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
353
354
355
356
357
358
359
360
361
362
|
/*
Copyright (c) 2009 Yahoo! Inc. All rights reserved. The copyrights
embodied in the content of this file are licensed under the BSD
(revised) open source license
The algorithm here is generally based on Jonathan Shewchuck's tutorial.
*/
#include <fstream>
#include <float.h>
#include <netdb.h>
#include <string.h>
#include <stdio.h>
#include <assert.h>
#include "parse_example.h"
#include "constant.h"
#include "sparse_dense.h"
#include "cg.h"
#include "cache.h"
#include "multisource.h"
#include "simple_label.h"
#include "delay_ring.h"
void quad_grad_update(weight* weights, feature& page_feature, v_array<feature> &offer_features, size_t mask, float g)
{
size_t halfhash = quadratic_constant * page_feature.weight_index;
float update = g * page_feature.x;
for (feature* ele = offer_features.begin; ele != offer_features.end; ele++)
{
weight* w=&weights[(halfhash + ele->weight_index) & mask];
w[1] += update * ele->x;
}
}
void quad_precond_update(weight* weights, feature& page_feature, v_array<feature> &offer_features, size_t mask, float g)
{
size_t halfhash = quadratic_constant * page_feature.weight_index;
float update = g * page_feature.x;
for (feature* ele = offer_features.begin; ele != offer_features.end; ele++)
{
weight* w=&weights[(halfhash + ele->weight_index) & mask];
w[4] += update * ele->x * ele->x;
}
}
// w[0] = weight
// w[1] = accumulated first derivative
// w[2] = step direction
// w[3] = old first derivative
// w[4] = preconditioner
float predict_and_gradient(regressor& reg, example* &ec)
{
float raw_prediction = inline_predict(reg,ec,0);
float fp = finalize_prediction(raw_prediction);
label_data* ld = (label_data*)ec->ld;
float loss_grad = reg.loss->first_derivative(fp,ld->label)*ld->weight;
size_t thread_mask = global.thread_mask;
weight* weights = reg.weight_vectors[0];
for (size_t* i = ec->indices.begin; i != ec->indices.end; i++)
{
feature *f = ec->subsets[*i][0];
for (; f != ec->subsets[*i][1]; f++)
{
weight* w = &weights[f->weight_index & thread_mask];
w[1] += loss_grad * f->x;
}
}
for (vector<string>::iterator i = global.pairs.begin(); i != global.pairs.end();i++)
{
if (ec->subsets[(int)(*i)[0]].index() > 0)
{
v_array<feature> temp = ec->atomics[(int)(*i)[0]];
temp.begin = ec->subsets[(int)(*i)[0]][0];
temp.end = ec->subsets[(int)(*i)[0]][1];
for (; temp.begin != temp.end; temp.begin++)
quad_grad_update(weights, *temp.begin, ec->atomics[(int)(*i)[1]], thread_mask, loss_grad);
}
}
return fp;
}
void update_preconditioner(regressor& reg, example* &ec)
{
label_data* ld = (label_data*)ec->ld;
float curvature = reg.loss->second_derivative(ec->final_prediction,ld->label) * ld->weight;
size_t thread_mask = global.thread_mask;
weight* weights = reg.weight_vectors[0];
for (size_t* i = ec->indices.begin; i != ec->indices.end; i++)
{
feature *f = ec->subsets[*i][0];
for (; f != ec->subsets[*i][1]; f++)
{
weight* w = &weights[f->weight_index & thread_mask];
w[4] += f->x * f->x * curvature;
}
}
for (vector<string>::iterator i = global.pairs.begin(); i != global.pairs.end();i++)
{
if (ec->subsets[(int)(*i)[0]].index() > 0)
{
v_array<feature> temp = ec->atomics[(int)(*i)[0]];
temp.begin = ec->subsets[(int)(*i)[0]][0];
temp.end = ec->subsets[(int)(*i)[0]][1];
for (; temp.begin != temp.end; temp.begin++)
quad_precond_update(weights, *temp.begin, ec->atomics[(int)(*i)[1]], thread_mask, curvature);
}
}
}
float dot_with_direction(regressor& reg, example* &ec)
{
float ret = 0;
weight* weights = reg.weight_vectors[0];
size_t thread_mask = global.thread_mask;
weights +=2;//direction vector stored two advanced
for (size_t* i = ec->indices.begin; i != ec->indices.end; i++)
{
feature *f = ec->subsets[*i][0];
for (; f != ec->subsets[*i][1]; f++)
ret += weights[f->weight_index & thread_mask] * f->x;
}
for (vector<string>::iterator i = global.pairs.begin(); i != global.pairs.end();i++)
{
if (ec->subsets[(int)(*i)[0]].index() > 0)
{
v_array<feature> temp = ec->atomics[(int)(*i)[0]];
temp.begin = ec->subsets[(int)(*i)[0]][0];
temp.end = ec->subsets[(int)(*i)[0]][1];
for (; temp.begin != temp.end; temp.begin++)
ret += one_pf_quad_predict(weights, *temp.begin, ec->atomics[(int)(*i)[1]], thread_mask);
}
}
return ret;
}
double derivative_magnitude(regressor& reg)
{//compute derivative magnitude & shift new derivative to old
double ret = 0.;
uint32_t length = 1 << global.num_bits;
size_t stride = global.stride;
weight* weights = reg.weight_vectors[0];//shift by one for ease of indexing.
for(uint32_t i = 0; i < length; i++)
{
ret += weights[stride*i+1]*weights[stride*i+1]*weights[stride*i+4];
weights[stride*i+3] = weights[stride*i+1];
weights[stride*i+1] = 0;
}
return ret;
}
void zero_derivative(regressor& reg)
{//compute derivative magnitude & shift new derivative to old
uint32_t length = 1 << global.num_bits;
size_t stride = global.stride;
weight* weights = reg.weight_vectors[0];//shift by one for ease of indexing.
for(uint32_t i = 0; i < length; i++)
weights[stride*i+1] = 0;
}
double direction_magnitude(regressor& reg)
{//compute direction magnitude
double ret = 0.;
uint32_t length = 1 << global.num_bits;
size_t stride = global.stride;
weight* weights = reg.weight_vectors[0];
for(uint32_t i = 0; i < length; i++)
ret += weights[stride*i+2]*weights[stride*i+2];
return ret;
}
double derivative_diff_mag(regressor& reg)
{//compute the derivative difference
double ret = 0.;
uint32_t length = 1 << global.num_bits;
size_t stride = global.stride;
weight* weights = reg.weight_vectors[0];
for(uint32_t i = 0; i < length; i++)
{
ret += weights[stride*i+1]*weights[stride*i+4]*
(weights[stride*i+1] - weights[stride*i+3]);
}
return ret;
}
double add_regularization(regressor& reg,float regularization)
{//compute the derivative difference
double ret = 0.;
uint32_t length = 1 << global.num_bits;
size_t stride = global.stride;
weight* weights = reg.weight_vectors[0];
for(uint32_t i = 0; i < length; i++)
weights[stride*i+1] += regularization*weights[stride*i];
return ret;
}
void finalize_preconditioner(regressor& reg,float regularization)
{
uint32_t length = 1 << global.num_bits;
size_t stride = global.stride;
weight* weights = reg.weight_vectors[0];
for(uint32_t i = 0; i < length; i++) {
weights[stride*i+4] += regularization;
if (weights[stride*i+4] > 0)
weights[stride*i+4] = 1. / weights[stride*i+4];
}
}
double derivative_in_direction(regressor& reg)
{
double ret = 0.;
uint32_t length = 1 << global.num_bits;
size_t stride = global.stride;
for(uint32_t i = 0; i < length; i++)
ret += reg.weight_vectors[0][stride*i+3]*reg.weight_vectors[0][stride*i+2];
return ret;
}
void update_direction(regressor& reg, float old_portion)
{
uint32_t length = 1 << global.num_bits;
size_t stride = global.stride;
weight* weights = reg.weight_vectors[0];
for(uint32_t i = 0; i < length; i++)
{
weights[stride*i+2] = weights[stride*i+3]*weights[stride*i+4] + old_portion * weights[stride*i+2];
}
}
void update_weight(regressor& reg, float step_size)
{
uint32_t length = 1 << global.num_bits;
size_t stride = global.stride;
for(uint32_t i = 0; i < length; i++)
{
reg.weight_vectors[0][stride*i] += step_size * reg.weight_vectors[0][stride*i+2];
}
}
void setup_cg(gd_thread_params t)
{
regressor reg = t.reg;
size_t thread_num = 0;
example* ec = NULL;
v_array<float> predictions;
size_t example_number=0;
double curvature=0.;
bool gradient_pass=true;
double loss_sum = 0;
float step_size = 0.;
double importance_weight_sum = 0.;
double previous_d_mag=0;
size_t current_pass = 0;
double previous_loss_sum = 0;
while ( true )
{
if ((ec = get_example(thread_num)) != NULL)//semiblocking operation.
{
assert(ec->in_use);
if (ec->pass != current_pass)//we need to do work on all features.
{
if (current_pass == 0)
finalize_preconditioner(reg,global.regularization*importance_weight_sum);
if (gradient_pass) // We just finished computing all gradients
if (current_pass > 0 && loss_sum > previous_loss_sum)
{// we stepped to far last time, step back
step_size *= 0.5;
cout << "backstepping, new step_size = " << step_size << endl;
update_weight(reg,- step_size);
zero_derivative(reg);
loss_sum = 0.;
}
else
{
previous_loss_sum = loss_sum;
loss_sum = 0.;
if (global.regularization > 0.)
add_regularization(reg,global.regularization*importance_weight_sum);
example_number = 0;
curvature = 0;
float mix_frac = 0;
if (current_pass != 0)
mix_frac = derivative_diff_mag(reg) / previous_d_mag;
if (mix_frac < 0 || isnan(mix_frac))
mix_frac = 0;
float new_d_mag = derivative_magnitude(reg);
previous_d_mag = new_d_mag;
update_direction(reg, mix_frac);
gradient_pass = false;//now start computing curvature
}
else // just finished all second gradients
{
if (global.regularization > 0.)
curvature += global.regularization*direction_magnitude(reg)*importance_weight_sum;
step_size = - derivative_in_direction(reg)/(max(curvature,1.));
predictions.erase();
update_weight(reg,step_size);
gradient_pass = true;
}//now start computing derivatives.
current_pass++;
}
if (gradient_pass)
{
ec->final_prediction = predict_and_gradient(reg,ec);
if (current_pass == 0)
{
label_data* ld = (label_data*)ec->ld;
importance_weight_sum += ld->weight;
update_preconditioner(reg,ec);
}
label_data* ld = (label_data*)ec->ld;
ec->loss = reg.loss->getLoss(ec->final_prediction, ld->label) * ld->weight;
loss_sum += ec->loss;
push(predictions,ec->final_prediction);
}
else //computing curvature
{
float d_dot_x = dot_with_direction(reg,ec);
label_data* ld = (label_data*)ec->ld;
ec->final_prediction = predictions[example_number];
ec->loss = reg.loss->getLoss(ec->final_prediction, ld->label) * ld->weight;
float sd = reg.loss->second_derivative(predictions[example_number++],ld->label);
curvature += d_dot_x*d_dot_x*sd*ld->weight;
}
finish_example(ec);
}
else if (thread_done(thread_num))
{
if (example_number == predictions.index())//do one last update
{
if (global.regularization > 0.)
curvature += global.regularization*direction_magnitude(reg)*importance_weight_sum;
float step_size = - derivative_in_direction(reg)/(max(curvature,1.));
update_weight(reg,step_size);
}
if (global.local_prediction > 0)
shutdown(global.local_prediction, SHUT_WR);
free(predictions.begin);
return;
}
else
;//busywait when we have predicted on all examples but not yet trained on all.
}
free(predictions.begin);
return;
}
void destroy_cg()
{
}
|