np.outer(meanRef, meanRef) + (X11-np.outer(xm1, meanRef)-np.outer(meanRef, xm1))/(len(data)-p*1.0)
1183 | # Z11 = np.dot(SRef.T, np.dot(K11, SRef))
1184 | # print Z11
1185 |
1186 |
1187 | #Test fitting:
1188 | #print W
1189 | # pre = np.dot(zeta, W.T) #np.dot(W, zeta.T).T
1190 | # res = z_p-pre
1191 | # print np.trace(np.dot(res.T, res))/len(res)
1192 | # print empiricalRawErrorRefImp(z, W)
1193 | # errComp = empiricalRawErrorComponentsRefImp(z, W)
1194 | # print errComp
1195 | # print np.sum(errComp)
1196 |
1197 | def compareIteration():
1198 | data = createTestData()
1199 | p = 2
1200 | print data
1201 | print "------------------Sphering------------------"
1202 | #Sphering:
1203 | meanRef, SRef, zRef = calcSpheringParametersAndDataRefImp(data)#, threshold = 0.0000001, offset = 0, length = -1, besselsCorrection = 0)
1204 |
1205 | print "------------------Fitting------------------"
1206 | #Fitting:
1207 | z_pRef = zRef[p:]
1208 | zetaRef = calcZetaDataRefImp(zRef, p)
1209 | WRef = calcRegressionCoeffRefImp(zRef, p)
1210 |
1211 | print "------------------PCA on Error Covariance------------------"
1212 | r = 1
1213 | #PCA on error covariance:
1214 | # X = np.dot(res.T, res)
1215 | # print X
1216 | # Ar = calcExtractionForErrorCovRefImp(X, r)
1217 | # m = np.dot(z, Ar)
1218 | # Wm = calcRegressionCoeffRefImp(m, p)
1219 | # print m
1220 | # errComp_m = empiricalRawErrorComponentsRefImp(m, Wm)
1221 | # print errComp_m
1222 | # print np.sum(errComp_m)
1223 | # print np.dot(m.T, m)/len(m)
1224 | #
1225 | # print "--------------------------------------------------"
1226 | # print "------------------Affine variant------------------"
1227 | # print "--------------------------------------------------"
1228 | # print "------------------Sphering------------------"
1229 | # print "like before"
1230 | #
1231 | # print "------------------Fitting------------------"
1232 | # #Fitting:
1233 | # zeta_c = calcZetacDataRefImp(z, p)
1234 | ## print zeta_c
1235 | #
1236 | ## z_p = z[p:]
1237 | # W_c = calcRegressionCoeffAffineRefImp(z, p)
1238 | # #Test fitting:
1239 | # #print W
1240 | # pre_c = np.dot(zeta_c, W_c.T) #np.dot(W, zeta.T).T
1241 | # res_c = z_p-pre_c
1242 | # print np.trace(np.dot(res_c.T, res_c))/len(res_c)
1243 | # print empiricalRawErrorAffineRefImp(z, W_c)
1244 | # errComp_c = empiricalRawErrorComponentsAffineRefImp(z, W_c)
1245 | # print errComp_c
1246 | # print np.sum(errComp_c)
1247 | # print W_c
1248 | #
1249 | # print "------------------Fitting Sinus------------------"
1250 | # sig_x = lambda t: [np.sin(t)]
1251 | # ts = range(0, 50)
1252 | # sin_x = np.array([sig_x(t) for t in ts])
1253 | # #print sin_x
1254 | # sin_mean, sin_S, sin_z = calcSpheringParametersAndDataRefImp(sin_x)#, threshold = 0.0000001, offset = 0, length = -1, besselsCorrection = 0)
1255 | # #Test Sphering:
1256 | # #print sin_z
1257 | # print "sin_z cov:"
1258 | # print np.multiply(np.dot(sin_z.T, sin_z), 1.0/len(sin_z))
1259 | #
1260 | # sin_p = 2
1261 | # sin_zeta_c = calcZetacDataRefImp(sin_z, sin_p)
1262 | ## print zeta_c
1263 | #
1264 | # sin_z_p = sin_z[sin_p:]
1265 | # sin_W_c = calcRegressionCoeffAffineRefImp(sin_z, sin_p)
1266 | # #Test fitting:
1267 | # #print W
1268 | # sin_pre_c = np.dot(sin_zeta_c, sin_W_c.T) #np.dot(W, zeta.T).T
1269 | # sin_res_c = sin_z_p-sin_pre_c
1270 | # print "sin_z, sin_W_c error:"
1271 | # print np.trace(np.dot(sin_res_c.T, sin_res_c))/len(sin_res_c)
1272 | # print empiricalRawErrorAffineRefImp(sin_z, sin_W_c)
1273 | # sin_errComp_c = empiricalRawErrorComponentsAffineRefImp(sin_z, sin_W_c)
1274 | # print sin_errComp_c
1275 | # print np.sum(sin_errComp_c)
1276 | # print "fitted sin W_c:"
1277 | # print sin_W_c
1278 | # print "analytic sin W_c:"
1279 | # cs = np.cos(1.0)
1280 | # sin_W_c_2 = np.array([[2*cs, -1.0, (-2.0+2.0*cs)*sin_S[0][0]*sin_mean[0]]])
1281 | # print empiricalRawErrorAffineRefImp(sin_z, sin_W_c_2)
1282 | # print sin_W_c_2
1283 | #
1284 | # print "------------------PCA on Error Covariance------------------"
1285 | # r_c = 1
1286 | # #PCA on error covariance:
1287 | # X_c = np.dot(res_c.T, res_c)
1288 | # print X_c
1289 | # Ar_c = calcExtractionForErrorCovRefImp(X_c, r_c)
1290 | # m_c = np.dot(z, Ar_c)
1291 | # Wm_c = calcRegressionCoeffAffineRefImp(m_c, p)
1292 | # print m_c
1293 | # print empiricalRawErrorAffineRefImp(m_c, Wm_c)
1294 | # print np.dot(m_c.T, m_c)/len(m_c)
1295 | #
1296 | # print "------------------PCA on Error Covariance - Noisy sine------------------"
1297 | # sig_x2 = lambda t: [np.sin(t), np.random.rand(1)[0]]
1298 | # #ts = range(0, 250)
1299 | # nsin_x = np.array([sig_x2(t) for t in ts])
1300 | # nsin_mean, nsin_S, nsin_z = calcSpheringParametersAndDataRefImp(nsin_x)#, threshold = 0.0000001, offset = 0, length = -1, besselsCorrection = 0)
1301 | # #Test Sphering:
1302 | # #print sin_z
1303 | # print "nsin_z cov:"
1304 | # print np.multiply(np.dot(nsin_z.T, nsin_z), 1.0/len(nsin_z))
1305 | #
1306 | # nsin_p = 2
1307 | # nsin_zeta_c = calcZetacDataRefImp(nsin_z, nsin_p)
1308 | ### print zeta_c
1309 | #
1310 | # nsin_z_p = nsin_z[nsin_p:]
1311 | # nsin_W_c = calcRegressionCoeffAffineRefImp(nsin_z, nsin_p)
1312 | # #Test fitting:
1313 | # #print W
1314 | # nsin_pre_c = np.dot(nsin_zeta_c, nsin_W_c.T) #np.dot(W, zeta.T).T
1315 | # nsin_res_c = nsin_z_p-nsin_pre_c
1316 | # print "nsin_z, nsin_W_c error:"
1317 | # print np.trace(np.dot(nsin_res_c.T, nsin_res_c))/len(nsin_res_c)
1318 | # print empiricalRawErrorAffineRefImp(nsin_z, nsin_W_c)
1319 | # nsin_errComp_c = empiricalRawErrorComponentsAffineRefImp(nsin_z, nsin_W_c)
1320 | # print nsin_errComp_c
1321 | # print np.sum(nsin_errComp_c)
1322 | ## print "fitted sin W_c:"
1323 | ## print nsin_W_c
1324 | ## print "analytic sin W_c:"
1325 | ## cs = np.cos(1.0)
1326 | ## nsin_W_c_2 = np.array([[2*cs, -1.0, (-2.0+2.0*cs)*nsin_S[0][0]*nsin_mean[0]]])
1327 | ## print empiricalRawErrorAffine(nsin_z, nsin_W_c_2)
1328 | ## print nsin_W_c_2
1329 | #
1330 | # #print "------------------PCA on Error Covariance------------------"
1331 | # nsin_r_c = 1
1332 | # #PCA on error covariance:
1333 | # nsin_X_c = np.dot(nsin_res_c.T, nsin_res_c)
1334 | # print nsin_X_c
1335 | # nsin_Ar_c = calcExtractionForErrorCovRefImp(nsin_X_c, nsin_r_c)
1336 | # nsin_m_c = np.dot(nsin_z, nsin_Ar_c)
1337 | # nsin_Wm_c = calcRegressionCoeffAffineRefImp(nsin_m_c, nsin_p)
1338 | # #print nsin_m_c
1339 | # print empiricalRawErrorAffineRefImp(nsin_m_c, nsin_Wm_c)
1340 | # print np.dot(nsin_m_c.T, nsin_m_c)/len(nsin_m_c)
1341 | # print "fitted nsin Wm_c:"
1342 | # print nsin_Wm_c
1343 | # print "analytic nsin Wm_c:"
1344 | # SI = LA.inv(nsin_S.T)
1345 | # factor = LA.norm(np.dot(np.array([1.0, 0]), SI))
1346 | # Ar2 = np.multiply(np.dot(np.array([1.0, 0]), SI), 1.0/factor)
1347 | # nsin_Wm_c_2 = np.array([[2*cs, -1.0, (-2.0+2.0*cs)*(1.0/factor)*nsin_mean[0]]])
1348 | # print empiricalRawErrorAffineRefImp(nsin_m_c, nsin_Wm_c_2)
1349 | # print nsin_Wm_c_2
1350 | #
1351 | print ""
1352 | print " _____________________________________________"
1353 | print "/ \\"
1354 | print "|---------------iteration stuff---------------|"
1355 | print "\\_____________________________________________/"
1356 |
1357 | print ""
1358 | #print zeta
1359 | zeta0Ref = calcZeta0DataRefImp(zRef, p)
1360 | #print zeta0
1361 | ZZ0Ref = np.dot(zeta0Ref.T, zetaRef)
1362 | zZRef = np.dot(z_pRef.T, zetaRef)
1363 | ZZRef = np.dot(zetaRef.T, zetaRef)
1364 | ZZIRef = invertByProjectionRefImp(ZZRef)
1365 | W0Ref = np.dot(zZRef, ZZIRef)
1366 | VRef = np.dot(ZZ0Ref, ZZIRef)
1367 | # print W0Ref
1368 | # print WRef
1369 | # print ""
1370 | # print VRef
1371 | # print buildVRefImp(WRef)
1372 | #
1373 | # print ""
1374 | # print np.dot(ZZRef, ZZIRef)
1375 |
1376 | print "calc iterated error coeff:"
1377 | k = 4
1378 | X_kRef = calcErrorCoeffRefImp(zRef, WRef, k)
1379 | Y_kRef = calcErrorCoeffConstLenRefImp(zRef, WRef, k)
1380 | #print XRef
1381 | print X_kRef
1382 | #print Y_kRef
1383 |
1384 |
1385 | X = np.zeros([len(WRef), len(WRef)])
1386 | corList = [np.dot(zRef[p:].T, zRef[p:])]
1387 |
1388 | #for i in range(1, k+p+1):
1389 | for i in range(1, p+1): #macht scheinbar eins zu viel)
1390 | corList.append(np.dot(zRef[p:].T, zRef[p-i:-i]))
1391 |
1392 | zetaList = []
1393 | lastLine = []#corList
1394 | zetaBlockList = []
1395 | for i in range(p):
1396 | zetaLine = []
1397 | for j in range(p):
1398 | if j < i:
1399 | zetaLine.append(zetaList[j][i].T)
1400 | else:
1401 | if i == 0:
1402 | zetaLine.append(corList[j]+np.outer(zRef[p-1], zRef[p-1-j])-np.outer(zRef[-1], zRef[-1-j]))
1403 | else:
1404 | #zetaLine.append(lastLine[j-1]-np.outer(zRef[-2], zRef[-2])+np.outer(zRef[p-2], zRef[p-2]))
1405 | zetaLine.append(lastLine[j-1]+np.outer(zRef[p-1-i], zRef[p-1-j])-np.outer(zRef[-1-i], zRef[-1-j]))
1406 |
1407 | zetaList.append(zetaLine)
1408 | lastLine = zetaLine
1409 |
1410 | # for i in range(p):
1411 | # zetaBlockList.append(np.hstack(zetaList[i]))
1412 | # ZZ0 = np.vstack(zetaBlockList)
1413 | #
1414 | ## for j in range(p):
1415 | ## lastLine[j] -= np.outer(zRef[-p-1], zRef[-2-j])
1416 | # for i in range(p):
1417 | # for j in range(p):
1418 | # if j < i:
1419 | # zetaList[i][j] = zetaList[j][i].T
1420 | # else:
1421 | # zetaList[i][j] -= np.outer(zRef[-2-i], zRef[-2-j])
1422 | #
1423 | # #print np.hstack(lastLine)
1424 | # for i in range(p):
1425 | # zetaBlockList[i] = np.hstack(zetaList[i])
1426 | # ZZ1 = np.vstack(zetaBlockList)
1427 | #print ZZ1
1428 |
1429 | # for i in range(p):
1430 | # for j in range(p):
1431 | # if j < i:
1432 | # zetaList[i][j] = zetaList[j][i].T
1433 | # else:
1434 | # zetaList[i][j] -= np.outer(zRef[-3-i], zRef[-3-j])
1435 | #
1436 | # #print np.hstack(lastLine)
1437 | # for i in range(p):
1438 | # zetaBlockList[i] = np.hstack(zetaList[i])
1439 | # ZZ2 = np.vstack(zetaBlockList)
1440 | # print ZZ2
1441 |
1442 | #z_pk = data[p+k:]
1443 | cov_p = np.dot(zRef[p:].T, zRef[p:])
1444 | WV = WRef
1445 | for i in range(0, k+1):
1446 | if i > 0:
1447 | WV = np.dot(WV, VRef)
1448 | zeta1_i = calcZetaDataRefImp(zRef, p, i)
1449 | # print zRef
1450 | # print zeta1_i
1451 | #print len(zeta1_i)
1452 | #X += np.dot((zRef[p+i:]-z_i_pre).T, zRef[p+i:]-z_i_pre)
1453 | # X += np.dot(zRef[p+i:].T, zRef[p+i:]) - np.dot(z_i_pre.T, zRef[p+i:]) - np.dot(zRef[p+i:].T, z_i_pre) + np.dot(z_i_pre.T, z_i_pre)
1454 | zZiRef = np.dot(zRef[p+i:].T, zeta1_i)
1455 | if i > 0:
1456 | for j in range(i+1, len(corList)):
1457 | corList[j] -= np.outer(zRef[p+i-1], zRef[p-j+i-1])
1458 | corList.append(np.dot(zRef[p+i:].T, zRef[:-p-i])) #ist scheinbar falsch
1459 | zZi = np.hstack(corList[1+i:])#p+i+1])
1460 | # print "+++"
1461 | # print zZiRef
1462 | # print zZi
1463 | # print "==="
1464 | ZZiRef = np.dot(zeta1_i.T, zeta1_i)
1465 |
1466 | #print np.hstack(lastLine)
1467 | if i == 0:
1468 | for l in range(p):
1469 | zetaBlockList.append(np.hstack(zetaList[l]))
1470 | else:
1471 | for l in range(p):
1472 | for j in range(p):
1473 | if j < l:
1474 | zetaList[l][j] = zetaList[j][l].T
1475 | else:
1476 | zetaList[l][j] -= np.outer(zRef[-(i+1)-l], zRef[-(i+1)-j])
1477 | for l in range(p):
1478 | zetaBlockList[l] = np.hstack(zetaList[l])
1479 | ZZi = np.vstack(zetaBlockList)
1480 |
1481 | # print "i = "+str(i)
1482 | # print zZiRef
1483 | # print zZi
1484 | # print "========="
1485 |
1486 | K = np.dot(zZi, WV.T)
1487 | X += cov_p -K-K.T + np.dot(np.dot(WV, ZZi), WV.T)
1488 | cov_p -= np.outer(zRef[p+i], zRef[p+i])
1489 |
1490 | print X
1491 | print calcErrorCoeff(zRef, WRef, k)
1492 |
1493 | print "----------const len-----------"
1494 | #Y = np.zeros([len(WRef), len(WRef)])
1495 | z_pk = zRef[p+k:]
1496 | WV = WRef
1497 | Y = np.dot(z_pk.T, z_pk)*(k+1)
1498 | corList = [np.dot(z_pk.T, z_pk)]
1499 | for i in range(1, k+p+1):
1500 | corList.append(np.dot(z_pk.T, zRef[p+k-i:-i]))
1501 |
1502 | zetaList = []
1503 | lastLine = []#corList
1504 | #zetaBlockList = []
1505 | for i in range(p):
1506 | zetaLine = []
1507 | for j in range(p):
1508 | if j < i:
1509 | zetaLine.append(zetaList[j][i].T)
1510 | else:
1511 | if i == 0:
1512 | zetaLine.append(corList[j]-np.outer(zRef[-1], zRef[-1-j])+np.outer(zRef[p+k-1], zRef[p+k-1-j]))
1513 | else:
1514 | #zetaLine.append(lastLine[j-1]-np.outer(zRef[-2], zRef[-2])+np.outer(zRef[p-2], zRef[p-2]))
1515 | zetaLine.append(lastLine[j-1]-np.outer(zRef[-1-i], zRef[-1-j])+np.outer(zRef[p+k-1-i], zRef[p+k-1-j]))
1516 |
1517 | zetaList.append(zetaLine)
1518 | lastLine = zetaLine
1519 |
1520 | for i in range(p):
1521 | zetaList[i] = np.hstack(zetaList[i])
1522 | ZZi = np.vstack(zetaList)
1523 | pnl = len(WRef)*(p-1)
1524 | for i in range(0, k+1):
1525 | if i > 0:
1526 | WV = np.dot(WV, VRef)
1527 | zeta1_i = calcZetaDataRefImp(zRef, p, i)[k-i:]
1528 | zZiRef = np.dot(z_pk.T, zeta1_i)
1529 | zZi = np.hstack(corList[1+i:p+i+1])
1530 | ZZiRef = np.dot(zeta1_i.T, zeta1_i)
1531 | #ZZi = ZZ0
1532 | # print "i = "+str(i)+":"
1533 | # print ZZi
1534 | # print ZZiRef
1535 | # print "+++"+str(k)
1536 | K = np.dot(zZi, WV.T)
1537 | Y += -K-K.T + np.dot(np.dot(WV, ZZi), WV.T)
1538 | if i < k:
1539 | for j in range(p):
1540 | lastLine[j] += -np.outer(zRef[-p-i-1], zRef[-j-2-i])+np.outer(zRef[k-1-i], zRef[p-i+k-j-2])
1541 | line = np.hstack(lastLine)
1542 | ZZi = np.vstack([np.hstack([ ZZi[pnl:].T[pnl:], line.T[:pnl]]), line])
1543 | # print "===="
1544 | # print np.hstack(lastLine)
1545 | print Y_kRef
1546 | print Y
1547 | print calcErrorCoeffConstLen(zRef, WRef, k)
1548 | print calcErrorCoeffConstLenRoll(zRef, WRef, k)
1549 |
1550 | print "-------------affine version---------------"
1551 | print ""
1552 | W_c = calcRegressionCoeffAffineRefImp(zRef, p)
1553 | #print zeta
1554 | zeta0_c = calcZeta0cDataRefImp(zRef, p)
1555 | # print zeta0_c
1556 | # print zeta_c
1557 | # ZZ0_c = np.dot(zeta0_c.T, zeta_c)
1558 | # zZ_c = np.dot(z_p.T, zeta_c)
1559 | # ZZ_c = np.dot(zeta_c.T, zeta_c)
1560 | # ZZI_c = invertByProjectionRefImp(ZZ_c)
1561 | # W0_c = np.dot(zZ_c, ZZI_c)
1562 | #The following idea of V_c is cumbersome, because it
1563 | #puts error on the constant-one component of zeta_c,
1564 | #if this reduces the overall error.
1565 | #But the constant component may not be compromized in any case, so
1566 | #this way to compute V_c is misleading:
1567 | # V_c_cumb = np.dot(ZZ0_c, ZZI_c)
1568 | # print V_c_cumb
1569 |
1570 | # print W0_c
1571 | print W_c
1572 | print ""
1573 | #Better way:
1574 | V_c = buildV_cRefImp(W_c)
1575 | print V_c
1576 | print "calc iterated error coeff:"
1577 | X_k_c = calcErrorCoeffAffineRefImp(zRef, W_c, 4)
1578 | print calcErrorCoeffAffine(zRef, W_c, 4)
1579 | print X_k_c
1580 |
1581 | print "const len:"
1582 | Y_k_c = calcErrorCoeffConstLenAffineRefImp(zRef, W_c, 3)
1583 | print calcErrorCoeffConstLenAffine(zRef, W_c, 3)
1584 | print Y_k_c
1585 | print calcErrorCoeffConstLenAffineRoll(zRef, W_c, 3)
1586 |
1587 | def insertRollTest():
1588 | print "insert test"
1589 | A = np.multiply(np.reshape(range(25), (5, 5)), 1.0)
1590 | print A
1591 | # B = np.array([[1.5, 2.5], [3.5, 4.5]])
1592 | # print B
1593 | print A[-3:-1, 3:5]
1594 | # A[2:4, 3:5] = B
1595 | # print A
1596 | # print np.roll(np.roll(A, 2, 0), 2, 1)
1597 |
1598 | if __name__ == "__main__":
1599 | PFAReferenceImplementationTest()
1600 | #compareIteration()
1601 | #insertRollTest()
1602 |
--------------------------------------------------------------------------------