/
base_runner.py
executable file
·406 lines (338 loc) · 17.2 KB
/
base_runner.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
import os
import random
import time
import datetime
from collections import defaultdict, namedtuple
import numpy as np
import torch
import torch.optim as optim
from tqdm import tqdm
from common.logger import TensorboardLogger, FileLogger
from common.utils import (
save_checkpoint,
EarlyStopping
)
from datasets import get_data_handler
import moser
import ffjord
class BaseRunner:
def __init__(self, config, run_dir):
self.cpu = config["cmd"]["cpu"]
if torch.cuda.is_available() and not self.cpu:
self.device = config["cmd"]["local_rank"]
torch.cuda.set_device(self.device)
else:
self.device = "cpu"
self.cpu = True # handle case when `--cpu` isn't specified
# but there are no gpu devices available
print("running on device %s" %self.device)
self.config = config
self.config["cmd"]["checkpoint_dir"] = str(run_dir / "checkpoints")
self.config["cmd"]["results_dir"] = str(run_dir / "results_dir")
self.config["cmd"]["logs_dir"] = str(run_dir / "logs_dir")
if not self.config["cmd"]["is_debug"]:
os.makedirs(self.config["cmd"]["checkpoint_dir"], exist_ok=True)
os.makedirs(self.config["cmd"]["results_dir"], exist_ok=True)
os.makedirs(self.config["cmd"]["logs_dir"], exist_ok=True)
# saves last epoch for cases of continuing a saved experiment
self.last_epoch = 0
self.load()
def load(self):
self.load_seed_from_config()
self.load_dataset()
self.load_model()
self.load_logger()
self.load_optimizer()
self.load_early_stopping()
def load_early_stopping(self):
if "max_time" in self.config["early_stop"]:
self.early_stopping = TimeEarlyStopper(**self.config["early_stop"])
else:
self.early_stopping = EarlyStopping(**self.config["early_stop"])
def load_seed_from_config(self):
# https://pytorch.org/docs/stable/notes/randomness.html
seed = self.config["cmd"]["seed"]
if seed is None:
return
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
def load_logger(self):
self.loggers = []
if not self.config["cmd"]["is_debug"]:
for logger_name in self.config.get("logger", []):
if logger_name == "tensorboard":
logger_class = TensorboardLogger
elif logger_name == "file":
logger_class = FileLogger
else:
raise ValueError("illegal logger %s" %self.config["logger"])
self.loggers.append(logger_class(self.config))
for logger in self.loggers:
logger.watch(self.model)
def load_dataset(self):
print("### Loading dataset: {}".format(self.config["dataset"]["name"]))
data_handler = get_data_handler(self.config["dataset"], eps=self.config["model"].get("eps", 0))
(
self.train_loader,
self.val_loader,
self.test_loader,
) = data_handler.get_dataloaders(
batch_size=int(self.config["optim"]["batch_size"]),
eval_batch_size=int(self.config["optim"]["eval_batch_size"]),
shuffle=self.config["optim"].get("shuffle", True)
)
def load_model(self):
raise NotImplementedError
def load_optimizer(self):
self.optimizer = optim.AdamW(
self.model.parameters(),
lr=self.config["optim"]["lr_initial"],
weight_decay=self.config["optim"]["weight_decay"]
)
if self.config["optim"].get("lr_milestones"):
self.scheduler = optim.lr_scheduler.MultiStepLR(
optimizer=self.optimizer,
milestones=self.config["optim"]["lr_milestones"],
gamma=self.config["optim"]["lr_gamma"]
)
else:
self.scheduler = None
def load_pretrained(self, checkpoint_path=None, ddp_to_dp=False):
if checkpoint_path is None or os.path.isfile(checkpoint_path) is False:
print(f"Checkpoint: {checkpoint_path} not found!")
return False
print("### Loading checkpoint from: {}".format(checkpoint_path))
checkpoint = torch.load(checkpoint_path, map_location=torch.device(self.device))
self.model.load_state_dict(checkpoint["state_dict"])
self.optimizer.load_state_dict(checkpoint["optimizer"])
self.early_stopping.load_state_dict(checkpoint.get("early_stopping", {}))
# self.config = checkpoint["config"]
try:
self.last_epoch = int(np.floor(checkpoint["epoch"]))
except TypeError:
self.last_epoch = 0
return True
def log(self, *args, **kwargs):
for logger in self.loggers:
logger.log(*args, **kwargs)
def save(self, epoch, metrics, name=None, split="val"):
if name is None and self.config["cmd"]["only_save_checkpoints"]:
name = epoch
if not self.config["cmd"]["is_debug"]:
save_checkpoint(
{
"epoch": epoch,
"state_dict": self.model.state_dict(),
"optimizer": self.optimizer.state_dict(),
"config": self.config,
"metrics": metrics,
"early_stopping": self.early_stopping.state_dict()
},
self.config["cmd"]["checkpoint_dir"],
name="%s_%s" %(name, split)
)
def calculate_loss(self, x):
raise NotImplementedError
def calc_metrics(self, *args, **kwargs):
return {}
def start(self):
self.train_loader.dataset.initial_plots(self.config["cmd"]["results_dir"], model=self.model)
self.cummulative_training_time = 0.
self.init_tik = time.perf_counter()
def train(self):
eval_every = self.config["optim"].get("eval_every", -1)
iters = 0
for epoch in range(self.last_epoch, self.config["optim"]["max_epochs"]):
# Print metrics, make plots.
if self.val_loader is not None and eval_every != -1 and epoch % eval_every == 0:
val_metrics = self.validate(
split="val",
epoch=epoch,
)
self.early_stopping(**val_metrics)
if self.config["optim"].get("save_all"):
self.save(epoch + 1, val_metrics)
if self.early_stopping.is_best:
self.save(epoch + 1, val_metrics, name="best")
if self.early_stopping.early_stop:
break
self.model.train()
average_losses = defaultdict(lambda: 0)
for (i, (x, _)) in enumerate(self.train_loader):
self.optimizer.zero_grad()
tik = time.perf_counter()
loss_dict = self.calculate_loss(x.to(self.device))
loss = loss_dict["loss"].mean()
loss.backward()
self.optimizer.step()
tok = time.perf_counter()
self.cummulative_training_time += tok - tik
log_dict = {"epoch": epoch + (i + 1) / len(self.train_loader), "training_step_time": tok-tik, "total_time": tok-self.init_tik, "cummulative_training_time": self.cummulative_training_time}
for key, value in loss_dict.items():
log_dict[key] = value.mean().item()
average_losses[key] += value.sum().item()
log_dict.update(self.calc_metrics())
# Evaluate on val set every `eval_every` iterations.
iters += 1
average_losses = {key: value / len(self.train_loader.dataset) for key, value in average_losses.items()}
if epoch % self.config["cmd"]["print_every"] == 0:
log_str = [
"{}: {:.4f}".format(k, v) for k, v in log_dict.items()
]
print(", ".join(log_str))
self.log(
log_dict,
step=epoch * len(self.train_loader) + i + 1,
split="train",
)
if self.config["optim"].get("save_every") and epoch % self.config["optim"].get("save_every") == 0:
train_metrics = log_dict.copy()
train_metrics.update(average_losses)
self.save(epoch + 1, train_metrics, split='train')
if log_dict["total_time"] > self.config["early_stop"].get("max_time", np.inf):
break
if self.scheduler:
self.scheduler.step()
torch.cuda.empty_cache()
self.validate(split='val', epoch=epoch)
self.finalize()
def validate(self, split="val", epoch=None):
print("### Evaluating on {}.".format(split))
self.model.eval()
if (epoch is None and self.config["cmd"].get("plot_at_test")) or (not self.config["cmd"]["only_save_checkpoints"]):
if epoch is None:
eval_dir = self.config["cmd"]["results_dir"]
else:
eval_dir = os.path.join(self.config["cmd"]["results_dir"], "epoch %s" %epoch)
if not os.path.exists(eval_dir):
os.makedirs(eval_dir)
self.train_loader.dataset.evaluate_model(self.model, epoch, eval_dir)
if split == "test":
self.train_loader.dataset.test_model(self.model, self.config["cmd"]["results_dir"], self.config["optim"].get("calc_ode_density", False))
loader = self.val_loader if split == "val" else self.test_loader
numel = 0
log_dict = defaultdict(lambda: 0.)
for (i, (x, _)) in tqdm(
enumerate(loader),
total=len(loader)
):
loss_dict = self.calculate_loss(x.to(self.device))
for key, value in loss_dict.items():
value = value.detach()
loss_dict[key] = value
log_dict[key] += value.sum().item()
numel += x.shape[0]
del loss_dict
torch.cuda.empty_cache()
log_str = []
for k,v in log_dict.items():
if isinstance(v, dict):
for k2, v2 in v.items():
v2 /= numel
log_dict[k][k2] = v2
log_str.append("{}: {:.4f}".format("%s_%s" %(k, k2), v2))
else:
v /= numel
log_dict[k] = v
log_str.append("{}: {:.4f}".format(k, v))
total_time = time.perf_counter() - self.init_tik
log_dict["total_time"] = total_time
log_dict["cummulative_training_time"] = self.cummulative_training_time
log_str.append("{}: {:.4f}".format("total_time", total_time))
print(", ".join(log_str))
# epoch is None for test split
if epoch is not None:
self.log(
log_dict,
step=(epoch + 1) * len(self.train_loader),
split=split,
)
return dict(log_dict)
def finalize(self):
for logger in self.loggers:
logger.close()
# load best model
if self.config["optim"].get("test_on_best_val"):
checkpoint_path = os.path.join(self.config["cmd"]["checkpoint_dir"], "checkpoint.pt")
if os.path.exists(checkpoint_path):
self.load_pretrained(checkpoint_path)
test_metrics = self.validate(split='test')
self.save(metrics=test_metrics, epoch=None, name="test")
class MoserEarlyStopper(EarlyStopping):
def __call__(self, **loss_dict):
loss_dict.pop("loss")
super().__call__(loss_dict["nll"], **loss_dict)
class TimeEarlyStopper(EarlyStopping):
def __init__(self, max_time, *args, **kwargs):
super().__init__(*args, **kwargs)
self.max_time = max_time
def __call__(self, loss, total_time, **kwargs):
super().__call__(loss, **kwargs)
if total_time > self.max_time:
self.early_stop = True
class MoserRunner(BaseRunner):
def load_early_stopping(self):
if "max_time" in self.config["early_stop"]:
self.early_stopping = TimeEarlyStopper(**self.config["early_stop"])
else:
self.early_stopping = MoserEarlyStopper(**self.config["early_stop"])
def load_model(self):
kwargs = {}
MODELS = {
"torus": moser.TorusMoserFlow,
"implicit": moser.ImplicitMoser
}
model_class = MODELS[self.config["model"]["manifold"]]
if self.config["model"]["manifold"] == "implicit":
surface = self.train_loader.dataset.surface
surface.to(self.device)
kwargs["surface"] = surface
self.model = model_class(self.config["dataset"]["input_dim"],
self.config["model"], self.device, **kwargs)
def calculate_loss(self, x):
nll = self.model(x.to(self.device))
mc_batch_size_scale = self.config["optim"].get("mc_batch_size_scale", 1)
lambda_plus = self.config["optim"].get("lambda_plus", 0)
lambda_minus = self.config["optim"].get("lambda_minus", 1)
positivity_loss = self.model.positivity_loss(lambda_plus, lambda_minus, x.shape[0] * mc_batch_size_scale)
loss = nll.mean() + positivity_loss.mean()
# Print metrics, make plots.
loss_dict = {
"loss": loss,
"nll": nll,
"positivity": positivity_loss,
}
return loss_dict
def validate(self, split="val", epoch=None):
val_metrics = super().validate(split, epoch)
if split == "test" and self.config["optim"].get("calc_ode_density"):
nll = 0.
numel = 0
for (i, (x, _)) in tqdm(
enumerate(self.test_loader),
total=len(self.test_loader)
):
nll -= self.model.direct_log_likelihood(x.to(self.device)).sum()
numel += x.shape[0]
nll /= numel
print("test nll by ode is %s, estimated nll is %s" % (nll, val_metrics["nll"]))
val_metrics["ode_nll"] = nll.item()
else:
val_metrics["ode_nll"] = 0.
return val_metrics
class FFJORDRunner(BaseRunner):
def __init__(self, *largs, **kwargs):
super(FFJORDRunner, self).__init__(*largs, **kwargs)
def load_model(self):
args_dict = self.config["model"]
args = namedtuple("Args", args_dict.keys())(**args_dict)
self.model = ffjord.FFJORDTorusModel(args, self.config["dataset"]["input_dim"], self.device)
def calculate_loss(self, x):
loss = self.model(x)
return {"loss": loss, "nll": loss}
def calc_metrics(self):
return {"nfe": ffjord.count_nfe(self.model)}