/
train_doge.py
197 lines (187 loc) · 9.76 KB
/
train_doge.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
from ensurepip import version
from genericpath import isfile
import os, argparse
import numpy as np
import torch
torch.set_default_dtype(torch.float32)
from datetime import datetime
import warnings
warnings.filterwarnings("ignore", ".*does not have many workers.*")
warnings.filterwarnings("ignore", ".*if you want to see logs for the training epoch.*")
from pytorch_lightning import Trainer, seed_everything
from pytorch_lightning.loggers import TensorBoardLogger
from pytorch_lightning.callbacks import ModelCheckpoint, EarlyStopping
import torch
torch.use_deterministic_algorithms(False)
from configs.defaults import get_cfg_defaults
from data.dataloader import get_ilp_gnn_loaders
from doge import DOGE
def get_final_config(args):
cfg = get_cfg_defaults()
if (hasattr(args, 'config_file')) and os.path.exists(args.config_file):
cfg.set_new_allowed(True)
cfg.merge_from_file(args.config_file)
orig_root_dir = cfg.OUTPUT_ROOT_DIR
cfg.merge_from_list(args.opts)
cfg.freeze()
output_dir = os.path.join(cfg.OUTPUT_ROOT_DIR, cfg.OUT_REL_DIR)
os.makedirs(output_dir, exist_ok=True)
path = os.path.join(output_dir, "config.yaml")
with open(path, 'w') as yaml_file:
cfg.dump(stream = yaml_file, default_flow_style=False)
print('USING FOLLOWING CONFIG:')
print(cfg)
print("Wrote config file at: {}".format(path))
return cfg, output_dir, orig_root_dir
def save_best_ckpt_cfg(cfg, output_dir, best_ckpt_path):
cfg.defrost()
cfg.MODEL.CKPT_PATH = best_ckpt_path
cfg.freeze()
path = os.path.join(output_dir, "config_best.yaml")
with open(path, 'w') as yaml_file:
cfg.dump(stream = yaml_file, default_flow_style=False)
def find_ckpt(root_dir, ckpt_rel_path, find_best_ckpt = False):
if ckpt_rel_path is not None and os.path.isfile(ckpt_rel_path):
return ckpt_rel_path
if ckpt_rel_path is not None and os.path.isfile(os.path.join(root_dir, ckpt_rel_path)):
return os.path.join(root_dir, ckpt_rel_path)
versions = os.path.join(root_dir, 'default')
if not os.path.isdir(versions):
return None
for folder in sorted(os.listdir(versions)):
ckpt_folder = os.path.join(versions, folder, 'checkpoints')
if not find_best_ckpt:
possible_path = os.path.join(ckpt_folder, 'last.ckpt')
if os.path.isfile(possible_path):
print(f'Found checkpoint: {possible_path}')
return possible_path
else:
for ckpt in os.listdir(ckpt_folder):
if ckpt.endswith('.ckpt') and 'epoch' in ckpt:
return os.path.join(ckpt_folder, ckpt)
return None
def main(args):
print(datetime.now().time())
cfg, output_dir, orig_output_dir = get_final_config(args)
seed_everything(cfg.SEED)
# gpus = 0
# if cfg.DEVICE == 'gpu':
# gpus = [0]
# gpu_id = get_freer_gpu()
# if gpu_id >= 0:
# print(f'Using GPU: {gpu_id}')
# gpus = [gpu_id]
# wandb.tensorboard.patch(root_logdir = output_dir)
# wandb.init(project=os.path.basename(output_dir), sync_tensorboard=True)
tb_logger = TensorBoardLogger(output_dir, default_hp_metric=False, max_queue = 1000, flush_secs = 60)
ckpt_path = None
if cfg.MODEL.CKPT_PATH is not None or (args.eval_only and not args.only_test_non_learned):
ckpt_path = find_ckpt(cfg.OUTPUT_ROOT_DIR, cfg.MODEL.CKPT_PATH, args.eval_best_ckpt)
if ckpt_path is None:
ckpt_path = find_ckpt(orig_output_dir, cfg.MODEL.CKPT_PATH, args.eval_best_ckpt)
assert ckpt_path is None or os.path.isfile(ckpt_path), f'CKPT: {ckpt_path} not found.'
checkpoint_callback = ModelCheckpoint(save_last = True, save_on_train_epoch_end = True, mode = 'max', save_top_k = 1, monitor = 'train_last_round_lb', verbose = True)
patience_mult = 4 if cfg.TRAIN.USE_REPLAY_BUFFER else 2 # train for more iterations if using replay buffer since samples from trajectories are 'off-policy'.
early_stopping = EarlyStopping('train_last_round_lb',
patience = patience_mult * cfg.TRAIN.MAX_NUM_EPOCHS // cfg.TRAIN.NUM_JOURNEYS,
check_on_train_epoch_end = True,
mode = 'max')
num_sanity_val_steps = 0
# if args.test_non_learned:
# num_sanity_val_steps = -1
trainer = Trainer(deterministic=False, # due to https://github.com/pyg-team/pytorch_geometric/issues/3175#issuecomment-1047886622
accelerator = 'gpu',
max_epochs = cfg.TRAIN.MAX_NUM_EPOCHS,
default_root_dir=output_dir,
check_val_every_n_epoch = cfg.TEST.VAL_PERIOD,
logger = tb_logger,
num_sanity_val_steps = num_sanity_val_steps,
log_every_n_steps=cfg.LOG_EVERY,
gradient_clip_val=cfg.TRAIN.GRAD_CLIP_VAL,
callbacks=[checkpoint_callback, early_stopping],
detect_anomaly = False)
combined_train_loader, val_loaders, val_datanames, test_loaders, test_datanames = get_ilp_gnn_loaders(cfg,
skip_dual_solved = True,
test_only = args.eval_only,
test_precision_double = not args.test_precision_float,
test_on_train = args.test_on_train,
train_precision_double = args.train_precision_double)
if 'SLURM_JOB_ID' in os.environ:
job_id = os.environ['SLURM_JOB_ID']
job_file_path = f'out_dual/slurm_new/{job_id}.out'
if os.path.isfile(job_file_path):
os.symlink(os.path.abspath(job_file_path), os.path.join(output_dir, f'{job_id}.out'))
if ckpt_path is not None:
print(f'Loading checkpoint and hyperparameters from: {ckpt_path}')
model = DOGE.load_from_checkpoint(ckpt_path,
num_test_rounds = cfg.TEST.NUM_ROUNDS,
num_dual_iter_test = cfg.TEST.NUM_DUAL_ITERATIONS,
dual_improvement_slope_test = cfg.TEST.DUAL_IMPROVEMENT_SLOPE,
val_datanames = val_datanames,
test_datanames = test_datanames,
non_learned_updates_test = args.test_non_learned,
only_test_non_learned = args.only_test_non_learned)
else:
print(f'Initializing from scratch.')
model = DOGE.from_config(cfg,
num_test_rounds = cfg.TEST.NUM_ROUNDS,
num_dual_iter_test = cfg.TEST.NUM_DUAL_ITERATIONS,
dual_improvement_slope_test = cfg.TEST.DUAL_IMPROVEMENT_SLOPE,
val_datanames = val_datanames,
test_datanames = test_datanames,
non_learned_updates_test = args.test_non_learned,
only_test_non_learned = args.only_test_non_learned)
if not args.eval_only:
if args.train_precision_double:
torch.set_default_dtype(torch.float64)
model = model.to(torch.float64)
else:
torch.set_default_dtype(torch.float32)
model = model.to(torch.float32)
trainer.fit(model, combined_train_loader, val_loaders)
save_best_ckpt_cfg(cfg, output_dir, checkpoint_callback.best_model_path)
# Use the checkpoint with the best training performance for testing.
model = DOGE.load_from_checkpoint(checkpoint_callback.best_model_path,
num_test_rounds = cfg.TEST.NUM_ROUNDS,
num_dual_iter_test = cfg.TEST.NUM_DUAL_ITERATIONS,
dual_improvement_slope_test = cfg.TEST.DUAL_IMPROVEMENT_SLOPE,
val_datanames = val_datanames,
test_datanames = test_datanames,
non_learned_updates_test = args.test_non_learned)
combined_train_loader = None
val_loaders = None
if args.test_precision_float:
model = model.to(torch.float32)
torch.set_default_dtype(torch.float32)
else:
model = model.to(torch.float64)
torch.set_default_dtype(torch.float64)
model.eval()
trainer.test(model, dataloaders = test_loaders)
else:
model.eval()
if not args.test_precision_float:
model = model.to(torch.float64)
torch.set_default_dtype(torch.float64)
trainer.test(model, dataloaders = test_loaders)
print(datetime.now().time())
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--config-file", default="", metavar="FILE", help="path to config file")
parser.add_argument("--eval-only", action="store_true", help="perform evaluation only")
parser.add_argument("--eval-best-ckpt", action="store_true", help="perform evaluation on best ckpt instead of last")
parser.add_argument("--test-non-learned", action="store_true", help="Runs FastDOG updates.")
parser.add_argument("--only-test-non-learned", action="store_true", help="Only runs FastDOG updates.")
parser.add_argument("--test-precision-float", action="store_true", help="Performs testing in FP32 format. Recommended to not set due to numerical issues in FP32.")
parser.add_argument("--test-on-train", action="store_true", help="Performs testing on training data.")
parser.add_argument('--train-precision-double', action="store_true", help="double precision training.")
parser.add_argument(
"opts",
help="Modify config options by adding 'KEY VALUE' pairs at the end of the command. ",
default=None,
nargs=argparse.REMAINDER,
)
args = parser.parse_args()
print("Command Line Args:")
print(args)
main(args)