-
Notifications
You must be signed in to change notification settings - Fork 8
/
main.py
504 lines (409 loc) · 19.3 KB
/
main.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
"""Main training function.
(A) NodeGAM training.
- Set name. E.g. 0615_bikeshare.
- Set dataset. E.g. bikeshare.
- Set arch. E.g. GAMAtt.
- Set ga2m = 1 if training a GA2M, or 0 for training GAM.
- Output:
- The final test accuracy is stored in the results/bikeshare_GAMAtt.csv.
- The best model is stored in the logs/0615_bikeshare/best.ckpt.
- The hyperparameter is stored in logs/hparams/0615_bikeshare.
- The training and validation loss figure is in loss_figs/0615_bikeshare.jpg.
- The training and validation results are stored in logs/0615_bikeshare/recorder.json,
loss_history.npy (training loss history per step), and err_history.npy (val err history).
"""
import argparse
import json
import os
import pickle
import shutil
import sys
import time
from os.path import join as pjoin, exists as pexists
from pathlib import Path
import matplotlib
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import torch
from qhoptim.pyt import QHAdam
from sklearn.model_selection import train_test_split
import nodegam
# Don't use multiple gpus; If more than 1 gpu, just use first one
if torch.cuda.device_count() > 1:
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
# Use it to create figure instead of interactive
matplotlib.use('Agg')
def get_args():
parser = argparse.ArgumentParser(description="")
parser.add_argument("--name",
default='debug',
help="Name of this run. Used for monitoring and checkpointing.")
# Load hparams from a model name
parser.add_argument("--load_from_hparams", type=str, default=None)
parser.add_argument('--seed', type=int, default=None,
help='seed for initializing training.')
# My own arguments
parser.add_argument("--dataset", default='bikeshare',
help="Choose the dataset.",
choices=['year', 'epsilon', 'a9a', 'higgs', 'microsoft',
'yahoo', 'click', 'mimic2', 'adult', 'churn',
'credit', 'compas', 'support2', 'mimic3',
'rossmann', 'wine', 'bikeshare'])
parser.add_argument('--fold', type=int, default=0,
help='Choose from 0 to 4, as we only support 5-fold CV.')
parser.add_argument("--arch", type=str, default='GAM',
choices=['ODST', 'GAM', 'GAMAtt'])
parser.add_argument("--num_trees", type=int, default=1024)
parser.add_argument("--num_layers", type=int, default=2)
parser.add_argument("--depth", type=int, default=1)
parser.add_argument("--addi_tree_dim", type=int, default=0)
parser.add_argument("--l2_lambda", type=float, default=0.)
parser.add_argument("--lr", type=float, default=0.02)
parser.add_argument("--lr_warmup_steps", type=int, default=-1)
parser.add_argument("--lr_decay_steps", type=int, default=-1,
help='Decay learning rate by 1/5 if not improving for this step')
parser.add_argument("--quantile_dist", type=str, default='normal',
choices=['normal', 'uniform'],
help='Which distribution to do qunatile transform')
parser.add_argument("--early_stopping_rounds", type=int, default=11000)
parser.add_argument("--max_rounds", type=int, default=-1)
parser.add_argument("--max_time", type=float, default=3600 * 20) # At most 20 hours
parser.add_argument("--report_frequency", type=int, default=100)
parser.add_argument("--batch_size", type=int, default=None)
parser.add_argument("--max_bs", type=int, default=2048,
help='If batch size is None, it automatically finds the right batch size '
'that fits into the GPU memory between max_bs and min_bs via binary '
'search.')
parser.add_argument("--min_bs", type=int, default=128)
parser.add_argument("--random_search", type=int, default=0)
parser.add_argument('--fp16', type=int, default=0,
help='Uses the 16-precision to train. Slows down 5~10%, but saves memory '
'and is slightly better.')
parser.add_argument('--data_subsample', type=float, default=1.,
help='Between 0 and 1. Percentage of training data used. If bigger than 1, '
'treats it as integer and select the specified number of samples.')
parser.add_argument('--ignore_prev_runs', type=int, default=0,
help='If 1, in random search, it ignores previous runs and reruns the '
'training even it was run before. Useful when fixing a bug.')
temp_args, _ = parser.parse_known_args()
# Remove stuff if in debug mode
if temp_args.name.startswith('debug'):
clean_up(temp_args.name)
# Load previous hparams arch
prev_hparams = load_from_prev_hparams(temp_args)
if prev_hparams is not None and all([not arg.startswith('--arch') for arg in sys.argv]):
temp_args.arch = prev_hparams['arch']
parser = getattr(nodegam.arch, temp_args.arch + 'Block').add_model_specific_args(parser)
args = parser.parse_args()
# If loading previous hparams, update prev hparams with user inputs
user_hparams = load_user_hparams(parser)
if prev_hparams is not None:
update_args(args, user_hparams, prev_hparams)
return args, user_hparams
def load_from_prev_hparams(args):
hparams = None
if pexists(pjoin('logs', 'hparams', args.name)):
with open(pjoin('logs', 'hparams', args.name)) as fp:
hparams = json.load(fp)
elif args.load_from_hparams is not None:
path = args.load_from_hparams
if '/' not in args.load_from_hparams:
path = pjoin('logs', 'hparams', args.load_from_hparams)
with open(path) as fp:
hparams = json.load(fp)
return hparams
def load_user_hparams(parser):
for action in parser._actions:
action.default = argparse.SUPPRESS
return vars(parser.parse_args())
def clean_up(name):
shutil.rmtree(pjoin('logs', name), ignore_errors=True)
shutil.rmtree(pjoin('lightning_logs', name), ignore_errors=True)
if pexists(pjoin('logs', 'hparams', name)):
os.remove(pjoin('logs', 'hparams', name))
def update_args(args, user_hparams, prev_hparams):
for k, v in prev_hparams.items():
if k not in user_hparams:
setattr(args, k, v)
def main():
args, user_hparams = get_args()
if args.random_search == 0:
try:
train(args)
finally:
if pexists(pjoin('is_running', args.name)): # release it
os.remove(pjoin('is_running', args.name))
sys.exit()
def get_rs_name(hparams, rs_hparams):
if isinstance(hparams, argparse.Namespace):
hparams = vars(hparams)
tmp = '_'.join([f'{v["short_name"]}{hparams[k]}'
for k, v in rs_hparams.items()])
tmp += ('' if hparams['data_subsample'] == 1 else f'_ds{hparams["data_subsample"]}')
return tmp
# Create a directory to record what is running
os.makedirs('is_running', exist_ok=True)
rs_hparams = getattr(nodegam.arch, args.arch + 'Block').get_model_specific_rs_hparams()
# This makes every random search as the same order!
if args.seed is not None:
nodegam.utils.seed_everything(args.seed)
orig_name, num_random_search = args.name, args.random_search
args.random_search = 0 # When sending jobs, not run the random search!!
unsearched_set = {k for k in user_hparams if k in rs_hparams and k not in ['seed']}
if len(unsearched_set) > 0:
print('Do not random search following attributes:', unsearched_set)
for r in range(num_random_search):
for _ in range(50): # Try 50 times if can't found, quit
for k, v in rs_hparams.items():
if 'gen' in v and v['gen'] is not None:
if k in user_hparams and k not in ['seed']:
continue
setattr(args, k, v['gen'](args))
args.name = orig_name + '_' + get_rs_name(args, rs_hparams)
if pexists(pjoin('is_running', args.name)):
continue
if (not args.ignore_prev_runs) and pexists(pjoin('logs', args.name, 'MY_IS_FINISHED')):
continue
Path(pjoin('is_running', args.name)).touch()
train(args)
break
else:
print('Can not find any more parameters! Quit.')
sys.exit()
def train(args) -> None:
# Create directory
os.makedirs(pjoin('logs', args.name), exist_ok=True)
if pexists(pjoin('logs', args.name, 'MY_IS_FINISHED')):
print('Quit! Already finish running for %s' % args.name)
return
# Set seed
if args.seed is not None:
nodegam.utils.seed_everything(args.seed)
device = 'cuda' if torch.cuda.is_available() else 'cpu'
# Data
with nodegam.utils.Timer(f'Load dataset {args.dataset}'):
data = nodegam.data.DATASETS[args.dataset.upper()](path='./data', fold=args.fold)
# Dataset-dependent quantile noise. If it's set too small, the categorical features
# will not get enough value. In general 1e-3 is a good value.
qn = data.get('quantile_noise', 1e-3)
preprocessor = nodegam.mypreprocessor.MyPreprocessor(
cat_features=data.get('cat_features', None),
y_normalize=(data['problem'] == 'regression'),
random_state=1337, quantile_transform=True,
output_distribution=args.quantile_dist,
quantile_noise=qn,
)
X_train, y_train = data['X_train'], data['y_train']
preprocessor.fit(X_train, y_train)
if args.data_subsample > 1.:
args.data_subsample = int(args.data_subsample)
if args.data_subsample != 1. and args.data_subsample < X_train.shape[0]:
print(f'Subsample the data by ds={args.data_subsample}')
X_train, _, y_train, _ = train_test_split(
X_train, y_train, train_size=args.data_subsample, random_state=1377,
stratify=(y_train if data['problem'] == 'classification' else None))
use_data_val = ('X_valid' in data and 'y_valid' in data)
if use_data_val:
X_valid, y_valid = data['X_valid'], data['y_valid']
else:
# Merge with the valid set, and cut it ourselves
if 'X_valid' in data:
X_train = pd.concat([X_train, data['X_valid']], axis=0)
y_train = np.concatenate([y_train, data['y_valid']], axis=0)
X_train, X_valid, y_train, y_valid = train_test_split(
X_train, y_train, test_size=0.2, random_state=1377,
stratify=(y_train if data['problem'] == 'classification' else None)
)
# Transform dataset
X_train, y_train = preprocessor.transform(X_train, y_train)
X_valid, y_valid = preprocessor.transform(X_valid, y_valid)
X_test, y_test = preprocessor.transform(data['X_test'], data['y_test'])
# Save preprocessor
with open(pjoin('logs', args.name, 'preprocessor.pkl'), 'wb') as op:
pickle.dump(preprocessor, op)
metric = data.get('metric', ('classification_error'
if data['problem'] == 'classification' else 'mse'))
# Modify args based on the dataset
args.in_features = X_train.shape[1]
args.problem = data['problem']
args.num_classes = data.get('num_classes', 1)
args.data_addi_tree_dim = data.get('addi_tree_dim', 0)
print(f'X_train: {X_train.shape}, X_valid: {X_valid.shape}, X_test: {X_test.shape}')
# Model
model, step_callbacks = getattr(nodegam.arch, args.arch + 'Block').load_model_by_hparams(
args, ret_step_callback=True)
# Initialize bias before sending to cuda
if 'init_bias' in args and args.init_bias and args.problem == 'classification':
model.set_bias(y_train)
model.to(device)
optimizer_params = {'nus': (0.7, 1.0), 'betas': (0.95, 0.998)}
trainer = nodegam.trainer.Trainer(
model=model,
experiment_name=args.name,
warm_start=True, # To handle the interruption on v server
Optimizer=QHAdam,
optimizer_params=optimizer_params,
lr=args.lr,
lr_warmup_steps=args.lr_warmup_steps,
verbose=False,
n_last_checkpoints=5,
step_callbacks=step_callbacks, # Temp annelaing
fp16=args.fp16,
problem=args.problem,
)
assert metric in ['negative_auc', 'classification_error', 'mse']
eval_fn = getattr(trainer, 'evaluate_' + metric)
# Before we start, we will need to select the batch size if unspecified
if args.batch_size is None or args.batch_size < 0:
assert device != 'cpu', 'Have to specify batch size when using CPU'
args.batch_size = choose_batch_size(trainer, X_train, y_train, device,
max_bs=args.max_bs, min_bs=args.min_bs)
else:
# trigger data-aware init
with torch.no_grad():
res = model(torch.as_tensor(X_train[:(2 * args.batch_size)], device=device))
# Then show hparams after deciding the batch size
print("experiment:", args.name)
print("Args:")
print(args)
# Then record hparams
saved_args = pjoin('logs', args.name, 'hparams.json')
json.dump(vars(args), open(saved_args, 'w'))
# record hparams again, since logs/{args.name} will be deleted!
os.makedirs(pjoin('logs', 'hparams'), exist_ok=True)
json.dump(vars(args), open(pjoin('logs', 'hparams', args.name), 'w'))
# To make sure when rerunning the err history and time are accurate,
# we save the whole history in training.json.
recorder = nodegam.recorder.Recorder(path=pjoin('logs', args.name))
st_time = time.time()
is_first_run = True
for batch in nodegam.utils.iterate_minibatches(X_train, y_train,
batch_size=args.batch_size,
shuffle=True, epochs=float('inf')):
# Handle removing missing by sampling from a Gaussian!
metrics = trainer.train_on_batch(*batch, device=device)
if recorder.loss_history is not None:
recorder.loss_history.append(float(metrics['loss']))
if trainer.step % args.report_frequency == 0:
trainer.save_checkpoint()
trainer.remove_old_temp_checkpoints()
trainer.average_checkpoints(out_tag='avg')
trainer.load_checkpoint(tag='avg')
err = eval_fn(X_valid, y_valid, device=device, batch_size=args.batch_size * 2)
if err < recorder.best_err:
recorder.best_err = err
recorder.best_step_err = trainer.step
trainer.save_checkpoint(tag='best')
if recorder.err_history is not None:
recorder.err_history.append(err)
recorder.step = trainer.step
recorder.run_time += float(time.time() - st_time)
st_time = time.time()
recorder.save_record()
trainer.load_checkpoint() # last
if recorder.loss_history is not None and recorder.err_history is not None:
save_loss_fig(recorder.loss_history, recorder.err_history,
pjoin('loss_figs', f'{args.name}.jpg'))
if is_first_run:
print("Step\tVal_Err\tTime(s)")
is_first_run = False
print('{}\t{}\t{:.0f}'.format(trainer.step, np.around(err, 5), recorder.run_time))
bstep = recorder.best_step_err
if isinstance(bstep, list):
bstep = np.max(bstep)
min_steps = max(bstep, getattr(args, 'anneal_steps', -1))
if trainer.step > min_steps + args.early_stopping_rounds:
print('BREAK. There is no improvment for {} steps'.format(args.early_stopping_rounds))
break
if args.lr_decay_steps > 0 \
and trainer.step > bstep + args.lr_decay_steps \
and trainer.step > (recorder.lr_decay_step + args.lr_decay_steps):
lr_before = trainer.lr
trainer.decrease_lr(ratio=0.2, min_lr=1e-6)
recorder.lr_decay_step = trainer.step
print('LR: %.2e -> %.2e' % (lr_before, trainer.lr))
if 0 < args.max_rounds < trainer.step:
print('End. It reaches the maximum rounds %d' % args.max_rounds)
break
if recorder.run_time > args.max_time:
print('End. It reaches the maximum run time %d (s)' % args.max_time)
break
print("Best step: ", recorder.best_step_err)
print("Best Val Error: ", recorder.best_err)
max_step = trainer.step
# Run test time
trainer.load_checkpoint(tag='best')
test_err = eval_fn(X_test, y_test, device=device, batch_size=2 * args.batch_size)
print("Test Error rate: {}".format(test_err))
# Save csv results
results = dict()
results['test_err'] = test_err
results['val_err'] = recorder.best_err
results['best_step_err'] = recorder.best_step_err
results['max_step'] = max_step
results['time(s)'] = '%d' % recorder.run_time
results['fold'] = args.fold
results['fp16'] = args.fp16
results['batch_size'] = args.batch_size
# Append the hyperparameters
rs_hparams = getattr(nodegam.arch, args.arch + 'Block').get_model_specific_rs_hparams()
for k in rs_hparams:
results[k] = getattr(args, k)
results = getattr(nodegam.arch, args.arch + 'Block').add_model_specific_results(results, args)
results['name'] = args.name
os.makedirs(f'results', exist_ok=True)
dataset_postfix = f'_ds{args.data_subsample}' if args.data_subsample != 1. else ''
csv_file = f'results/{args.dataset}{dataset_postfix}_{args.arch}.csv'
nodegam.utils.output_csv(csv_file, results)
print('output results to %s' % csv_file)
# Clean up
open(pjoin('logs', args.name, 'MY_IS_FINISHED'), 'a')
trainer.remove_old_temp_checkpoints(number_ckpts_to_keep=0)
def choose_batch_size(trainer, X_train, y_train, device, max_bs=4096, min_bs=64):
def clean_up_memory():
for p in trainer.model.parameters():
p.grad = None
torch.cuda.empty_cache()
# Starts with biggest batch size. Capped by training size
bs = min(max_bs, X_train.shape[0])
min_bs = min(min_bs, X_train.shape[0])
shuffle_indices = np.random.permutation(X_train.shape[0])
while True:
try:
if bs < min_bs:
raise RuntimeError('The batch size %d is smaller than mininum %d' % (bs, min_bs))
print('Trying batch size %d ...' % bs)
trainer.train_on_batch(
X_train[shuffle_indices[:bs]], y_train[shuffle_indices[:bs]],
device=device, update=False)
break
except RuntimeError as e:
if 'out of memory' not in str(e):
raise e
print('| batch size %d failed.' % (bs))
bs = bs // 2
if bs < min_bs:
raise e
continue
finally:
clean_up_memory()
print('Choose batch size %d.' % (bs))
return bs
def save_loss_fig(loss_history, err_history, path):
os.makedirs(os.path.dirname(path), exist_ok=True)
# At last, save the loss figure
plt.figure(figsize=[18, 6])
plt.subplot(1, 2, 1)
plt.plot(loss_history)
plt.title('Loss')
plt.grid()
plt.subplot(1, 2, 2)
plt.plot(err_history)
plt.title('Error')
plt.grid()
plt.savefig(path, bbox_inches='tight')
# plt.show()
plt.close()
if __name__ == '__main__':
main()