-
Notifications
You must be signed in to change notification settings - Fork 2
/
gulf.py
520 lines (432 loc) · 18.7 KB
/
gulf.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
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
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
import torch
import numpy as np
from torch.optim import SGD
import torchnet as tnt
import torch.nn.functional as F
from utils.utils import cast
from utils.utils0 import logging, timeLog, Clock, raise_if_nonpositive, raise_if_nan, Global_state, Local_state, stem_name, add_if_absent_, raise_if_absent
#--- constants
Ddim = 0; Cdim = 1
Target_index = 1
#----------------------------------------------------------
# input: opt.ini_type in { iniRand, iniBase, iniBase/2, file, file/2 }.
# output: opt.do_iniBase, opt.fc_scale
#----------------------------------------------------------
def interpret_ini_type_(opt):
add_if_absent_(opt, ['ini_type','initial'], '')
opt.do_iniBase = opt.ini_type.startswith('iniBase')
opt.fc_scale = 0.5 if opt.ini_type.endswith('/2') else -1
if opt.ini_type.startswith('file') and not opt.initial:
raise ValueError("ini_type=%s requires 'initial' to specify a pathname of the initial model.")
if opt.initial and not opt.ini_type.startswith('file'):
raise ValueError("'initial' requires 'ini_type' to be 'file' or 'file/2'.")
def is_gulf(opt):
return is_gulf1(opt) or is_gulf2(opt)
def is_gulf1(opt):
return opt.m > 0
def is_gulf2(opt):
return (not is_gulf1(opt)) and opt.alpha != 1
#----------------------------------------------------------
def base_update(opt, clock, net, params, trn_data, test_dss, g_st, l_st):
timeLog('base_update --------------------------------')
doing_epo = opt.do_count_epochs
doing_upd = not doing_epo
max_count = opt.max_count
ti_upd,ti_epo = get_test_interval(opt)
epoch,upd,lr_coeff = l_st.get()
optimizer = create_optim(opt, lr_coeff, params)
mtr_loss = tnt.meter.AverageValueMeter()
#-
def test_net():
is_last = (doing_upd and upd >= max_count) or (doing_epo and epoch >= max_count)
mys = eval(clock, doing_upd, g_st.epo(epoch), g_st.upd(upd), upd, net, test_dss, opt,
trn_data=trn_data, params=params, is_last=is_last)
logging(mys, opt.csv_fn)
#-
num_data = 0
while (doing_upd and upd < max_count) or (doing_epo and epoch < max_count):
optimizer,lr_coeff = change_lr_if_needed(opt, optimizer, lr_coeff, params, epoch=epoch)
timeLog('epoch ' + str(epoch) + ' upd ' + str(upd))
for sample in trn_data:
num_data += sample[0].size(0)
if doing_upd and upd >= max_count:
break
optimizer,lr_coeff = change_lr_if_needed(opt, optimizer, lr_coeff, params, upd=upd)
loss,_ = net(sample, is_train=True)
loss.backward()
mtr_loss.add(float(loss))
if opt.max_grad_norm > 0:
torch.nn.utils.clip_grad_norm_(params.values(), opt.max_grad_norm)
optimizer.step(); upd += 1; optimizer.zero_grad()
#--- show progress
if opt.inc > 0 and upd % opt.inc == 0:
s = ' ... %d, %.5f, #data,%d' % (upd, mtr_loss.value()[0], num_data)
timeLog(s)
raise_if_nan(mtr_loss.value()[0])
mtr_loss.reset()
#--- test and save
is_last = doing_upd and upd >= max_count
if is_last or ti_upd > 0 and upd % ti_upd == 0:
test_net()
save_net(opt, is_last, params, g_st, Local_state(epoch, upd, lr_coeff))
epoch += 1
#--- test and save
is_last = doing_epo and epoch >= max_count
if is_last or ti_epo > 0 and epoch % ti_epo == 0:
test_net()
save_net(opt, is_last, params, g_st, Local_state(epoch, upd, lr_coeff))
g_st.update(epoch, upd)
#----------------------------------------------------------
def scale_fc(params, name, fc_scale, what=None):
if what is not None:
timeLog('Scaling %s of %s by factor %.5f ...' % (name,what,fc_scale))
else:
timeLog('Scaling %s by factor %.5f ...' % (name,fc_scale))
with torch.no_grad():
for type in ['weight', 'bias']:
params[name+'.'+type].data *= fc_scale
#----------------------------------------------------------
# base model if num_stages == 1
# base-loop if num_stages > 1
#----------------------------------------------------------
def train_base_model(opt, net, params, trn_data, test_dss):
check_opt_(opt); show_model_info(opt, params)
assert opt.alpha == 1 and opt.m <= 0
if opt.fc_scale > 0:
raise ValueError('train_base_model: No support fc_scale.')
clock = Clock()
g_st = Global_state()
l_st = Local_state()
if opt.resume != '':
g_st,l_st = load(opt.resume, params, do_show_st=True)
elif opt.initial != '':
load(opt.initial, params)
optim = None
while g_st.lc() < opt.num_stages:
base_update(opt, clock, net, params, trn_data, test_dss, g_st, l_st)
l_st.reset()
#----------------------------------------------------------
def train_gulf_model(opt, i_net, i_params, o_net, o_params, trn_data, test_dss):
check_opt_(opt); show_model_info(opt, o_params)
raise_if_nonpositive(opt.alpha, 'alpha')
if is_gulf2(opt):
if opt.alpha > 1:
raise ValueError('alpha must be no greater than 1 for GULF2.')
clock = Clock()
g_st = Global_state()
l_st = Local_state()
if opt.resume != '': # resume training ...
if opt.do_iniBase or opt.initial:
logging('!WARNING!: do_iniBase or opt.initial is ignored as resume is given ...')
logging('Resuming training ... ')
g_st,l_st = load(opt.resume, o_params, i_params, do_show_st=True)
else:
if opt.initial: # initialize parameter by loading from a file.
load(opt.initial, o_params, i_params)
if opt.do_iniBase: # initialize parameter by regular training.
base_g_st = Global_state(); base_l_st = Local_state()
base_update(opt, clock, o_net, o_params, trn_data, test_dss, base_g_st, base_l_st)
copy_params(src=o_params, dst=i_params)
if opt.fc_scale > 0: # scale the last linear layer.
scale_fc(i_params, opt.fc_name, opt.fc_scale, 'i_params')
scale_fc(o_params, opt.fc_name, opt.fc_scale, 'o_params')
while g_st.lc() < opt.num_stages:
gulf_update(opt, clock, i_net, i_params, o_net, o_params,
trn_data, test_dss, g_st, l_st)
copy_params(src=o_params, dst=i_params)
l_st.reset()
#----------------------------------------------------------
def gulf_update(opt, clock, i_net, i_params, o_net, o_params,
trn_data, test_dss, g_st, l_st):
gulf12 = 'GULF1' if is_gulf1(opt) else 'GULF2'
timeLog('gulf_update (%s) --------------------------------' % gulf12)
loss_function = o_net(None)
doing_epo = opt.do_count_epochs
doing_upd = not doing_epo
max_count = opt.max_count
ti_upd,ti_epo = get_test_interval(opt)
epoch, upd, lr_coeff = l_st.get()
optimizer = create_optim(opt, lr_coeff, o_params)
mtr_o_loss = tnt.meter.AverageValueMeter()
mtr_g_loss = tnt.meter.AverageValueMeter()
mtr_tar_loss = tnt.meter.AverageValueMeter()
#-
def test_o_net():
is_last = (doing_upd and upd >= max_count) or (doing_epo and epoch >= max_count)
mys = eval(clock, doing_upd, g_st.epo(epoch), g_st.upd(upd), upd, o_net, test_dss, opt,
trn_data=trn_data, params=o_params, is_last=is_last)
logging(mys, opt.csv_fn)
#-
num_data = 0
while (doing_upd and upd < max_count) or (doing_epo and epoch < max_count):
optimizer,lr_coeff = change_lr_if_needed(opt, optimizer, lr_coeff, o_params, epoch=epoch)
logging('epoch ' + str(epoch) + ' upd ' + str(upd))
for sample in trn_data:
bsz = sample[0].size(0)
num_data += bsz
if doing_upd and upd >= max_count:
break
optimizer,lr_coeff = change_lr_if_needed(opt, optimizer, lr_coeff, o_params, upd=upd)
targets = cast(sample[Target_index], 'long')
with torch.no_grad():
i_output = i_net(sample)
o_loss, o_output = o_net(sample, is_train=True)
#*** GULF2 (h(p)=L_y(p))
# Let f = f_theta and f' = f_{theta_t}
# D_{L_y}(f,f') + alpha nabla L_y(f')^T f
# = L_y(f) - (1-alpha) nabla L_y(f')^T f + c where c is a constant that does not depend on theta.
if opt.m <= 0:
i_output.detach_(); assert i_output.grad is None
i_output.requires_grad = True
i_loss = loss_function(i_output, targets)
i_loss.backward()
g_loss = o_loss - (1 - opt.alpha)*(o_output*i_output.grad.data).sum()
#*** GULF1 (h(u)=|u|^2/2)
else:
#--- m steps of functional gradient descent
tar = i_output
assert tar.size(Ddim) == bsz
for i in range(opt.m):
tar.detach_();
if tar.grad is not None:
tar.grad.zero_()
tar.requires_grad = True
tar_loss = loss_function(tar, targets)
tar_loss.backward()
tar.detach_(); tar.data -= opt.alpha * (bsz*tar.grad.data)
mtr_tar_loss.add(float(tar_loss))
#--- || f(theta;x) - f_m^*(x,y) ||^2/2
g_loss = ((o_output-tar)**2).sum()/2/bsz
#---
g_loss.backward()
mtr_o_loss.add(float(o_loss))
mtr_g_loss.add(float(g_loss))
if opt.max_grad_norm > 0:
torch.nn.utils.clip_grad_norm_(o_params.values(), opt.max_grad_norm)
optimizer.step(); upd += 1; optimizer.zero_grad()
#---- show progress
if opt.inc > 0 and upd % opt.inc == 0:
s = ' ... %d, g_loss,%.5f, o_loss,%.5f' % (upd, mtr_g_loss.value()[0], mtr_o_loss.value()[0])
if mtr_tar_loss.n > 0:
s += ', t_loss,%.5f' % mtr_tar_loss.value()[0]
timeLog(s)
raise_if_nan(mtr_o_loss.value()[0])
mtr_o_loss.reset(); mtr_g_loss.reset(); mtr_tar_loss.reset()
is_last = doing_upd and upd >= max_count
if is_last or ti_upd > 0 and upd % ti_upd == 0:
test_o_net()
save_net(opt, is_last, o_params, g_st, Local_state(epoch, upd, lr_coeff),
i_params)
epoch += 1
is_last = doing_epo and epoch >= max_count
if is_last or ti_epo > 0 and epoch % ti_epo == 0:
test_o_net()
save_net(opt, is_last, o_params, g_st, Local_state(epoch, upd, lr_coeff),
i_params)
g_st.update(epoch, upd)
#----------------------------------------------------------
def save_net(opt, is_end_of_loop, o_params, g_st, l_st, i_params=None):
fname = opt.save
if fname == None or fname == '':
return
ext = '.pth'
stem = stem_name(fname, ext)
save(stem+ext, o_params, g_st, l_st, i_params)
if opt.do_noow_save and is_end_of_loop:
slim_save(stem+'-glc'+str(g_st.lc()+1)+'-slim'+ext,
o_params, g_st, l_st)
#----------------------------------------------------------
def slim_save(fname, o_params, g_st, l_st):
if fname == None or fname == '':
return
timeLog('Saving (slim): ' + fname + ' ... ')
torch.save(dict(o_params=o_params, i_params=None,
global_state=g_st.to_list(), local_state=l_st.to_list(),
optimizer=None),
fname)
#----------------------------------------------------------
def save(fname, o_params, g_st, l_st, i_params=None):
if fname == None or fname == '':
return
timeLog('Saving (fat): ' + fname + ' ... ')
torch.save(dict(o_params=o_params, i_params=i_params,
global_state=g_st.to_list(), local_state=l_st.to_list(),
optimizer=None),
fname)
#----------------------------------------------------------
def load(fname, o_params, i_params=None, do_show_st=False):
timeLog('Loading ' + fname + ' ... ')
if torch.cuda.is_available():
d = torch.load(fname)
else:
d = torch.load(fname, map_location='cpu')
copy_params(src=d['o_params'], dst=o_params)
if i_params != None:
if d['i_params'] == None:
logging('***WARNING***: Copying o_params to i_params since i_params was not saved ...')
copy_params(src=d['o_params'], dst=i_params)
else:
copy_params(src=d['i_params'], dst=i_params)
g_st = Global_state(inplist=d['global_state'])
l_st = Local_state(inplist=d['local_state'])
if do_show_st:
logging('g_st= %s' % str(g_st))
logging('l_st= %s' % str(l_st))
return g_st,l_st
#----------------------------------------------------------
def test(net, data, opt, name):
timeLog('testing ... '+name)
topk = [1,5] if opt.do_top5 else [1]
inc = opt.test_inc
sum_loss = 0
mtr_err = tnt.meter.ClassErrorMeter(topk=topk, accuracy=False)
data_num = 0; count = 0
for sample in data:
bsz=sample[0].size(0)
data_num += bsz; count += 1
with torch.no_grad():
loss, output = net(sample, is_train=False)
mtr_err.add(output.data, sample[Target_index])
sum_loss += float(loss)*bsz
if inc > 0 and count % inc == 0:
s = '... testing ... %d (%d): %s' % (count, data_num, str(float(mtr_err.value()[0])))
timeLog(s)
return mtr_err.value(), sum_loss/data_num
#----------------------------------------------------------
def get_loss(net, data, inc, info_max):
sum_loss = 0; data_num = 0; count = 0
for sample in data:
if info_max > 0 and data_num >= info_max:
break
bsz=sample[0].size(0)
data_num += bsz; count += 1
with torch.no_grad():
loss, output = net(sample, is_train=False)
sum_loss += float(loss)*bsz
if inc > 0 and count % inc == 0:
timeLog('... getting loss ... %d (%d) ... ' % (count,data_num))
return sum_loss/data_num
#-----------------------------------------
def get_l2(params):
sum2 = 0
for v in params.values():
if v.requires_grad:
sum2 += float((v.data**2).sum())
return sum2
#----------------------------------------------------------
def copy_params(src, dst):
for key, value in dst.items():
value.data.copy_(src[key])
#----------------------------------------------------------
def clone_params(src):
return {
key: torch.zeros_like(value).data.copy_(value)
for key, value in src.items()
}
#----------------------------------------------------------
def print_params(params):
if len(params) <= 0:
return
logging('Parameters: ---------------------------------------------------------------------')
kmax = max(len(key) for key in params.keys())
for (key, v) in sorted(params.items()):
print(key.ljust(kmax+3), str(tuple(v.shape)).ljust(23), torch.typename(v), v.requires_grad)
logging('---------------------------------------------------------------------------------')
#----------------------------------------------------------
def setup_optim_params(opt, params, lam):
keys = [ k for k,v in params.items() if v.requires_grad ]
if lam > 0:
k_do_reg = keys
k_dont_reg = []
else:
k_do_reg = []
k_dont_reg = keys
optim_param =[]
if len(k_do_reg ) > 0:
optim_param += [ {'params': [ v for k,v in params.items() if k in k_do_reg ], 'weight_decay': lam} ]
if len(k_dont_reg) > 0:
optim_param += [ {'params': [ v for k,v in params.items() if k in k_dont_reg ], 'weight_decay': 0.0} ]
return optim_param,k_do_reg
#----------------------------------------------------------
def create_optim(opt, lr_coeff, params):
lr = opt.lr*lr_coeff
lam = opt.weight_decay
optim_param,_ = setup_optim_params(opt, params, lam)
timeLog('Creating optimizer with lr=%.5f and lam=%s' % (lr,str(lam)))
optim = SGD(optim_param, lr, momentum=0.9, weight_decay=lam, nesterov=False)
optim.zero_grad()
return optim
#----------------------------------------------------------
def change_lr_if_needed(opt, optimizer, lr_coeff, params, epoch=-1, upd=-1):
do_change = False
if upd >= 0 and not opt.do_count_epochs:
assert epoch == -1
if upd in opt.decay_lr_at:
do_change = True
elif epoch >= 0 and opt.do_count_epochs:
assert upd == -1
if epoch in opt.decay_lr_at:
do_change = True
if do_change:
lr_coeff *= opt.lr_decay_ratio
optimizer = create_optim(opt, lr_coeff, params) # this is from WRN code
return optimizer, lr_coeff
#----------------------------------------------------------
def show_err_loss(name, pfx, errs, loss):
s = ","+name+"_err"+pfx
for err in errs:
s += ',%.3f' % (err)
s += ","+name+"_loss"+pfx+"," + '%.5f' % (loss)
return s
#----------------------------------------------------------
def eval(clock, doing_upd, epo, upd, ite, net, test_dss, opt,
trn_data, params, is_last, do_mark_last=True):
clk_tim = clock.suspend()
pfx = '_t' if is_last and do_mark_last else ''
s = (',epoch,' + str(epo)) if not doing_upd else ''
mys = clk_tim + ",upd,"+str(upd) + s + ",ite,"+str(ite)
for dsinfo in test_dss:
if opt.do_reduce_testing and dsinfo['name'] == 'test' and not is_last:
continue
errs, loss = test(net, dsinfo['data'], opt, dsinfo['name'])
mys += show_err_loss(dsinfo['name'], pfx, errs, loss)
if is_last and opt.do_collect_info:
timeLog('Getting training loss ...')
trn_loss = get_loss(net, trn_data, opt.test_inc, opt.collect_info_max)
l2 = get_l2(params)
mys += ',train_loss%s,%.5f,l2%s,%.3f' % (pfx, trn_loss, pfx, l2)
clock.resume()
return mys
#----------------------------------------------------------
def get_test_interval(opt):
if opt.test_interval <= 0:
raise ValueError("test_interval must be positive.")
ti_upd = -1; ti_epo = -1
if opt.do_count_epochs:
ti_epo = opt.test_interval
else:
ti_upd = opt.test_interval
return ti_upd, ti_epo
#----------------------------------------------------------
#----------------------------------------------------------
def check_opt_(opt):
#--- required attributes
names = [ 'do_iniBase','fc_scale','alpha','m','num_stages','weight_decay','max_count','do_count_epochs','lr','test_interval']
raise_if_absent(opt, names, who='gulf')
add_if_absent_(opt, ['fc_name'], 'fc')
#--- optional attributes
add_if_absent_(opt, ['do_reduce_testing','do_top5','do_collect_info','do_noow_save','verbse'], False)
add_if_absent_(opt, ['inc','test_inc','collect_info_max'], -1)
add_if_absent_(opt, ['max_grad_norm','lr_decay_ratio'], -1)
add_if_absent_(opt, ['resume','save','initial','csv_fn'], '')
add_if_absent_(opt, ['decay_lr_at'],[])
if opt.verbose:
logging('gulf.check_opt_: opt -------------------------------------------')
logging({**vars(opt)})
logging('----------------------------------------------------------------')
#----------------------------------------------------------
def show_model_info(opt, params):
if opt.verbose:
print_params(params)
n_parameters = sum(p.numel() for p in params.values() if p.requires_grad)
logging('#parameters:' + str(n_parameters))