-
Notifications
You must be signed in to change notification settings - Fork 0
/
traintest.py
278 lines (216 loc) · 11.3 KB
/
traintest.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
import cPickle
import itertools
import math
import pymongo as pm
import scipy as sp
from bson import SON
import tabular as tb
import gridfs
from starflow.utils import uniqify, ListUnion
from dbutils import get_most_recent_files, dict_union
import numpy as np
import tabular as tb
import copy
"""
train / test
"""
def combine_things(a,b):
for k in b:
if k == '$where' and k in a:
a[k] = a[k].strip('; ') + ' && ' + b[k]
elif k == '$or' and k in a:
pass
elif hasattr(b[k],'keys') and ('$in' in b[k] or '$nin' in b[k] or '$ne' in b[k]):
pass
else:
a[k] = b[k]
def generate_splits(dbname,collectionname,task_query,N,ntrain,ntest,
ntrain_pos = None, ntest_pos = None, universe = None,
use_negate = False, overlap=None):
task_query = copy.deepcopy(task_query)
print('Generating splits ...')
if universe is None:
universe = SON([])
connection = pm.Connection(document_class=SON)
db = connection[dbname]
data = db[collectionname + '.files']
fs = gridfs.GridFS(db,collection=collectionname)
combine_things(task_query,universe)
print('T',task_query)
task_data = list(data.find(task_query,fields=["filename"]))
task_fnames = [str(x['filename']) for x in task_data]
N_task = len(task_data)
if use_negate:
task_fnames = np.array(task_fnames)
all_data = list(data.find(universe,fields=["filename"]))
all_fnames = np.array([str(x['filename']) for x in all_data])
I = np.invert(tb.isin(all_fnames,task_fnames)).nonzero()[0]
nontask_data = [all_data[ind] for ind in I]
nontask_fnames = [str(x['filename']) for x in nontask_data]
assert set(task_fnames).intersection(nontask_fnames) == set([]), set(task_fnames).intersection(nontask_fnames)
else:
nontask_query = {'filename':{'$nin':task_fnames}}
nontask_query.update(universe)
nontask_data = get_most_recent_files(data,nontask_query)
N_nontask = len(nontask_data)
assert ntrain + ntest <= N_task + N_nontask, "Not enough training and/or testing examples " + str([N_task,N_nontask])
splits = []
for ind in range(N):
print('... split', ind)
if ntrain_pos is not None:
ntrain_neg = ntrain - ntrain_pos
assert ntrain_pos <= N_task, "Not enough positive training examples, there are: " + str(N_task)
assert ntrain_neg <= N_nontask, "Not enough negative training examples, there are: " + str(N_nontask)
perm_pos = sp.random.permutation(len(task_data))
perm_neg = sp.random.permutation(len(nontask_data))
train_data = [task_data[i] for i in perm_pos[:ntrain_pos]] + [nontask_data[i] for i in perm_neg[:ntrain_neg]]
if ntest_pos is not None:
ntest_neg = ntest - ntest_pos
assert ntest_pos <= N_task - ntrain_pos, "Not enough positive test examples, there are: " + str(N_task - ntrain_pos)
assert ntest_neg <= N_nontask - ntrain_neg, "Not enough negative test examples, there are: " + str(N_nontask - ntrain_neg)
test_data = [task_data[i] for i in perm_pos[ntrain_pos:ntrain_pos + ntest_pos]] + [nontask_data[i] for i in perm_neg[ntrain_neg:ntrain_neg + ntest_neg]]
else:
nontrain_data = [task_data[i] for i in perm_pos[ntrain_pos:]] + [nontask_data[i] for i in perm_neg[ntrain_neg:]]
new_perm = sp.random.permutation(len(nontrain_data))
test_data = [nontrain_data[i] for i in new_perm[:ntest]]
else:
if ntest_pos is not None:
ntest_neg = ntest - ntest_pos
assert ntest_pos <= N_task, "Not enough positive test examples, there are: " + str(N_task)
assert ntest_neg <= N_nontask, "Not enough negative test examples, there are: " + str(N_nontask)
perm_pos = sp.random.permutation(len(task_data))
perm_neg = sp.random.permutation(len(nontask_data))
test_data = [task_data[i] for i in perm_pos[:ntest_pos]] + [nontask_data[i] for i in perm_neg[:ntest_neg]]
nontest_data = [task_data[i] for i in perm_pos[ntest_pos:]] + [nontask_data[i] for i in perm_neg[ntest_neg:]]
new_perm = sp.random.permutation(len(nontest_data))
train_data = [nontest_data[i] for i in new_perm[:ntrain]]
else:
all_data = task_data + nontask_data
perm = sp.random.permutation(len(all_data))
train_data = [all_data[i] for i in perm[:ntrain]]
test_data = [all_data[i] for i in perm[ntrain:ntrain + ntest]]
train_filenames = np.array([str(_t['filename']) for _t in train_data])
test_filenames = np.array([str(_t['filename']) for _t in test_data])
train_labels = tb.isin(train_filenames,task_fnames)
test_labels = tb.isin(test_filenames,task_fnames)
#train_labels = sp.array([x['filename'] in task_fnames for x in train_data])
#test_labels = sp.array([x['filename'] in task_fnames for x in test_data])
assert set(train_filenames).intersection(test_filenames) == set([]), str(set(train_filenames).intersection(test_filenames))
split = {'train_data': train_data, 'test_data' : test_data, 'train_labels':train_labels,'test_labels':test_labels}
splits.append(split)
return splits
def validate(idseq):
ids = ListUnion(idseq)
ids1 = [id[1] for id in ids]
assert len(uniqify(ids1)) == sum([len(X) for X in idseq]), 'Classes are not disjoint.'
return ids
def generate_multi_splits(dbname, collectionname, task_queries, N, ntrain,
ntest, universe=None, labels=None, overlap=None,
balance = None, kfold=None):
nq = len(task_queries)
if labels is None:
labels = range(nq)
task_queries = [copy.deepcopy(task_query) for task_query in task_queries]
print('Generating splits ...')
if universe is None:
universe = SON([])
connection = pm.Connection(document_class=SON)
db = connection[dbname]
data = db[collectionname + '.files']
fs = gridfs.GridFS(db,collection=collectionname)
for task_query in task_queries:
combine_things(task_query,universe)
task_data = [list(data.find(task_query,fields=["filename"])) for task_query in task_queries]
if kfold is not None:
return generate_multi_kfold_splits(labels,kfold,task_data)
else:
return generate_multi_random_splits(labels,ntrain,ntest,task_data,overlap,N,balance)
def generate_multi_kfold_splits(labels,kfold,task_data):
total = sum([len(td) for td in task_data])
perms = [sp.random.permutation(len(td)) for td in task_data]
indsets = []
for td in task_data:
indset = []
l = len(td)
ld = l/kfold
for i in range(kfold-1):
indset.append(range(ld*i,ld*(i+1)))
indset.append(range(ld*(kfold-1),l))
indsets.append(indset)
splits = []
for ind in range(kfold):
l_ind_out = range(kfold)
l_ind_out.pop(ind)
train_data = []
train_labels = []
test_data = []
test_labels = []
for indset, td, p, label in zip(indsets,task_data,perms,labels):
for _ind in l_ind_out:
new = [td[p[i]] for i in indset[_ind]]
train_data.extend(new)
train_labels.extend([label]*len(new))
new = [td[p[i]] for i in indset[ind]]
test_data.extend(new)
test_labels.extend([label]*len(new))
train_filenames = np.array([str(_t['filename']) for _t in train_data])
test_filenames = np.array([str(_t['filename']) for _t in test_data])
assert set(train_filenames).intersection(test_filenames) == set([]), str(set(train_filenames).intersection(test_filenames))
assert len(train_data) + len(test_data) == total
split = {'train_data': train_data, 'test_data' : test_data, 'train_labels':train_labels,'test_labels':test_labels}
splits.append(split)
test_datas = [split['test_data'] for split in splits]
test_filenames = [[str(_t['filename']) for _t in tds] for tds in test_datas]
all_test_filenames = list(itertools.chain(*test_filenames))
assert len(all_test_filenames) == total
assert len(np.unique(all_test_filenames)) == total
return splits
def generate_multi_random_splits(labels,ntrain,ntest,task_data,overlap,N,balance):
floor = math.floor
if balance is None:
ntrain_vec = [ntrain/nq]*(nq - 1) + [ntrain - (ntrain/nq)*(nq-1)]
ntest_vec = [ntest/nq]*(nq - 1) + [ntest - (ntest/nq)*(nq-1)]
else:
assert len(balance) == nq - 1
ntrain_vec = [int(floor(ntrain*b)) for b in balance]
ntrain_vec = ntrain_vec + [ntrain - sum(ntrain_vec)]
ntest_vec = [int(floor(ntest*b)) for b in balance]
ntest_vec = ntest_vec + [ntest - sum(ntest_vec)]
for (tq,td,ntr,nte) in zip(task_queries,task_data,ntrain_vec,ntest_vec):
assert ntr + nte <= len(td), 'not enough examples to train/test for %s, %d needed, but only have %d' % (repr(tq),ntr+nte,len(td))
if overlap:
assert isinstance(overlap,float) and 0 < overlap <= 1, 'overlap must be a float in (0,1]'
for (_i,(td,ntr,nte)) in enumerate(zip(task_data,ntrain_vec,ntest_vec)):
eff_n = int(math.ceil((ntr+nte)*(1/overlap)))
#add given category intersections and check for error
perm = sp.random.permutation(len(td))
task_data[_i] = [td[_j] for _j in perm[:eff_n]]
splits = []
for ind in range(N):
print('... split', ind)
train_data = []
test_data = []
train_labels = []
test_labels = []
for (label,td,ntr,nte) in zip(labels,task_data,ntrain_vec,ntest_vec):
td = get_correct_td(td,test_data)
assert len(td) >= ntr + nte, 'problem with %s, need %d have %d' % (label,ntr+nte,len(td))
perm = sp.random.permutation(len(td))
train_data.extend([td[i] for i in perm[:ntr]])
td = get_correct_td(td,train_data)
assert len(td) >= nte, 'problem with %s test, need %d have %d' % (label,nte,len(td))
perm = sp.random.permutation(len(td))
test_data.extend([td[i] for i in perm[:nte]])
train_labels.extend([label]*ntr)
test_labels.extend([label]*nte)
train_filenames = np.array([str(_t['filename']) for _t in train_data])
test_filenames = np.array([str(_t['filename']) for _t in test_data])
assert set(train_filenames).intersection(test_filenames) == set([]), str(set(train_filenames).intersection(test_filenames))
split = {'train_data': train_data, 'test_data' : test_data, 'train_labels':train_labels,'test_labels':test_labels}
splits.append(split)
return splits
def get_correct_td(td,T):
tf = np.array([str(t['filename']) for t in td])
Tf = np.array([str(t['filename']) for t in T])
good_inds = np.invert(tb.fast.isin(tf,Tf)).nonzero()[0]
return [td[ind] for ind in good_inds]