forked from apache/systemds
-
Notifications
You must be signed in to change notification settings - Fork 0
/
MultiLogReg.dml
353 lines (314 loc) · 13.5 KB
/
MultiLogReg.dml
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
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
#-------------------------------------------------------------
#
# Licensed to the Apache Software Foundation (ASF) under one
# or more contributor license agreements. See the NOTICE file
# distributed with this work for additional information
# regarding copyright ownership. The ASF licenses this file
# to you under the Apache License, Version 2.0 (the
# "License"); you may not use this file except in compliance
# with the License. You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing,
# software distributed under the License is distributed on an
# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
# KIND, either express or implied. See the License for the
# specific language governing permissions and limitations
# under the License.
#
#-------------------------------------------------------------
# Solves Multinomial Logistic Regression using Trust Region methods.
# (See: Trust Region Newton Method for Logistic Regression, Lin, Weng and Keerthi, JMLR 9 (2008) 627-650)
# INPUT PARAMETERS:
# --------------------------------------------------------------------------------------------
# NAME TYPE DEFAULT MEANING
# --------------------------------------------------------------------------------------------
# X String --- Location to read the matrix of feature vectors
# Y String --- Location to read the matrix with category labels
# B String --- Location to store estimated regression parameters (the betas)
# Log String " " Location to write per-iteration variables for log/debugging purposes
# icpt Int 0 Intercept presence, shifting and rescaling X columns:
# 0 = no intercept, no shifting, no rescaling;
# 1 = add intercept, but neither shift nor rescale X;
# 2 = add intercept, shift & rescale X columns to mean = 0, variance = 1
# reg Double 0.0 regularization parameter (lambda = 1/C); intercept is not regularized
# tol Double 0.000001 tolerance ("epsilon")
# moi Int 100 max. number of outer (Newton) iterations
# mii Int 0 max. number of inner (conjugate gradient) iterations, 0 = no max
# fmt String "text" Matrix output format, usually "text" or "csv" (for matrices only)
# --------------------------------------------------------------------------------------------
# The largest label represents the baseline category; if label -1 or 0 is present, then it is
# the baseline label (and it is converted to the largest label).
#
# The Log file, when requested, contains the following per-iteration variables in CSV format,
# each line containing triple (NAME, ITERATION, VALUE) with ITERATION = 0 for initial values:
#
# NAME MEANING
# -------------------------------------------------------------------------------------------
# LINEAR_TERM_MIN The minimum value of X %*% B, used to check for overflows
# LINEAR_TERM_MAX The maximum value of X %*% B, used to check for overflows
# NUM_CG_ITERS Number of inner (Conj.Gradient) iterations in this outer iteration
# IS_TRUST_REACHED 1 = trust region boundary was reached, 0 = otherwise
# POINT_STEP_NORM L2-norm of iteration step from old point (i.e. matrix B) to new point
# OBJECTIVE The loss function we minimize (negative regularized log-likelihood)
# OBJ_DROP_REAL Reduction in the objective during this iteration, actual value
# OBJ_DROP_PRED Reduction in the objective predicted by a quadratic approximation
# OBJ_DROP_RATIO Actual-to-predicted reduction ratio, used to update the trust region
# IS_POINT_UPDATED 1 = new point accepted; 0 = new point rejected, old point restored
# GRADIENT_NORM L2-norm of the loss function gradient (omitted if point is rejected)
# TRUST_DELTA Updated trust region size, the "delta"
# -------------------------------------------------------------------------------------------
#
# Script invocation example:
# hadoop jar SystemML.jar -f MultiLogReg.dml -nvargs icpt=2 reg=1.0 tol=0.000001 moi=100 mii=20
# X=INPUT_DIR/X123 Y=INPUT_DIR/Y123 B=OUTPUT_DIR/B123 fmt=csv Log=OUTPUT_DIR/log
fileX = $X;
fileY = $Y;
fileB = $B;
fileLog = ifdef ($Log, " ");
fmtB = ifdef ($fmt, "text");
intercept_status = ifdef ($icpt, 0); # $icpt = 0;
regularization = ifdef ($reg, 0.0); # $reg = 0.0;
tol = ifdef ($tol, 0.000001); # $tol = 0.000001;
maxiter = ifdef ($moi, 100); # $moi = 100;
maxinneriter = ifdef ($mii, 0); # $mii = 0;
tol = as.double (tol);
print ("BEGIN MULTINOMIAL LOGISTIC REGRESSION SCRIPT");
print ("Reading X...");
X = read (fileX);
print ("Reading Y...");
Y_vec = read (fileY);
eta0 = 0.0001;
eta1 = 0.25;
eta2 = 0.75;
sigma1 = 0.25;
sigma2 = 0.5;
sigma3 = 4.0;
psi = 0.1;
N = nrow (X);
D = ncol (X);
# Introduce the intercept, shift and rescale the columns of X if needed
if (intercept_status == 1 | intercept_status == 2) # add the intercept column
{
X = cbind (X, matrix (1, rows = N, cols = 1));
D = ncol (X);
}
scale_lambda = matrix (1, rows = D, cols = 1);
if (intercept_status == 1 | intercept_status == 2)
{
scale_lambda [D, 1] = 0;
}
if (intercept_status == 2) # scale-&-shift X columns to mean 0, variance 1
{ # Important assumption: X [, D] = matrix (1, rows = N, cols = 1)
avg_X_cols = t(colSums(X)) / N;
var_X_cols = (t(colSums (X ^ 2)) - N * (avg_X_cols ^ 2)) / (N - 1);
is_unsafe = var_X_cols <= 0;
scale_X = 1.0 / sqrt (var_X_cols * (1 - is_unsafe) + is_unsafe);
scale_X [D, 1] = 1;
shift_X = - avg_X_cols * scale_X;
shift_X [D, 1] = 0;
rowSums_X_sq = (X ^ 2) %*% (scale_X ^ 2) + X %*% (2 * scale_X * shift_X) + sum (shift_X ^ 2);
} else {
scale_X = matrix (1, rows = D, cols = 1);
shift_X = matrix (0, rows = D, cols = 1);
rowSums_X_sq = rowSums (X ^ 2);
}
# Henceforth we replace "X" with "X %*% (SHIFT/SCALE TRANSFORM)" and rowSums(X ^ 2)
# with "rowSums_X_sq" in order to preserve the sparsity of X under shift and scale.
# The transform is then associatively applied to the other side of the expression,
# and is rewritten via "scale_X" and "shift_X" as follows:
#
# ssX_A = (SHIFT/SCALE TRANSFORM) %*% A --- is rewritten as:
# ssX_A = diag (scale_X) %*% A;
# ssX_A [D, ] = ssX_A [D, ] + t(shift_X) %*% A;
#
# tssX_A = t(SHIFT/SCALE TRANSFORM) %*% A --- is rewritten as:
# tssX_A = diag (scale_X) %*% A + shift_X %*% A [D, ];
# Convert "Y_vec" into indicator matrix:
max_y = max (Y_vec);
if (min (Y_vec) <= 0) {
# Category labels "0", "-1" etc. are converted into the largest label
Y_vec = Y_vec + (- Y_vec + max_y + 1) * (Y_vec <= 0);
max_y = max_y + 1;
}
Y = table (seq (1, N, 1), Y_vec, N, max_y);
K = ncol (Y) - 1; # The number of non-baseline categories
lambda = (scale_lambda %*% matrix (1, rows = 1, cols = K)) * regularization;
delta = 0.5 * sqrt (D) / max (sqrt (rowSums_X_sq));
B = matrix (0, rows = D, cols = K); ### LT = X %*% (SHIFT/SCALE TRANSFORM) %*% B;
### LT = cbind (LT, matrix (0, rows = N, cols = 1));
### LT = LT - rowMaxs (LT) %*% matrix (1, rows = 1, cols = K+1);
P = matrix (1, rows = N, cols = K+1); ### exp_LT = exp (LT);
P = P / (K + 1); ### P = exp_LT / (rowSums (exp_LT) %*% matrix (1, rows = 1, cols = K+1));
obj = N * log (K + 1); ### obj = - sum (Y * LT) + sum (log (rowSums (exp_LT))) + 0.5 * sum (lambda * (B_new ^ 2));
Grad = t(X) %*% (P [, 1:K] - Y [, 1:K]);
if (intercept_status == 2) {
Grad = diag (scale_X) %*% Grad + shift_X %*% Grad [D, ];
}
Grad = Grad + lambda * B;
norm_Grad = sqrt (sum (Grad ^ 2));
norm_Grad_initial = norm_Grad;
if (maxinneriter == 0) {
maxinneriter = D * K;
}
iter = 1;
# boolean for convergence check
converge = (norm_Grad < tol) | (iter > maxiter);
print ("-- Initially: Objective = " + obj + ", Gradient Norm = " + norm_Grad + ", Trust Delta = " + delta);
if (fileLog != " ") {
log_str = "OBJECTIVE,0," + obj;
log_str = append (log_str, "GRADIENT_NORM,0," + norm_Grad);
log_str = append (log_str, "TRUST_DELTA,0," + delta);
} else {
log_str = " ";
}
while (! converge)
{
# SOLVE TRUST REGION SUB-PROBLEM
S = matrix (0, rows = D, cols = K);
R = - Grad;
V = R;
delta2 = delta ^ 2;
inneriter = 1;
norm_R2 = sum (R ^ 2);
innerconverge = (sqrt (norm_R2) <= psi * norm_Grad);
is_trust_boundary_reached = 0;
while (! innerconverge)
{
if (intercept_status == 2) {
ssX_V = diag (scale_X) %*% V;
ssX_V [D, ] = ssX_V [D, ] + t(shift_X) %*% V;
} else {
ssX_V = V;
}
Q = P [, 1:K] * (X %*% ssX_V);
HV = t(X) %*% (Q - P [, 1:K] * (rowSums (Q) %*% matrix (1, rows = 1, cols = K)));
if (intercept_status == 2) {
HV = diag (scale_X) %*% HV + shift_X %*% HV [D, ];
}
HV = HV + lambda * V;
alpha = norm_R2 / sum (V * HV);
Snew = S + alpha * V;
norm_Snew2 = sum (Snew ^ 2);
if (norm_Snew2 <= delta2)
{
S = Snew;
R = R - alpha * HV;
old_norm_R2 = norm_R2
norm_R2 = sum (R ^ 2);
V = R + (norm_R2 / old_norm_R2) * V;
innerconverge = (sqrt (norm_R2) <= psi * norm_Grad);
} else {
is_trust_boundary_reached = 1;
sv = sum (S * V);
v2 = sum (V ^ 2);
s2 = sum (S ^ 2);
rad = sqrt (sv ^ 2 + v2 * (delta2 - s2));
if (sv >= 0) {
alpha = (delta2 - s2) / (sv + rad);
} else {
alpha = (rad - sv) / v2;
}
S = S + alpha * V;
R = R - alpha * HV;
innerconverge = TRUE;
}
inneriter = inneriter + 1;
innerconverge = innerconverge | (inneriter > maxinneriter);
}
# END TRUST REGION SUB-PROBLEM
# compute rho, update B, obtain delta
gs = sum (S * Grad);
qk = - 0.5 * (gs - sum (S * R));
B_new = B + S;
if (intercept_status == 2) {
ssX_B_new = diag (scale_X) %*% B_new;
ssX_B_new [D, ] = ssX_B_new [D, ] + t(shift_X) %*% B_new;
} else {
ssX_B_new = B_new;
}
LT = cbind ((X %*% ssX_B_new), matrix (0, rows = N, cols = 1));
if (fileLog != " ") {
log_str = append (log_str, "LINEAR_TERM_MIN," + iter + "," + min (LT));
log_str = append (log_str, "LINEAR_TERM_MAX," + iter + "," + max (LT));
}
LT = LT - rowMaxs (LT) %*% matrix (1, rows = 1, cols = K+1);
exp_LT = exp (LT);
P_new = exp_LT / (rowSums (exp_LT) %*% matrix (1, rows = 1, cols = K+1));
obj_new = - sum (Y * LT) + sum (log (rowSums (exp_LT))) + 0.5 * sum (lambda * (B_new ^ 2));
# Consider updating LT in the inner loop
# Consider the big "obj" and "obj_new" rounding-off their small difference below:
actred = (obj - obj_new);
rho = actred / qk;
is_rho_accepted = (rho > eta0);
snorm = sqrt (sum (S ^ 2));
if (fileLog != " ") {
log_str = append (log_str, "NUM_CG_ITERS," + iter + "," + (inneriter - 1));
log_str = append (log_str, "IS_TRUST_REACHED," + iter + "," + is_trust_boundary_reached);
log_str = append (log_str, "POINT_STEP_NORM," + iter + "," + snorm);
log_str = append (log_str, "OBJECTIVE," + iter + "," + obj_new);
log_str = append (log_str, "OBJ_DROP_REAL," + iter + "," + actred);
log_str = append (log_str, "OBJ_DROP_PRED," + iter + "," + qk);
log_str = append (log_str, "OBJ_DROP_RATIO," + iter + "," + rho);
}
if (iter == 1) {
delta = min (delta, snorm);
}
alpha2 = obj_new - obj - gs;
alpha = ifelse(alpha2 <= 0, sigma3, max(sigma1, -0.5 * gs / alpha2));
if (rho < eta0)
delta = min (max (alpha, sigma1) * snorm, sigma2 * delta);
else if (rho < eta1)
delta = max (sigma1 * delta, min (alpha * snorm, sigma2 * delta));
else if (rho < eta2)
delta = max (sigma1 * delta, min (alpha * snorm, sigma3 * delta));
else
delta = max (delta, min (alpha * snorm, sigma3 * delta));
if (is_trust_boundary_reached == 1)
{
print ("-- Outer Iteration " + iter + ": Had " + (inneriter - 1) + " CG iterations, trust bound REACHED");
} else {
print ("-- Outer Iteration " + iter + ": Had " + (inneriter - 1) + " CG iterations");
}
print (" -- Obj.Reduction: Actual = " + actred + ", Predicted = " + qk +
" (A/P: " + (round (10000.0 * rho) / 10000.0) + "), Trust Delta = " + delta);
if (is_rho_accepted)
{
B = B_new;
P = P_new;
Grad = t(X) %*% (P [, 1:K] - Y [, 1:K]);
if (intercept_status == 2) {
Grad = diag (scale_X) %*% Grad + shift_X %*% Grad [D, ];
}
Grad = Grad + lambda * B;
norm_Grad = sqrt (sum (Grad ^ 2));
obj = obj_new;
print (" -- New Objective = " + obj + ", Beta Change Norm = " + snorm + ", Gradient Norm = " + norm_Grad);
if (fileLog != " ") {
log_str = append (log_str, "IS_POINT_UPDATED," + iter + ",1");
log_str = append (log_str, "GRADIENT_NORM," + iter + "," + norm_Grad);
}
} else {
if (fileLog != " ") {
log_str = append (log_str, "IS_POINT_UPDATED," + iter + ",0");
}
}
if (fileLog != " ") {
log_str = append (log_str, "TRUST_DELTA," + iter + "," + delta);
}
iter = iter + 1;
converge = ((norm_Grad < (tol * norm_Grad_initial)) | (iter > maxiter) |
((is_trust_boundary_reached == 0) & (abs (actred) < (abs (obj) + abs (obj_new)) * 0.00000000000001)));
if (converge) { print ("Termination / Convergence condition satisfied."); } else { print (" "); }
}
if (intercept_status == 2) {
B_out = diag (scale_X) %*% B;
B_out [D, ] = B_out [D, ] + t(shift_X) %*% B;
} else {
B_out = B;
}
write (B_out, fileB, format=fmtB);
if (fileLog != " ") {
write (log_str, fileLog);
}