-
Notifications
You must be signed in to change notification settings - Fork 23
/
tree.py
729 lines (605 loc) · 32 KB
/
tree.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
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
"""
Main script for training an adaptive neural tree (ANT).
"""
from __future__ import print_function
import argparse
import os
import sys
import json
import time
import numpy as np
import torch
from torch.autograd import Variable
import torch.nn.functional as F
from torch import optim
import matplotlib
matplotlib.use('agg')
from data import get_dataloaders, get_dataset_details
from models import Tree, One
from ops import get_params_node
from utils import define_node, get_scheduler, set_random_seed
from visualisation import visualise_routers_behaviours
# Experiment settings
parser = argparse.ArgumentParser(description='Adaptive Neural Trees')
parser.add_argument('--experiment', '-e', dest='experiment', default='tree', help='experiment name')
parser.add_argument('--subexperiment','-sube', dest='subexperiment', default='', help='experiment name')
parser.add_argument('--dataset', default='mnist', help='dataset type')
parser.add_argument('--no-cuda', action='store_true', default=False, help='disables CUDA training')
parser.add_argument('--gpu', type=str, default="", help='which GPU to use')
parser.add_argument('--seed', type=int, default=0, metavar='S', help='random seed')
parser.add_argument('--num_workers', type=int, default=0, metavar='N', help='number of threads for data-loader')
# Optimization settings:
parser.add_argument('--batch-size', type=int, default=256, metavar='N', help='input batch size for training')
parser.add_argument('--log-interval', type=int, default=10, metavar='N', help='how many batches to wait before logging training status')
parser.add_argument('--augmentation_on', action='store_true', default=False, help='perform data augmentation')
parser.add_argument('--lr', type=float, default=0.001, metavar='LR', help='learning rate')
parser.add_argument('--scheduler', type=str, default="", help='learning rate scheduler')
parser.add_argument('--momentum', type=float, default=0.5, metavar='M', help='SGD momentum')
parser.add_argument('--valid_ratio', '-vr', dest='valid_ratio', type=float, default=0.1, metavar='LR', help='validation set ratio')
parser.add_argument('--criteria', default='avg_valid_loss', help='growth criteria')
parser.add_argument('--epochs_node', type=int, default=50, metavar='N', help='max number of epochs to train per node during the growth phase')
parser.add_argument('--epochs_finetune', type=int, default=100, metavar='N', help='number of epochs for the refinement phase')
parser.add_argument('--epochs_patience', type=int, default=5, metavar='N', help='number of epochs to be waited without improvement at each node during the growth phase')
parser.add_argument('--maxdepth', type=int, default=10, help='maximum depth of tree')
parser.add_argument('--finetune_during_growth', action='store_true', default=False, help='refine the tree globally during the growth phase')
parser.add_argument('--epochs_finetune_node', type=int, default=1, metavar='N', help='number of epochs to perform global refinement at each node during the growth phase')
# Solver, router and transformer modules:
parser.add_argument('--router_ver', '-r_ver', dest='router_ver', type=int, default=1, help='default router version')
parser.add_argument('--router_ngf', '-r_ngf', dest='router_ngf', type=int, default=1, help='number of feature maps in routing function')
parser.add_argument('--router_k', '-r_k', dest='router_k', type=int, default=28, help='kernel size in routing function')
parser.add_argument('--router_dropout_prob', '-r_drop', dest='router_dropout_prob', type=float, default=0.0, help='drop-out probabilities for router modules.')
parser.add_argument('--transformer_ver', '-t_ver', dest='transformer_ver', type=int, default=1, help='default transformer version: identity')
parser.add_argument('--transformer_ngf', '-t_ngf', dest='transformer_ngf', type=int, default=3, help='number of feature maps in residual transformer')
parser.add_argument('--transformer_k', '-t_k', dest='transformer_k', type=int, default=5, help='kernel size in transfomer function')
parser.add_argument('--transformer_expansion_rate', '-t_expr', dest='transformer_expansion_rate', type=int, default=1, help='default transformer expansion rate')
parser.add_argument('--transformer_reduction_rate', '-t_redr', dest='transformer_reduction_rate', type=int, default=2, help='default transformer reduction rate')
parser.add_argument('--solver_ver', '-s_ver', dest='solver_ver', type=int, default=1, help='default router version')
parser.add_argument('--solver_inherit', '-s_inh', dest='solver_inherit', action='store_true', help='inherit the parameters of the solver when defining two new ones for splitting a node')
parser.add_argument('--solver_dropout_prob', '-s_drop', dest='solver_dropout_prob', type=float, default=0.0, help='drop-out probabilities for solver modules.')
parser.add_argument('--downsample_interval', '-ds_int', dest='downsample_interval', type=int, default=0, help='interval between two downsampling operations via transformers i.e. 0 = downsample at every transformer')
parser.add_argument('--batch_norm', '-bn', dest='batch_norm', action='store_true', default=False, help='turn batch norm on')
# Visualisation:
parser.add_argument('--visualise_split', action='store_true', help='visuliase how the test dist is split by the routing function')
args = parser.parse_args()
# GPUs devices:
args.cuda = not args.no_cuda and torch.cuda.is_available()
if args.gpu:
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu
# Set the seed for repeatability
set_random_seed(args.seed, args.cuda)
# Define a dictionary for post-inspection of the model:
records = vars(args)
records['time'] = 0.0
records['counter'] = 0 # number of optimization steps
records['train_nodes'] = [] # node indices for each logging interval
records['train_loss'] = [] # avg. train. loss for each log interval
records['train_best_loss'] = np.inf # best train. loss
records['train_epoch_loss'] = [] # epoch wise train loss
records['valid_nodes'] = []
records['valid_best_loss_nodes'] = []
records['valid_best_loss_nodes_split'] = []
records['valid_best_loss_nodes_ext'] = []
records['valid_best_root_nosplit'] = np.inf
records['valid_best_loss'] = np.inf
records['valid_best_accuracy'] = 0.0
records['valid_epoch_loss'] = []
records['valid_epoch_accuracy'] = []
records['test_best_loss'] = np.inf
records['test_best_accuracy'] = 0.0
records['test_epoch_loss'] = []
records['test_epoch_accuracy'] = []
# ----------------------------- Data loaders ---------------------------------
train_loader, valid_loader, test_loader, NUM_TRAIN, NUM_VALID = get_dataloaders(
args.dataset, args.batch_size, args.augmentation_on,
cuda=args.cuda, num_workers=args.num_workers,
)
args.input_nc, args.input_width, args.input_height, args.classes = \
get_dataset_details(args.dataset)
args.no_classes = len(args.classes)
# ----------------------------- Components ----------------------------------
def train(model, data_loader, optimizer, node_idx):
""" Train step"""
model.train()
train_loss = 0
no_points = 0
train_epoch_loss = 0
# train the model
for batch_idx, (x, y) in enumerate(data_loader):
optimizer.zero_grad()
if args.cuda:
x, y = x.cuda(), y.cuda()
x, y = Variable(x), Variable(y)
y_pred, p_out = model(x)
loss = F.nll_loss(y_pred, y)
train_epoch_loss += loss.data[0] * y.size(0)
train_loss += loss.data[0] * y.size(0)
loss.backward()
optimizer.step()
records['counter'] += 1
no_points += y.size(0)
if batch_idx % args.log_interval == 0:
# show the interval-wise average loss:
train_loss /= no_points
records['train_loss'].append(train_loss)
records['train_nodes'].append(node_idx)
sys.stdout.flush()
sys.stdout.write('\t [{}/{} ({:.0f}%)] Loss: {:.6f} \r'.
format(batch_idx*len(x), NUM_TRAIN,
100. * batch_idx / NUM_TRAIN, train_loss))
train_loss = 0
no_points = 0
# compute average train loss for the epoch
train_epoch_loss /= NUM_TRAIN
records['train_epoch_loss'].append(train_epoch_loss)
if train_epoch_loss < records['train_best_loss']:
records['train_best_loss'] = train_epoch_loss
print('\nTrain set: Average loss: {:.4f}'.format(train_epoch_loss))
def valid(model, data_loader, node_idx, struct):
""" Validation step """
model.eval()
valid_epoch_loss = 0
correct = 0
for data, target in data_loader:
if args.cuda:
data, target = data.cuda(), target.cuda()
data, target = Variable(data, volatile=True), Variable(target)
output = model(data)
# sum up batch loss
valid_epoch_loss += F.nll_loss(
output, target, size_average=False,
).data[0]
pred = output.data.max(1, keepdim=True)[1]
correct += pred.eq(target.data.view_as(pred)).cpu().sum()
valid_epoch_loss /= NUM_VALID
valid_epoch_accuracy = 100. * correct / NUM_VALID
records['valid_epoch_loss'].append(valid_epoch_loss)
records['valid_epoch_accuracy'].append(valid_epoch_accuracy)
if valid_epoch_loss < records['valid_best_loss']:
records['valid_best_loss'] = valid_epoch_loss
if valid_epoch_accuracy > records['valid_best_accuracy']:
records['valid_best_accuracy'] = valid_epoch_accuracy
# see if the current node is root and undergoing the initial training
# prior to the growth phase.
is_init_root_train = not model.split and not model.extend and node_idx == 0
# save the best split model during node-wise training as model_tmp.pth
if not is_init_root_train and model.split and \
valid_epoch_loss < records['valid_best_loss_nodes_split'][node_idx]:
records['valid_best_loss_nodes_split'][node_idx] = valid_epoch_loss
checkpoint_model('model_tmp.pth', model=model)
checkpoint_msc(struct, records)
# save the best extended model during node-wise training as model_ext.pth
if not is_init_root_train and model.extend and \
valid_epoch_loss < records['valid_best_loss_nodes_ext'][node_idx]:
records['valid_best_loss_nodes_ext'][node_idx] = valid_epoch_loss
checkpoint_model('model_ext.pth', model=model)
checkpoint_msc(struct, records)
# separately store best performance for the initial root training
if is_init_root_train \
and valid_epoch_loss < records['valid_best_root_nosplit']:
records['valid_best_root_nosplit'] = valid_epoch_loss
checkpoint_model('model_tmp.pth', model=model)
checkpoint_msc(struct, records)
# saving model during the refinement (fine-tuning) phase
if not is_init_root_train and \
valid_epoch_loss < records['valid_best_loss_nodes'][node_idx]:
records['valid_best_loss_nodes'][node_idx] = valid_epoch_loss
if not model.split and not model.extend:
checkpoint_model('model_tmp.pth', model=model)
checkpoint_msc(struct, records)
end = time.time()
records['time'] = end - start
print(
'Valid set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)'
'\nTook {} seconds. '.format(
valid_epoch_loss, correct, NUM_VALID,
100. * correct / NUM_VALID, records['time'],
),
)
return valid_epoch_loss
def test(model, data_loader):
""" Test step """
model.eval()
test_loss = 0
correct = 0
for data, target in data_loader:
if args.cuda:
data, target = data.cuda(), target.cuda()
data, target = Variable(data, volatile=True), Variable(target)
output = model(data)
test_loss += F.nll_loss(output, target, size_average=False).data[0]
pred = output.data.max(1, keepdim=True)[1]
correct += pred.eq(target.data.view_as(pred)).cpu().sum()
test_loss /= len(data_loader.dataset)
test_accuracy = 100. * correct / len(data_loader.dataset)
records['test_epoch_loss'].append(test_loss)
records['test_epoch_accuracy'].append(test_accuracy)
if test_loss < records['test_best_loss']:
records['test_best_loss'] = test_loss
if test_accuracy > records['test_best_accuracy']:
records['test_best_accuracy'] = test_accuracy
end = time.time()
print(
'Test set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)'
'\nTook {} seconds. '.format(
test_loss, correct, len(data_loader.dataset),
100. * correct / len(data_loader.dataset), end - start,
),
)
def _load_checkpoint(model_file_name):
save_dir = "./experiments/{}/{}/{}/{}".format(
args.dataset, args.experiment, args.subexperiment, 'checkpoints',
)
model = torch.load(save_dir + '/' + model_file_name)
if args.cuda:
model.cuda()
return model
def checkpoint_model(model_file_name, struct=None, modules=None, model=None, figname='hist.png', data_loader=None):
if not(os.path.exists(os.path.join("./experiments", args.dataset, args.experiment, args.subexperiment))):
os.makedirs(os.path.join("./experiments", args.dataset, args.experiment, args.subexperiment, 'figures'))
os.makedirs(os.path.join("./experiments", args.dataset, args.experiment, args.subexperiment, 'checkpoints'))
# If model is not given, then build one.
if not(model) and modules and struct:
model = Tree(struct, modules, cuda_on=args.cuda)
# save the model:
save_dir = "./experiments/{}/{}/{}/{}".format(args.dataset, args.experiment, args.subexperiment, 'checkpoints')
model_path = save_dir + '/' + model_file_name
torch.save(model, model_path)
print("Model saved to {}".format(model_path))
# save tree histograms:
if args.visualise_split and not(data_loader is None):
save_hist_dir = "./experiments/{}/{}/{}/{}".format(args.dataset, args.experiment, args.subexperiment, 'figures')
visualise_routers_behaviours(model, data_loader, fig_scale=6, axis_font=20, subtitle_font=20,
cuda_on=args.cuda, objects=args.classes, plot_on=False,
save_as=save_hist_dir + '/' + figname)
def checkpoint_msc(struct, data_dict):
""" Save structural information of the model and experimental results.
Args:
struct (list) : list of dictionaries each of which contains
meta information about each node of the tree.
data_dict (dict) : data about the experiment (e.g. loss, configurations)
"""
if not(os.path.exists(os.path.join("./experiments", args.dataset, args.experiment, args.subexperiment))):
os.makedirs(os.path.join("./experiments", args.dataset, args.experiment, args.subexperiment, 'figures'))
os.makedirs(os.path.join("./experiments", args.dataset, args.experiment, args.subexperiment, 'checkpoints'))
# save the tree structures as a json file:
save_dir = "./experiments/{}/{}/{}/{}".format(args.dataset,args.experiment,args.subexperiment,'checkpoints')
struct_path = save_dir + "/tree_structures.json"
with open(struct_path, 'w') as f:
json.dump(struct, f)
print("Tree structure saved to {}".format(struct_path))
# save the dictionary as jason file:
dict_path = save_dir + "/records.json"
with open(dict_path, 'w') as f_d:
json.dump(data_dict, f_d)
print("Other data saved to {}".format(dict_path))
def get_decision(criteria, node_idx, tree_struct):
""" Define the splitting criteria
Args:
criteria (str): Growth criteria.
node_idx (int): Index of the current node.
tree_struct (list) : list of dictionaries each of which contains
meta information about each node of the tree.
Returns:
The function returns one of the following strings
'split': split the node
'extend': extend the node
'keep': keep the node as it is
"""
if criteria == 'always': # always split or extend
if tree_struct[node_idx]['valid_accuracy_gain_ext'] > tree_struct[node_idx]['valid_accuracy_gain_split'] > 0.0:
return 'extend'
else:
return 'split'
elif criteria == 'avg_valid_loss':
if tree_struct[node_idx]['valid_accuracy_gain_ext'] > tree_struct[node_idx]['valid_accuracy_gain_split'] and \
tree_struct[node_idx]['valid_accuracy_gain_ext'] > 0.0:
print("Average valid loss is reduced by {} ".format(tree_struct[node_idx]['valid_accuracy_gain_ext']))
return 'extend'
elif tree_struct[node_idx]['valid_accuracy_gain_split'] > 0.0:
print("Average valid loss is reduced by {} ".format(tree_struct[node_idx]['valid_accuracy_gain_split']))
return 'split'
else:
print("Average valid loss is aggravated by split/extension."
" Keep the node as it is.")
return 'keep'
else:
raise NotImplementedError(
"specified growth criteria is not available. ",
)
def optimize_fixed_tree(
model, tree_struct, train_loader,
valid_loader, test_loader, no_epochs, node_idx,
):
""" Train a tree with fixed architecture.
Args:
model (torch.nn.module): tree model
tree_struct (list): list of dictionaries which contain information
about all nodes in the tree.
train_loader (torch.utils.data.DataLoader) : data loader of train data
valid_loader (torch.utils.data.DataLoader) : data loader of valid data
test_loader (torch.utils.data.DataLoader) : data loader of test data
no_epochs (int): number of epochs for training
node_idx (int): index of the node you want to optimize
Returns:
returns the trained model and newly added nodes (if grown).
"""
# get if the model is growing or fixed
grow = (model.split or model.extend)
# define optimizer and trainable parameters
params, names = get_params_node(grow, node_idx, model)
for i, (n, p) in enumerate(model.named_parameters()):
if not(n in names):
# print('(Fix) ' + n)
p.requires_grad = False
else:
# print('(Optimize) ' + n)
p.requires_grad = True
for i, p in enumerate(params):
if not(p.requires_grad):
print("(Grad not required)" + names[i])
optimizer = optim.Adam(
filter(lambda p: p.requires_grad, params), lr=args.lr,
)
if args.scheduler:
scheduler = get_scheduler(args.scheduler, optimizer, grow)
# monitor nodewise best valid loss:
if not(not(grow) and node_idx==0) and len(records['valid_best_loss_nodes']) == node_idx:
records['valid_best_loss_nodes'].append(np.inf)
if not(not(grow) and node_idx==0) and len(records['valid_best_loss_nodes_split']) == node_idx:
records['valid_best_loss_nodes_split'].append(np.inf)
if not(not(grow) and node_idx==0) and len(records['valid_best_loss_nodes_ext']) == node_idx:
records['valid_best_loss_nodes_ext'].append(np.inf)
# start training
min_improvement = 0.0 # acceptable improvement in loss for early stopping
valid_loss = np.inf
patience_cnt = 1
for epoch in range(1, no_epochs + 1):
print("\n----- Layer {}, Node {}, Epoch {}/{}, Patience {}/{}---------".
format(tree_struct[node_idx]['level'], node_idx,
epoch, no_epochs, patience_cnt, args.epochs_patience))
train(model, train_loader, optimizer, node_idx)
valid_loss_new = valid(model, valid_loader, node_idx, tree_struct)
# learning rate scheduling:
if args.scheduler == 'plateau':
scheduler.step(valid_loss_new)
elif args.scheduler == 'step_lr':
scheduler.step()
test(model, test_loader)
if not((valid_loss-valid_loss_new) > min_improvement) and grow:
patience_cnt += 1
valid_loss = valid_loss_new*1.0
if patience_cnt > args.epochs_patience > 0:
print('Early stopping')
break
# load the node-wise best model based on validation accuracy:
if no_epochs > 0 and grow:
if model.extend:
print('return the node-wise best extended model')
model = _load_checkpoint('model_ext.pth')
else:
print('return the node-wise best split model')
model = _load_checkpoint('model_tmp.pth')
# return the updated models:
tree_modules = model.update_tree_modules()
if model.split:
child_left, child_right = model.update_children()
return model, tree_modules, child_left, child_right
elif model.extend:
child_extension = model.update_children()
return model, tree_modules, child_extension
else:
return model, tree_modules
def grow_ant_nodewise():
"""The main function for optimising an ANT """
# ############## 0: Define the root node and optimise ###################
# define the root node:
tree_struct = [] # stores graph information for each node
tree_modules = [] # stores modules for each node
root_meta, root_module = define_node(
args, node_index=0, level=0, parent_index=-1, tree_struct=tree_struct,
)
tree_struct.append(root_meta)
tree_modules.append(root_module)
# train classifier on root node (no split no extension):
model = Tree(
tree_struct, tree_modules, split=False, extend=False, cuda_on=args.cuda,
)
if args.cuda:
model.cuda()
# optimise
model, tree_modules = optimize_fixed_tree(
model, tree_struct,
train_loader, valid_loader, test_loader, args.epochs_node, node_idx=0,
)
checkpoint_model('model.pth', struct=tree_struct, modules=tree_modules)
checkpoint_msc(tree_struct, records)
# ######################## 1: Growth phase starts ########################
nextind = 1
last_node = 0
for lyr in range(args.maxdepth):
print("---------------------------------------------------------------")
print("\nAt layer " + str(lyr))
for node_idx in range(len(tree_struct)):
change = False
if tree_struct[node_idx]['is_leaf'] and not(tree_struct[node_idx]['visited']):
print("\nProcessing node " + str(node_idx))
# -------------- Define children candidate nodes --------------
# ---------------------- (1) Split ----------------------------
# left child
identity = True
meta_l, node_l = define_node(
args,
node_index=nextind, level=lyr+1,
parent_index=node_idx, tree_struct=tree_struct,
identity=identity,
)
# right child
meta_r, node_r = define_node(
args,
node_index=nextind+1, level=lyr+1,
parent_index=node_idx, tree_struct=tree_struct,
identity=identity,
)
# inheriting solver modules to facilitate optimization:
if args.solver_inherit and meta_l['identity'] and meta_r['identity'] and not(node_idx == 0):
node_l['classifier'] = tree_modules[node_idx]['classifier']
node_r['classifier'] = tree_modules[node_idx]['classifier']
# define a tree with a new split by adding two children nodes:
model_split = Tree(tree_struct, tree_modules,
split=True, node_split=node_idx,
child_left=node_l, child_right=node_r,
extend=False,
cuda_on=args.cuda)
# -------------------- (2) Extend ----------------------------
# define a tree with node extension
meta_e, node_e = define_node(
args,
node_index=nextind,
level=lyr+1,
parent_index=node_idx,
tree_struct=tree_struct,
identity=False,
)
# Set the router at the current node as one-sided One().
# TODO: this is not ideal as it changes tree_modules
tree_modules[node_idx]['router'] = One()
# define a tree with an extended edge by adding a node
model_ext = Tree(tree_struct, tree_modules,
split=False,
extend=True, node_extend=node_idx,
child_extension=node_e,
cuda_on=args.cuda)
# ---------------------- Optimise -----------------------------
best_tr_loss = records['train_best_loss']
best_va_loss = records['valid_best_loss']
best_te_loss = records['test_best_loss']
print("\n---------- Optimizing a binary split ------------")
if args.cuda:
model_split.cuda()
# split and optimise
model_split, tree_modules_split, node_l, node_r \
= optimize_fixed_tree(model_split, tree_struct,
train_loader, valid_loader, test_loader,
args.epochs_node,
node_idx)
best_tr_loss_after_split = records['train_best_loss']
best_va_loss_adter_split = records['valid_best_loss_nodes_split'][node_idx]
best_te_loss_after_split = records['test_best_loss']
tree_struct[node_idx]['train_accuracy_gain_split'] \
= best_tr_loss - best_tr_loss_after_split
tree_struct[node_idx]['valid_accuracy_gain_split'] \
= best_va_loss - best_va_loss_adter_split
tree_struct[node_idx]['test_accuracy_gain_split'] \
= best_te_loss - best_te_loss_after_split
print("\n----------- Optimizing an extension --------------")
if not(meta_e['identity']):
if args.cuda:
model_ext.cuda()
# make deeper and optimise
model_ext, tree_modules_ext, node_e \
= optimize_fixed_tree(model_ext, tree_struct,
train_loader, valid_loader, test_loader,
args.epochs_node,
node_idx)
best_tr_loss_after_ext = records['train_best_loss']
best_va_loss_adter_ext = records['valid_best_loss_nodes_ext'][node_idx]
best_te_loss_after_ext = records['test_best_loss']
# TODO: record the gain from split/extra depth:
# need separately record best losses for split & depth
tree_struct[node_idx]['train_accuracy_gain_ext'] \
= best_tr_loss - best_tr_loss_after_ext
tree_struct[node_idx]['valid_accuracy_gain_ext'] \
= best_va_loss - best_va_loss_adter_ext
tree_struct[node_idx]['test_accuracy_gain_ext'] \
= best_te_loss - best_te_loss_after_ext
else:
print('No extension as '
'the transformer is an identity function.')
# ---------- Decide whether to split, extend or keep -----------
criteria = get_decision(args.criteria, node_idx, tree_struct)
if criteria == 'split':
print("\nSplitting node " + str(node_idx))
# update the parent node
tree_struct[node_idx]['is_leaf'] = False
tree_struct[node_idx]['left_child'] = nextind
tree_struct[node_idx]['right_child'] = nextind+1
tree_struct[node_idx]['split'] = True
# add the children nodes
tree_struct.append(meta_l)
tree_modules_split.append(node_l)
tree_struct.append(meta_r)
tree_modules_split.append(node_r)
# update tree_modules:
tree_modules = tree_modules_split
nextind += 2
change = True
elif criteria == 'extend':
print("\nExtending node " + str(node_idx))
# update the parent node
tree_struct[node_idx]['is_leaf'] = False
tree_struct[node_idx]['left_child'] = nextind
tree_struct[node_idx]['extended'] = True
# add the children nodes
tree_struct.append(meta_e)
tree_modules_ext.append(node_e)
# update tree_modules:
tree_modules = tree_modules_ext
nextind += 1
change = True
else:
# revert weights back to state before split
print("No splitting at node " + str(node_idx))
print("Revert the weights to the pre-split state.")
model = _load_checkpoint('model.pth')
tree_modules = model.update_tree_modules()
# record the visit to the node
tree_struct[node_idx]['visited'] = True
# save the model and tree structures:
checkpoint_model('model.pth', struct=tree_struct, modules=tree_modules,
data_loader=test_loader,
figname='hist_split_node_{:03d}.png'.format(node_idx))
checkpoint_msc(tree_struct, records)
last_node = node_idx
# global refinement prior to the next growth
# NOTE: this is an option not included in the paper.
if args.finetune_during_growth and (criteria == 1 or criteria == 2):
print("\n-------------- Global refinement --------------")
model = Tree(tree_struct, tree_modules,
split=False, node_split=last_node,
extend=False, node_extend=last_node,
cuda_on=args.cuda)
if args.cuda:
model.cuda()
model, tree_modules = optimize_fixed_tree(
model, tree_struct,
train_loader, valid_loader, test_loader,
args.epochs_finetune_node, node_idx,
)
# terminate the tree growth if no split or extend in the final layer
if not change: break
# ############### 2: Refinement (finetuning) phase starts #################
print("\n\n------------------- Fine-tuning the tree --------------------")
best_valid_accuracy_before = records['valid_best_accuracy']
model = Tree(tree_struct, tree_modules,
split=False,
node_split=last_node,
child_left=None, child_right=None,
extend=False,
node_extend=last_node, child_extension=None,
cuda_on=args.cuda)
if args.cuda:
model.cuda()
model, tree_modules = optimize_fixed_tree(model, tree_struct,
train_loader, valid_loader, test_loader,
args.epochs_finetune,
last_node)
best_valid_accuracy_after = records['valid_best_accuracy']
# only save if fine-tuning improves validation accuracy
if best_valid_accuracy_after - best_valid_accuracy_before > 0:
checkpoint_model('model.pth', struct=tree_struct, modules=tree_modules,
data_loader=test_loader,
figname='hist_split_node_finetune.png')
checkpoint_msc(tree_struct, records)
# --------------------------- Start growing an ANT! ---------------------------
start = time.time()
grow_ant_nodewise()