-
Notifications
You must be signed in to change notification settings - Fork 1
/
probsub.py
200 lines (174 loc) · 7.02 KB
/
probsub.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
# probsub.py
# General probably submodular learning interface
from __future__ import absolute_import, division, print_function # We require Python 2.6 or later
from tqdm import tqdm
# packages for SLIC superpixels
from skimage.segmentation import slic
from skimage.segmentation import mark_boundaries
import graphfuncs
import numpy as np
from one_slack_ssvm_hard import OneSlackSSVMHard
params = {
'slic': {
'segments': 500, # number of superpixels
'sigma': 3,
},
'ssvm':{
'C': 1.,
},
}
import os
try:
import cPickle as pickle
except:
import pickle
def segments(im, params_slic):
return slic(im, n_segments = params_slic['segments'], sigma = params_slic['sigma']) + 1 # so min. label is 1
def build_graph(params, dataset, example, memoize=False, load=True):
"""
add feature graph to an example of the dataset
"""
params_slic = params['slic']
def do_it():
im = dataset['helper'].get_image(dataset, example)
segments = segments(im, params_slic)
if 'mask' in example:
m = dataset['helper'].get_mask(dataset, example)
graph = graphfuncs.rag_histograms(im, segments, gt=m)
else:
graph = graphfuncs.rag_histograms(im, segments)
x_pystruct_features = np.vstack([graph.node[n]['feat'] for n in graph])
x_pystruct_edges = np.array(graph.edges()) - 1
x_pystruct_edgefeatures = np.vstack([graph.edge[n1][n2]['weight'] for (n1, n2) in graph.edges_iter()])
example['x'] = (x_pystruct_features, x_pystruct_edges, x_pystruct_edgefeatures)
if 'mask' in example:
example['y'] = np.array([graph.node[n]['gt'] for n in graph], dtype=int)
return example
if memoize is False:
return do_it()
else:
cachef = os.path.join(memoize, example["name"] + ".pkl")
lock = cachef + ".lock"
if os.path.isfile(lock):
return "locked"
else:
try:
if load:
example_cached = pickle.load(open(cachef, "rb"))
example.clear()
example.update( t for t in example_cached.viewitems() ) # change in-place!
return example
else:
open(cachef, "rb").close()
return "exists"
except (KeyboardInterrupt, SystemExit):
raise
except: # actually do the computation
try:
open(lock, 'a').close()
do_it()
pickle.dump(example, open(cachef, "wb"), pickle.HIGHEST_PROTOCOL)
return example
finally:
os.remove(lock)
def build_graphs(params, dataset, examples=-1, memoize=False, load=True):
"""
add feature graphs to examples of the dataset
if examples = -1, apply to all dataset['loaded_examples']
"""
if examples == -1:
examples = dataset['loaded_examples']
for example in tqdm(examples, desc='Building graphs'):
build_graph(params, dataset, example, memoize, load)
return examples
def binarize(dset, examples, cat, copy=True):
c = dset["helper"].labels.index(cat)
if copy:
examples = deepcopy(examples)
for ex in tqdm(examples, desc='Binarizing masks'):
ex['y_src'] = ex['y'].copy()
ex['class'] = cat
ex['y'][ex['y'] != c] = 0
ex['y'][ex['y'] == c] = 1
return examples
def get_constraints(constraints=()):
if constraints == ():
return {'type': 'empty'}
else:
cset, cex = constraints
cdict = {'type': cset}
n_states = cdict['n_states'] = len(set(l for ex in cex for l in ex['y']))
ufeatdim = cdict['ufeatdim'] = cex[0]['x'][0].shape[1] # graph.node[next(iter(graph))]['feat'].shape[0]
efeatdim = cdict['efeatdim'] = cex[0]['x'][2].shape[1]
jfeatdim = cdict['jfeatdim'] = n_states * ufeatdim + n_states * n_states * efeatdim
pairj = cdict['pairj'] = n_states * ufeatdim
# negative indices
cdict['negind'] = np.ogrid[pairj + efeatdim : pairj + 3*efeatdim]
# positive indices
cdict['posind'] = np.hstack((np.ogrid[pairj : pairj + efeatdim],
np.ogrid[pairj + 3*efeatdim:jfeatdim]))
cdict['examples'] = cex
return cdict
import one_slack_ssvm_hard
import edge_features_loss_graph
SSVM = one_slack_ssvm_hard.OneSlackSSVMHard
CRF = edge_features_loss_graph.EdgeFeaturesLossGraph
import probsub_helpers
from copy import deepcopy
def get_learner(params, cdict, warmstartfrom=None):
paramsssvm = params['ssvm']
crf = CRF(inference_method=('ogm', {'alg': 'gc'})) # graph-cuts in OpenGM
if cdict['type'] == 'empty':
negative = positive = []
initialize = generate = None
elif cdict['type'] == 'C0':
negative = positive = np.hstack((cdict['negind'], cdict['posind']))
initialize = generate = None
elif cdict['type'] == 'C1':
negative = np.hstack((cdict['negind'], cdict['posind']))
positive = cdict['posind']
initialize = generate = None
elif cdict['type'] == 'C2':
negative = cdict['negind']
positive = cdict['posind']
initialize = generate = None
elif cdict['type'] == 'C3':
negative = positive = []
initialize, generate = probsub_helpers.get_helpers(params, cdict)
elif cdict['type'] == 'C4':
negative = positive = []
initialize, generate = probsub_helpers.get_helpers(params, cdict)
if warmstartfrom is not None:
learner = deepcopy(warmstartfrom)
learner.hard_satisfied = False
learner.converged = False
learner.negativity_constraint = negative
learner.positivity_constraint = positive
learner.initialize_constraints = initialize
learner.generate_hard_constraints = generate
else:
learner = SSVM(crf, inference_cache=0, C=paramsssvm['C'], tol=paramsssvm['tol'], max_iter=paramsssvm['max_iter'],
n_jobs=1, negativity_constraint=negative, positivity_constraint=positive,
initialize_constraints=initialize, generate_hard_constraints=generate,
show_loss_every=params['log']['loss_every'], check_constraints=True,
break_on_bad=paramsssvm['break_on_bad'], verbose=params['log']['SSVM_verbose'])
return learner
def fit(examples, constraints, params, startfrom = []):
if startfrom:
fitted = copy.deepcopy(startfrom)
else:
fitted = []
if isinstance(constraints, tuple): # 1 constraint set
constraints = [constraints]
X = [ex['x'] for ex in examples]
Y = [ex['y'] for ex in examples]
for constraint in constraints:
cdict = get_constraints(constraints=constraint)
if not fitted:
learner = get_learner(params, cdict)
learner.fit(X, Y)
else:
learner = get_learner(params, cdict, warmstartfrom=fitted[-1])
learner.fit(X, Y, warm_start=True)
fitted.append(learner)
return fitted