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
|
#!/usr/bin/env python2.6
"""
mmert v0.2
multi-task mert
Copyright 2011
Patrick Simianer
Heidelberg University, ICL
"""
# This file is part of MMERT.
#
# MMERT is free software: you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
#
# MMERT is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU General Public License
# along with MMERT. If not, see <http://www.gnu.org/licenses/>.
import os, sys, shutil
def read_vec_from_file(fname):
""" File looks like: "w1 w2 ... wn" (on one line). """
return [float(el) for el in open(fname, 'r').read().strip().split()]
def vec2str(vector):
""" vec looks like: [0., 1.1, 2.2, ..., d] """
return " ".join([str(el) for el in vector])
def get_biggest_change(vector1, vector2):
""" Biggest difference of two vectors (element-wise) in abs value. """
biggest = 0
idx_ = 0
for element in vector1:
diff = abs(element - vector2[idx_])
if diff > biggest:
biggest = diff
idx_ += 1
return biggest
def get_avg_vec(vecs, vec_len):
""" 'vecs' is a list of vectors; calc arithmetic mean element-wise. """
avg_vec = []
num_vecs = len(vecs)
for idx_ in range(vec_len):
mysum = 0.
for vector in vecs:
mysum += vector[idx_]
avg = mysum/num_vecs
avg_vec.append(avg)
return avg_vec
def get_best_points_line_idx(lines):
""" Line number where to find best point in a mert.log. """
idx_ = 0
for line in lines:
if line.startswith("Best point: "):
break
idx_ += 1
return idx_
def get_weights_from_log(fname):
""" Get weights from a mert.log. """
logfile = open(fname, "r")
lines = logfile.readlines()
logfile.close()
idx_ = get_best_points_line_idx(lines)
return [float(el) for el in \
lines[idx_].split("Best point: ")[1].split("=>")[0].split()]
def write_new_log_file(fname, vec_):
""" Write a new mert.log file with weights found in vec. """
logfile = open(fname, "r")
lines = logfile.readlines()
logfile.close()
idx_ = get_best_points_line_idx(lines)
log = "".join(lines[0:idx_])
log += "Best point: "+vec2str(vec_)+" => 99\n"
log += "".join(lines[idx_+1:])
open(fname, "w+").write(log)
# parameters
DIR_PREFIX = sys.argv[1] # directory prefix
# (for task specific directories)
WORKDIR = sys.argv[2] # work directory
TASKS = sys.argv[3].split(",") # task list "a,b,c..."
CUR_IT = int(sys.argv[4]) # current (global) iteration
NUM_WEIGHTS = int(sys.argv[5]) # dimensionality of a weight vector?
MIN_CHANGE = float(sys.argv[6]) # min change in a vector
# to still continue
LAMBDA = float(sys.argv[7]) # lambda for regularization
FIRST_AVERAGE = bool(int(sys.argv[8])) # first average vector,
# 0 or self provided
print " CUR_IT %s" % str(CUR_IT)
if CUR_IT == 1:
print " DIR_PREFIX %s" % DIR_PREFIX
print " WORKDIR %s" % WORKDIR
print " TASKS %s" % str(TASKS)
print " NUM_WEIGHTS %s" % str(NUM_WEIGHTS)
print " MIN_CHANGE %s" % str(MIN_CHANGE)
print " LAMBDA %s" % str(LAMBDA)
print " FIRST_AVERAGE %s" % str(FIRST_AVERAGE)
if not FIRST_AVERAGE:
print " first average is 0 vector, creating file.."
open(WORKDIR+"/run0.avector.txt", "w+").write(\
" ".join(["0.0" for i in range(NUM_WEIGHTS)]))
else:
print " looking for pre-defined first average.."
if not os.path.exists(WORKDIR+"/run0.avector.txt"):
sys.stderr.write(sys.argv[0] + \
" ERROR: You have to provide your own run0.avector.txt \
containing the first average vector! arg FIRST_AVERAGE=" \
+str(FIRST_AVERAGE)+"\n")
sys.exit (1)
ITERATIONS = {} # current iteration(s)
CUR_VECS = {} # current weight vectors (from latest mert iteration)
VEC_LEN = 0 # check vector length
for t in TASKS:
print " reading vector from task %s" % t
ITERATIONS[t] = int(open(DIR_PREFIX + "_" + t +\
"/finished_step.txt").read().strip())
cur = get_weights_from_log(DIR_PREFIX+"_"+t+"/run"+str((ITERATIONS[t])) + \
".mert.log")
print cur
VEC_LEN = len(cur)
if VEC_LEN != NUM_WEIGHTS:
sys.stderr.write(sys.argv[0] + \
" ERROR: vector lengths differ! (task: "+t+", len: "+str(VEC_LEN) + \
", expected: "+str(NUM_WEIGHTS)+")\n")
CUR_VECS[t] = cur
# get previous average vector
PREV_AVG_VEC = read_vec_from_file(WORKDIR+"/run"+str(CUR_IT-1)+".avector.txt")
print " Previous average vector (clip against):"
print PREV_AVG_VEC
# new weight vecs
print " Calculating new weight vectors (clipping).."
NEXT_VECS = {}
for (t, vec) in CUR_VECS.items():
print " Task %s, before (current)" % t
print vec
idx = 0
nxt = []
for w in vec:
if w == 0:
nxt.append(0.0)
idx += 1
continue
if (w - PREV_AVG_VEC[idx] > 0):
nxt.append(max(PREV_AVG_VEC[idx], w - LAMBDA))
elif (w - PREV_AVG_VEC[idx] < 0):
nxt.append(min(PREV_AVG_VEC[idx], w + LAMBDA))
else:
nxt.append(w)
idx += 1
NEXT_VECS[t] = nxt
print " Task %s, after (next)" % t
print nxt
# get current average vector
AVG_VEC = get_avg_vec(NEXT_VECS.values(), VEC_LEN)
print " Current average vector (after clipping):"
print AVG_VEC
# convergence?
print " Previous average:", PREV_AVG_VEC
print " Current average:", AVG_VEC
BIGGEST_CHANGE = get_biggest_change(PREV_AVG_VEC, AVG_VEC)
print " Biggest change:", BIGGEST_CHANGE
# note: use weights from latest mert.log for eval!
if BIGGEST_CHANGE < MIN_CHANGE:
print " Converged!"
open(WORKDIR+"/CONVERGED", "w+").close()
# overwrite (and backup) mert.log
for t in TASKS:
logfilename = DIR_PREFIX+"_"+t+"/run"+str(ITERATIONS[t])+".mert.log"
shutil.copy(logfilename, logfilename+".bak")
write_new_log_file(logfilename, NEXT_VECS[t])
open(DIR_PREFIX+"_"+t+"/weightshist", "a+").write(str(ITERATIONS[t])+"\n"+ \
"cur "+str(CUR_VECS[t])+"\n"+ \
"prev avg "+str(PREV_AVG_VEC)+"\n"+ \
"next "+str(NEXT_VECS[t])+"\n"+ \
"cur avg "+str(AVG_VEC)+"\n---\n")
# write out current avg vector
open(WORKDIR+"/run"+str(CUR_IT)+".avector.txt", "w+").write(vec2str(AVG_VEC))
|
