-
Notifications
You must be signed in to change notification settings - Fork 5
/
train_distributed.py
613 lines (553 loc) · 31.1 KB
/
train_distributed.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
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
import torch
import os
from tqdm import tqdm
import argparse
from utils import *
from datasets import MNIST, Kink, Cifar10, Slab, SlabLinear, SlabNonlinear4
from fastargs import Section, Param, get_current_config
from fastargs.validation import OneOf
from fastargs.decorators import param, section
from optimizer import NelderMead, PatternSearch
import time
import json
from sql import *
Section("dataset", "Dataset parameters").params(
name=Param(str, OneOf(("mnist", "kink", "cifar10", "slab", "slab_nonlinear_3", "slab_nonlinear_4", "slab_linear")), default="kink"),
)
Section("dataset.kink", "Dataset parameters for kink").enable_if(
lambda cfg: cfg['dataset.name'] == 'kink'
).params(
margin=Param(float),
noise=Param(float)
)
Section("dataset.mnistcifar", "Dataset parameters for mnist/cifar").params(
num_classes=Param(int)
)
Section("model", "Model architecture parameters").params(
arch=Param(str, OneOf(("mlp", "lenet")), default="mlp"),
model_count_times_batch_size=Param(int, default=20000*16),
init=Param(str, OneOf(("uniform", "regular", "uniform2", "uniform5", "sphere100", "sphere200")), default="uniform")
)
Section("model.lenet", "Model architecture parameters").params(
width=Param(float),
feature_dim=Param(float)
)
Section("model.mlp", "Model architecture parameters").enable_if(lambda cfg: cfg['model.arch'] == 'mlp').params(
hidden_units=Param(int),
layers=Param(int)
)
Section("optimizer").params(
name=Param(str, OneOf(["SGD", "SGDPoison", "Adam", "RMSProp", "guess", "GD"]), default='guess'),
es_u=Param(float, default=float('inf')),
es_l=Param(float, default=-float('inf')),
grad_norm_thres=Param(float, desc='only accept models with gradient norm smaller than specified'),
lr=Param(float, desc='learning rate'),
momentum=Param(float, desc='momentum', default=0),
epochs=Param(int, desc='number of epochs to optimize for'),
es_acc=Param(float, desc='stop the training when average training acc reaches this level'),
batch_size=Param(int, desc='number of epochs ot optimize for', default=3),
scheduler=Param(int, desc='whether to use a scheduler', default=False),
poison_factor=Param(float, desc='level of poisoning applied'),
print_intermediate_test_acc=Param(int, default=0, desc='whether to print intermediate test acc')
)
# TODO: write logic for excluded_cells
Section("distributed").params(
loss_thres=Param(str, default="0.3,0.4,0.5"),
num_samples=Param(str, default="2,4,8"),
excluded_cells=Param(str, default="", desc='ex: 32_(0.3, 0.35)/16_(0.3, 0.35)'),
target_model_count_subrun=Param(int, default=1),
training_seed=Param(int, default=None, desc='If there is no training seed, then the training seed increment with every new runs'),
data_seed=Param(int, default=None, desc='If there is no data seed, then the training seed increment with every new runs, otherwise, it is fix')
)
Section("output", "arguments associated with output").params(
target_model_count=Param(int, default=1),
folder=Param(str, default='test_distributed')
)
@section('dataset')
@param('name')
@param('mnistcifar.num_classes')
@param('kink.noise')
@param('kink.margin')
def get_dataset(name, num_samples, seed, num_classes=None, noise=None, margin=0.25):
if name =="mnist":
name = MNIST(batch_size=num_samples, threads=1, aug='none', train_count=num_samples, num_classes=num_classes, seed=seed)
train_data, train_labels = next(iter(name.train))
test_data, test_labels = next(iter(name.test))
train_data, train_labels, test_data, test_labels = train_data.cuda(), train_labels.cuda(), test_data.cuda(), test_labels.cuda()
test_all_data, test_all_labels = name.test_all_data, name.test_all_labels
elif name == "cifar10":
name = Cifar10(batch_size=num_samples, threads=1, aug='none', train_count=num_samples, num_classes=num_classes, seed=seed)
train_data, train_labels = next(iter(name.train))
test_data, test_labels = next(iter(name.test))
train_data, train_labels, test_data, test_labels = train_data.cuda(), train_labels.cuda(), test_data.cuda(), test_labels.cuda()
test_all_data, test_all_labels = name.test_all_data, name.test_all_labels
elif name == "kink":
train_data = torch.tensor(
Kink(train=True, samples=num_samples, seed=seed, noise=noise, margin=margin).data).float().cuda()
train_labels = torch.tensor(
Kink(train=True, samples=num_samples, seed=seed, noise=noise, margin=margin).labels).long().cuda()
test_data = torch.tensor(
Kink(train=False, samples=num_samples, seed=seed, noise=noise, margin=margin).data).float().cuda()
test_labels = torch.tensor(
Kink(train=False, samples=num_samples, seed=seed, noise=noise, margin=margin).labels).long().cuda()
test_all_data, test_all_labels = test_data, test_labels
elif name == "slab":
train_data = torch.tensor(
Slab(train=True, samples=num_samples, seed=seed, noise=noise, margin=margin).data).float().cuda()
train_labels = torch.tensor(
Slab(train=True, samples=num_samples, seed=seed, noise=noise, margin=margin).labels).long().cuda()
test_data = torch.tensor(
Slab(train=False, samples=num_samples, seed=seed, noise=noise, margin=margin).data).float().cuda()
test_labels = torch.tensor(
Slab(train=False, samples=num_samples, seed=seed, noise=noise, margin=margin).labels).long().cuda()
test_all_data, test_all_labels = test_data, test_labels
elif name == "slab_nonlinear_4":
dataset = SlabNonlinear4(samples=num_samples)
train_data = torch.tensor(dataset.data).float().cuda()
train_labels = torch.tensor(dataset.labels).long().cuda()
test_data = train_data
test_labels = train_labels
test_all_data, test_all_labels = train_data, train_labels
elif name == "slab_linear":
dataset = SlabLinear(samples=num_samples)
train_data = torch.tensor(dataset.data).float().cuda()
train_labels = torch.tensor(dataset.labels).long().cuda()
test_data = train_data
test_labels = train_labels
test_all_data, test_all_labels = train_data, train_labels
return train_data, train_labels, test_data, test_labels, test_all_data, test_all_labels
@section('model')
@param('arch')
def get_model(arch, model_count, device):
if arch == "mlp":
model = MLPModels(input_dim=2, output_dim=2,
layers=config['model.mlp.layers'], hidden_units=config['model.mlp.hidden_units'],
model_count=model_count, device=device)
elif arch == "lenet":
model = LeNetModels(output_dim=config['dataset.mnistcifar.num_classes'],
width_factor=config['model.lenet.width'],
model_count=model_count,
dataset=config['dataset.name'],
feature_dim=config['model.lenet.feature_dim']).to(device)
elif arch == "linear":
model = LinearModels(input_dim=(28*28 if config['dataset.name'] == "mnist" else 32*32*3),
output_dim=config['dataset.mnistcifar.num_classes'],
model_count=model_count, device=device)
return model
@section('optimizer')
@param('name')
@param('lr')
@param('momentum')
@param('scheduler')
def get_optimizer_and_scheduler(name, model, scheduler=False, lr=None, momentum=0):
if name in ["SGD", "GD", "RMSProp", "Adam", "SGDPoison"]:
if name == "RMSProp":
optimizer = torch.optim.RMSprop(model.parameters(), lr=lr)
elif name == "Adam":
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
elif name == "SGDPoisons":
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
else:
optimizer = torch.optim.SGD(model.parameters(), lr=lr, momentum=momentum)
if scheduler == False:
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=[9999999], gamma=0.2)
else:
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=[60, 120, 160], gamma=0.2)
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=[500, 1000, 2000, 3000], gamma=0.5)
elif name == "guess":
optimizer = None
scheduler = None
elif name == "NelderMead":
optimizer = NelderMead(model.parameters(), alpha=1, gamma=2, rho=0.5, sigma=0.5)
scheduler = None
elif name == "PatternSearch":
optimizer = PatternSearch(model.parameters())
scheduler = None
else:
optimizer = None
scheduler = None
return optimizer, scheduler
@section('optimizer')
@param('epochs')
@param('batch_size')
@param('es_u')
@param('es_acc')
@param('print_intermediate_test_acc')
def train_sgd(
train_data, train_labels, test_data, test_labels, model, loss_func, optimizer, scheduler, batch_size, epochs, es_u, es_acc=1, print_intermediate_test_acc=0):
for epoch in range(epochs):
idx_list = torch.randperm(len(train_data))
for st_idx in range(0, len(train_data), batch_size):
idx = idx_list[st_idx:min(st_idx + batch_size, len(train_data))]
train_loss, train_acc = calculate_loss_acc(train_data[idx], train_labels[idx], model, loss_func)
if es_u != float('inf'):
with torch.no_grad():
train_loss_all, train_acc_all = calculate_loss_acc(train_data, train_labels,
model.forward_normalize,
loss_func)
train_loss = torch.where((train_loss_all > es_u) | (train_acc_all < 1), train_loss,
torch.zeros_like(train_loss))
train_loss = train_loss[~train_loss.isnan()]
optimizer.zero_grad()
train_loss.sum().backward()
optimizer.step()
scheduler.step()
with torch.no_grad():
if epoch % (epochs // 100 + 1) == 0:
train_loss, train_acc = calculate_loss_acc(train_data, train_labels, model.forward_normalize, loss_func)
test_loss, test_acc = calculate_loss_acc(test_data, test_labels, model.forward_normalize, loss_func)
if len(train_loss[train_acc==1]) > 0:
print(f"train loss range: {train_loss[train_acc==1].max().item()} {train_loss[train_acc==1].min().item()}")
train_loss = train_loss[~train_loss.isnan()]
test_loss = test_loss[~test_loss.isnan()]
print(
f"epoch {epoch} - train_acc: {train_acc.mean().cpu().detach().item(): 0.2f}, train_loss: {train_loss.mean().cpu().detach().item(): 0.4f}")
print(
f"epoch {epoch} - test acc: {test_acc.mean().item(): 0.2f}, test loss: {test_loss.mean().item(): 0.2f}")
if print_intermediate_test_acc:
_, test_acc = calculate_loss_acc(test_all_data.cuda(), test_all_labels.cuda(), model, loss_func, batch_size=batch_size)
print("test acc (all):", test_acc)
if train_acc.mean() >= es_acc:
break
optimizer.zero_grad()
@section('optimizer')
@param('epochs')
@param('batch_size')
@param('es_u')
@param('es_acc')
@param('poison_factor')
def train_sgd_poison(
train_data, train_labels, test_data, test_labels, model, loss_func, optimizer, scheduler, batch_size, epochs, es_u, es_acc=1, poison_factor=None,
test_all_data=None, test_all_labels=None):
test_all_data, test_all_labels = test_all_data.cuda(), test_all_labels.cuda()
poison_test_labels = torch.tensor([1,0], device=test_all_labels.device)[test_all_labels]
repeats = 10
poison_data = torch.cat([train_data.repeat_interleave(repeats, dim=0), test_all_data], dim=0)
poison_labels = torch.cat([train_labels.repeat_interleave(repeats), poison_test_labels], dim=0)
for epoch in range(epochs):
idx_list = torch.randperm(len(poison_data))
for st_idx in range(0, len(poison_data), batch_size):
idx = idx_list[st_idx:min(st_idx + batch_size, len(poison_data))]
train_loss, train_acc = calculate_loss_acc(poison_data[idx], poison_labels[idx], model, loss_func)
optimizer.zero_grad()
train_loss.sum().backward()
optimizer.step()
scheduler.step()
with torch.no_grad():
if epoch % (epochs // 100 + 1) == 0:
train_loss, train_acc = calculate_loss_acc(train_data, train_labels, model.forward_normalize, loss_func)
test_loss, test_acc = calculate_loss_acc(test_all_data, test_all_labels, model.forward_normalize, loss_func)
poison_train_loss, poison_train_acc = calculate_loss_acc(poison_data, poison_labels, model.forward_normalize, loss_func)
if len(train_loss[train_acc==1]) > 0:
print(f"train loss range: {train_loss[train_acc==1].max().item()} {train_loss[train_acc==1].min().item()}")
train_loss = train_loss[~train_loss.isnan()]
test_loss = test_loss[~test_loss.isnan()]
print(
f"epoch {epoch} - train_acc: {train_acc.mean().cpu().detach().item(): 0.2f}, train_loss: {train_loss.mean().cpu().detach().item(): 0.4f}")
print(
f"epoch {epoch} - test acc all (max, min): {test_acc.max().item(): 0.2f}, {test_acc.min().item(): 0.2f}")
print(
f"epoch {epoch} - poison_train_acc: {poison_train_acc.mean().cpu().detach().item(): 0.2f}, train_loss: {poison_train_loss.mean().cpu().detach().item(): 0.4f}")
if poison_train_acc.mean() >= es_acc:
break
optimizer.zero_grad()
@section('optimizer')
@param('epochs')
@param('es_acc')
@param('es_u')
def train_gd(train_data, train_labels, test_data, test_labels, model, loss_func, optimizer, scheduler, epochs, es_u, es_acc=2):
for epoch in range(epochs):
train_loss, train_acc = calculate_loss_acc(train_data, train_labels, model, loss_func)
if es_u != float('inf'):
train_loss = torch.where((train_loss > es_u) | (train_acc < 1),
train_loss, torch.zeros_like(train_loss))
optimizer.zero_grad()
train_loss.sum().backward()
optimizer.step()
scheduler.step()
with torch.no_grad():
if epoch % (epochs // 100 + 1) == 0:
train_loss, train_acc = calculate_loss_acc(train_data, train_labels, model.forward_normalize, loss_func)
test_loss, test_acc = calculate_loss_acc(test_data, test_labels, model.forward_normalize, loss_func)
print(
f"epoch {epoch} - train_loss: {train_loss.mean().cpu().detach().item(): 0.4f}, train_acc: {train_acc.mean().cpu().detach().item(): 0.2f}")
print(
f"epoch {epoch} - test acc: {test_acc.mean().item(): 0.2f}, test loss: {test_loss.mean().item(): 0.2f}")
if train_acc.mean() >= es_acc:
break
optimizer.zero_grad()
@section('optimizer')
@param('epochs')
@param('es_acc')
@torch.no_grad()
def train_nm(train_data, train_labels, test_data, test_labels, model, loss_func, optimizer, epochs, es_acc=2):
for epoch in range(epochs):
optimizer.step(lambda: calculate_loss_acc(train_data, train_labels, model, loss_func)[0][0])
if epoch % (epochs // 100 + 1) == 0:
train_loss, train_acc = calculate_loss_acc(train_data, train_labels, model.forward_normalize, loss_func)
test_loss, test_acc = calculate_loss_acc(test_data, test_labels, model.forward_normalize, loss_func)
print(
f"epoch {epoch} - train_loss: {train_loss.mean().cpu().detach().item(): 0.4f}, train_acc: {train_acc.mean().cpu().detach().item(): 0.2f}")
print(
f"epoch {epoch} - test acc: {test_acc.mean().item(): 0.2f}, test loss: {test_loss.mean().item(): 0.2f}")
if train_acc >= es_acc:
break
@section('optimizer')
@param('epochs')
@param('es_acc')
@torch.no_grad()
def train_ps(train_data, train_labels, test_data, test_labels, model, loss_func, optimizer, epochs, es_acc=2):
for epoch in range(epochs):
optimizer.step(lambda: calculate_loss_acc(train_data, train_labels, model.forward_normalize, loss_func)[0][0])
if epoch % (epochs // 100) == 0:
train_loss, train_acc = calculate_loss_acc(train_data, train_labels, model.forward_normalize, loss_func)
test_loss, test_acc = calculate_loss_acc(test_data, test_labels, model.forward_normalize, loss_func)
print(
f"epoch {epoch} - train_loss: {train_loss.mean().cpu().detach().item(): 0.4f}, train_acc: {train_acc.mean().cpu().detach().item(): 0.2f}")
print(
f"epoch {epoch} - test acc: {test_acc.mean().item(): 0.2f}, test loss: {test_loss.mean().item(): 0.2f}")
if train_acc >= es_acc:
break
@section('optimizer')
@param('epochs')
@param('es_acc')
@torch.no_grad()
def train_ps_fast(train_data, train_labels, test_data, test_labels, model, loss_func, optimizer, epochs, es_acc=2):
for epoch in range(epochs):
model.pattern_search(train_data, train_labels, loss_func)
if epoch % (epochs // 100) == 0:
train_loss, train_acc = calculate_loss_acc(train_data, train_labels, model.forward_normalize, loss_func)
test_loss, test_acc = calculate_loss_acc(test_data, test_labels, model.forward_normalize, loss_func)
print(
f"epoch {epoch} - train_loss: {train_loss.mean().cpu().detach().item(): 0.4f}, train_acc: {train_acc.mean().cpu().detach().item(): 0.2f}")
print(
f"epoch {epoch} - test acc: {test_acc.mean().item(): 0.2f}, test loss: {test_loss.mean().item(): 0.2f}")
if train_acc.mean() >= es_acc:
break
return model.get_model_subsets([0]).to(train_data.device)
@section('optimizer')
@param('epochs')
@param('es_acc')
@torch.no_grad()
def train_greedy_random(train_data, train_labels, test_data, test_labels, model, loss_func, optimizer, epochs, es_acc=2):
for epoch in range(epochs):
model.greedy_random(train_data, train_labels, loss_func)
if epoch % (epochs // 300) == 0:
train_loss, train_acc = calculate_loss_acc(train_data, train_labels, model.forward_normalize, loss_func)
test_loss, test_acc = calculate_loss_acc(test_data, test_labels, model.forward_normalize, loss_func)
print(
f"epoch {epoch} - train_loss: {train_loss.mean().cpu().detach().item(): 0.4f}, train_acc: {train_acc.mean().cpu().detach().item(): 0.2f}")
print(
f"epoch {epoch} - test acc: {test_acc.mean().item(): 0.2f}, test loss: {test_loss.mean().item(): 0.2f}")
if train_acc.mean() >= es_acc:
break
return model.get_model_subsets([0]).to(train_data.device)
@section('optimizer')
@param('name')
def train(train_data, train_labels, test_data, test_labels, model, loss_func, optimizer, scheduler, name, batch_size=None, es_u=None, test_all_data=None, test_all_labels=None):
if name in ["SGD", "RMSProp", "Adam"]:
train_sgd(train_data, train_labels, test_data, test_labels, model, loss_func, optimizer, scheduler, batch_size=batch_size, es_u=es_u)
elif name in ["SGDPoison"]:
train_sgd_poison(train_data, train_labels, test_data, test_labels, model, loss_func, optimizer, scheduler, batch_size=batch_size, es_u=es_u, test_all_data=test_all_data, test_all_labels=test_all_labels)
elif name == "GD":
train_gd(train_data, train_labels, test_data, test_labels, model, loss_func, optimizer, scheduler, es_u=es_u)
elif name == "NelderMead":
train_nm(train_data, train_labels, test_data, test_labels, model, loss_func, optimizer)
elif name == "PatternSearch":
train_ps(train_data, train_labels, test_data, test_labels, model, loss_func, optimizer)
elif name == "PatternSearchFast":
model = train_ps_fast(train_data, train_labels, test_data, test_labels, model, loss_func, optimizer)
elif name == "GreedyRandom":
model = train_greedy_random(train_data, train_labels, test_data, test_labels, model, loss_func, optimizer)
else:
pass
return model
def convert_config_to_dict(config):
config_dict = {}
for path in config.entries.keys():
try:
value = config[path]
if value is not None:
config_dict['.'.join(path)] = config[path]
except:
pass
return config_dict
def build_model_output_path(config, training_seed, data_seed, cur_num_samples):
output_path = f"{config['output.folder']}/models/"
output_path += f"{config['dataset.name']}_s{cur_num_samples}_"
if config['model.arch'] == "lenet":
output_path += f"lenet_w{config['model.lenet.width']}_"
elif config['model.arch'] == 'linear':
output_path += f"linear_"
elif config['model.arch'] == 'mlp':
output_path += f"mlp_h{config['model.mlp.hidden_units']}"\
f"l{config['model.mlp.layers']}_"
output_path += f"opt{config['optimizer.name'] }_"
if config['optimizer.grad_norm_thres']:
output_path += '_gnorm'
output_path += f"dseed{data_seed}_tseed{training_seed}"
return output_path
if __name__ == "__main__":
parser = argparse.ArgumentParser()
# the config needs to change to a vector of sample size (potential model counts as well) x loss bins??
# it will then devote compute to the specific sample size & loss bins combination with the smallest number of trained models
# always select a random training seed & data seed
config = get_current_config()
config.augment_argparse(parser)
config.collect_argparse_args(parser)
config.summary()
config_ns = config.get()
loss_thres = [float(v) for v in config['distributed.loss_thres'].split(",")]
loss_bins = [(low, up) for low, up in zip(loss_thres[:-1], loss_thres[1:])]
num_samples = [int(v) for v in config['distributed.num_samples'].split(",")]
# create the table with counts
os.makedirs(config['output.folder'], exist_ok=True)
db_path = os.path.join(config['output.folder'], "model_stats.db")
create_model_stats_table(db_path)
while True:
next_config = get_next_config(db_path=db_path, loss_bins=loss_bins, num_samples=num_samples)
model_id, cur_loss_bin, cur_num_samples, data_seed, training_seed, cur_smallest_model_count = next_config
if cur_smallest_model_count >= config['output.target_model_count']:
print(f"Found models greater than target model count:{config['output.target_model_count']}, so ending the search")
break
if config['optimizer.name'] in ["SGD"]:
cur_batch_size = min(cur_num_samples//2, config['optimizer.batch_size'])
cur_model_count = config['model.model_count_times_batch_size']//cur_batch_size
elif config['optimizer.name'] in ["SGDPoison"]:
cur_batch_size = config['optimizer.batch_size']
cur_model_count = config['model.model_count_times_batch_size']//cur_batch_size
elif config['optimizer.name']=="guess":
cur_batch_size = None
cur_model_count = config['model.model_count_times_batch_size']//cur_num_samples
else:
cur_batch_size = None
cur_model_count = config['model.model_count_times_batch_size']//cur_num_samples
get_model_stats_summary(db_path)
print("seed:", training_seed, data_seed)
print("next config:",
json.dumps(
{"num_samples":cur_num_samples,
"loss_bin": cur_loss_bin,
"model_count": cur_model_count}))
es_l, es_u = cur_loss_bin
train_data, train_labels, test_data, test_labels, test_all_data, test_all_labels = get_dataset(num_samples=cur_num_samples, seed=data_seed)
torch.manual_seed(training_seed)
device = torch.device('cuda:0')
model = get_model(model_count=cur_model_count,device=device)
perfect_model_count = 0
perfect_model_weights = []
loss_func = nn.CrossEntropyLoss(reduction='none')
target_model_count_subrun = config['distributed.target_model_count_subrun']
start_time = time.time()
tested_model_count = 0
prior_max = 0
while perfect_model_count < target_model_count_subrun:
# TODO: update this section to take in variable mult + simplifying the way that initializaiton is selected
if config['model.init'] == "uniform":
model.reinitialize()
elif config['model.init'] == "uniform2":
model.reinitialize(mult=2)
elif config['model.init'] == "uniform5":
model.reinitialize(mult=5)
elif config['model.init'] == "sphere100":
model.reinitialize_sphere(mult=100)
elif config['model.init'] == "sphere200":
model.reinitialize_sphere(mult=200)
elif config['model.init'] == "regular":
model.reset_parameters()
optimizer, scheduler = get_optimizer_and_scheduler(model=model)
model_result = train(
train_data, train_labels, test_data, test_labels, model, loss_func, optimizer, scheduler,
batch_size=cur_batch_size, es_u=es_u, test_all_data=test_all_data, test_all_labels=test_all_labels)
with torch.no_grad():
train_loss, train_acc = calculate_loss_acc(train_data, train_labels, model_result.forward_normalize, loss_func, batch_size=cur_batch_size)
if train_acc.max() > prior_max:
print("max train acc:", train_acc.max().detach().cpu().item())
prior_max = train_acc.max()
print("tested_model_count", tested_model_count)
# filtering models based on loss threshold
perfect_model_idxs = ((es_l< train_loss) & (train_loss <= es_u) & (train_acc == 1.0))
perfect_model_count_cur = perfect_model_idxs.sum().detach().cpu().item()
perfect_model_count += perfect_model_count_cur
tested_model_count += cur_model_count
if perfect_model_idxs.sum() > 0:
if perfect_model_count > target_model_count_subrun:
remain_count = perfect_model_count_cur - (perfect_model_count - target_model_count_subrun)
tested_model_count -= ((perfect_model_count - target_model_count_subrun)/perfect_model_count_cur)*cur_model_count
perfect_model_weights.append(model_result.get_weights_by_idx(perfect_model_idxs.nonzero().squeeze(1)[:remain_count]))
else:
perfect_model_weights.append(model_result.get_weights_by_idx(perfect_model_idxs))
if len(perfect_model_weights) == 0:
print(f"Failed to find a good model for set up {cur_num_samples} {cur_loss_bin}")
else:
train_time = time.time() - start_time
print("="*50)
# test that the model weights can be reloaded
# concatenating the weights learned into a single model
good_models_state_dict = dict()
cat_dim = 1 if config['model.arch'] in ["mlp", "linear"] else 0
for k in perfect_model_weights[0].keys():
good_models_state_dict[k] = torch.cat(
[d[k].cpu() for d in perfect_model_weights], dim=cat_dim
)
if config['model.arch'] == "mlp":
kwargs = {"input_dim": 2,
"output_dim": 2,
"layers": config['model.mlp.layers'],
"hidden_units": config['model.mlp.hidden_units'],
"model_count": target_model_count_subrun}
new_models = MLPModels(**kwargs, device=torch.device('cpu'))
elif config['model.arch'] == "linear":
kwargs = {"input_dim": model.input_dim,
"output_dim": model.output_dim,
"model_count": target_model_count_subrun}
new_models = LinearModels(**kwargs, device=torch.device('cpu'))
elif config['model.arch'] == "lenet":
kwargs = {"output_dim": config['dataset.mnistcifar.num_classes'],
"width_factor": config['model.lenet.width'],
"model_count": target_model_count_subrun,
"dataset": config['dataset.name'],
"feature_dim": config['model.lenet.feature_dim']}
new_models = LeNetModels(**kwargs)
new_models.load_state_dict(good_models_state_dict)
# show norm of the model
model_l2_norm = 0
model_linf_norm = 0
for para in new_models.parameters():
model_l2_norm += (para**2).sum()
model_linf_norm = max(para.abs().max(), model_linf_norm)
model_l2_norm = model_l2_norm ** 0.5
print(f"model l2 norm: {model_l2_norm}")
print(f"model linf norm: {model_linf_norm}")
with torch.no_grad():
loss, acc = calculate_loss_acc(train_data.cpu(), train_labels.cpu(), new_models, loss_func, batch_size=1)
test_loss, test_acc = calculate_loss_acc(test_all_data.cpu(), test_all_labels.cpu(), new_models, loss_func, batch_size=1)
print(f"verify that train acc is 100%: {acc.mean().item()}")
print(f"test acc: {test_acc.mean().item(): 0.3f} ({test_acc.max().item(): 0.3f} , {test_acc.min().item(): 0.3f} )")
# saving the models
os.makedirs(os.path.join(config['output.folder'], "models"), exist_ok=True)
output_path = build_model_output_path(config, training_seed, data_seed, cur_num_samples)
print(f"Saving models at: {output_path}")
# run specific features that are saved only for evaluate_minimas.py,these are used for resumming models
saveconfig = convert_config_to_dict(config)
saveconfig['dataset.num_samples'] = cur_num_samples
saveconfig['training.seed'] = training_seed
saveconfig['dataset.seed'] = data_seed
saveconfig['training.es_l'], saveconfig['training.es_u'] = cur_loss_bin
# save the model
torch.save({"kwargs": kwargs,
"good_models_state_dict": good_models_state_dict,
"config": saveconfig},
output_path)
update_model_stats_table(
db_path,
model_id=model_id, data_seed=data_seed,
training_seed=training_seed,
num_training_samples=cur_num_samples,
loss_bin_l=es_l,
loss_bin_u=es_u,
test_acc=test_acc.mean().item(),
train_time=train_time,
perfect_model_count=target_model_count_subrun,
tested_model_count=tested_model_count,
save_path=output_path, status="COMPLETE")