-
Notifications
You must be signed in to change notification settings - Fork 5
/
optuna_pelican_cov.py
303 lines (232 loc) · 13.3 KB
/
optuna_pelican_cov.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
import logging
import os
import sys
from src.trainer import which
if which('nvidia-smi') is not None:
min=40000
deviceid = 0
name, mem = os.popen('"nvidia-smi" --query-gpu=gpu_name,memory.total --format=csv,nounits,noheader').read().split('\n')[deviceid].split(',')
print(mem)
mem = int(mem)
if mem < min:
print('Less GPU memory than requested. Terminating.')
sys.exit()
import torch
from torch.utils.data import DataLoader
import optuna
from optuna.trial import TrialState
from src.models import PELICANRegression
from src.models import tests
from src.trainer import Trainer
from src.trainer import init_argparse, init_file_paths, init_logger, init_cuda, logging_printout, fix_args
from src.trainer import init_optimizer, init_scheduler
from src.models.metrics_cov import metrics, minibatch_metrics, minibatch_metrics_string
from src.models.lorentz_metric import normsq4, dot4
from src.dataloaders import initialize_datasets, collate_fn
# This makes printing tensors more readable.
torch.set_printoptions(linewidth=1000, threshold=100000, sci_mode=False)
logger = logging.getLogger('')
def suggest_params(args, trial):
# # args.lr_init = trial.suggest_loguniform("lr_init", 0.0005, 0.005)
# # args.num_epoch = trial.suggest_int("num_epoch", 40, 80, step=10)
# # args.lr_final = trial.suggest_loguniform("lr_final", 1e-8, 1e-5)
# # args.scale = trial.suggest_loguniform("scale", 1e-2, 3)
# # args.sig = trial.suggest_categorical("sig", [True, False])
args.drop_rate = trial.suggest_float("drop_rate", 0, 0.05, step=0.01)
args.drop_rate_out = trial.suggest_float("drop_rate_out", 0, 0.25, step=0.05)
args.weight_decay = trial.suggest_float("weight_decay", 0, 0.025, step=0.005)
# # args.layernorm = trial.suggest_categorical("layernorm", [True, False])
# # args.lr_decay_type = trial.suggest_categorical("lr_decay_type", ['exp', 'cos'])
# # args.batch_size = trial.suggest_categorical("batch_size", [16, 32])
args.double = trial.suggest_categorical("double", [False, True])
# args.factorize = trial.suggest_categorical("factorize", [False, True])
# args.nobj = trial.suggest_int("nobj", 50, 90)
# # args.ir_safe = trial.suggest_categorical("ir_safe", [False, True])
# args.masked = trial.suggest_categorical("masked", [False, True])
# args.config = trial.suggest_categorical("config", ["s", "m", "S", "M"]) # , "sM", "Sm"]) #, "S", "m", "M", "sS", "mM", "sM", "Sm", "SM"]) #, "mx", "Mx", "sSm", "sSM", "smM", "sMmM", "mxn", "mXN", "mxMX", "sXN", "smxn"])
# args.config_out = trial.suggest_categorical("config_out", ["s", "m", "S", "M"]) # , "sM", "Sm"]) #, "S", "m", "M", "sS", "mM", "sM", "Sm", "SM"]) #, "mx", "Mx", "sSm", "sSM", "smM", "sMmM", "mxn", "mXN", "mxMX", "sXN", "smxn"])
# n_layers1 = trial.suggest_int("n_layers1", 2, 6)
n_layers1 = 5
# # n_layersm = trial.suggest_int("n_layersm", 1, 2)
n_layersm = 1
args.num_channels_m = [[trial.suggest_int('n_channelsm['+str(k)+']', 20, 60) for k in range(n_layersm)]] * n_layers1
# n_layersm = [trial.suggest_int("n_layersm", 1, 2) for i in range(n_layers1)]
# args.num_channels_m = [[trial.suggest_int('n_channelsm['+str(i)+', '+str(k)+']', 10, 50) for k in range(n_layersm[i])] for i in range(n_layers1)]
# n_layersm_out = trial.suggest_int("n_layersm2", 1, 2)
# args.num_channels_m_out = [trial.suggest_int('n_channelsm_out['+str(k)+']', 10, 50) for k in range(n_layersm_out)]
args.num_channels_2to2 = [trial.suggest_int("n_channels1["+str(i)+"]", 10, 40) for i in range(n_layers1)]
# # args.num_channels_2to2 = [trial.suggest_int("n_channels1", 3, 30)]
# # args.num_channels_2to2 = args.num_channels_2to2 * (n_layers1) + [args.num_channels_m[0][0] if n_layersm > 0 else args.num_channels_2to2[0]]
# # args.num_channels_2to2 = [trial.suggest_int("n_channels1", 1, 10)] * n_layers1
# # args.num_channels_m = [[trial.suggest_int("n_channels1", 1, 10), args.num_channels_2to2[0]*15*len(args.config)]] * n_layers1
# # args.num_channels_2to2 = args.num_channels_2to2 + [args.num_channels_m[0][0]]
args.num_channels_m_out = [trial.suggest_int("n_channels_m_out["+str(i)+"]", 20, 60) for i in range(2)]
# n_layers2 = trial.suggest_int("n_layers2", 1, 2)
n_layers2 = 1
args.num_channels_out = [trial.suggest_int("n_channels2["+str(i)+"]", 20, 60) for i in range(n_layers2)]
# # args.activation = trial.suggest_categorical("activation", ["elu", "leakyrelu"]) #, "relu", "silu", "selu", "tanh"])
# # args.optim = trial.suggest_categorical("optim", ["adamw", "sgd", "amsgrad", "rmsprop", "adam"])
# # args.activate_agg = trial.suggest_categorical("activate_agg", [True, False])
# # args.activate_lin = trial.suggest_categorical("activate_lin", [True, False])
# # args.dropout = trial.suggest_categorical("dropout", [True])
# # args.batchnorm = trial.suggest_categorical("batchnorm", ['b'])
# trial.suggest_float("c1", 0., 0.002)
# trial.suggest_float("c2", 0., 0.002)
return args
def define_model(trial):
# Initialize arguments
args = init_argparse()
# Initialize file paths
args = init_file_paths(args)
# Initialize logger
init_logger(args)
# Suggest parameters to optuna to optimize over
args = suggest_params(args, trial)
# Write input paramaters and paths to log
git_status = logging_printout(args, trial)
# Fix possible inconsistencies in arguments
args = fix_args(args)
trial.set_user_attr("git_status", git_status)
trial.set_user_attr("args", vars(args))
# Initialize device and data type
device, dtype = init_cuda(args)
# Initialize model
model = PELICANRegression(args.num_channels_m, args.num_channels_2to2, args.num_channels_out, args.num_channels_m_out, num_targets=args.num_targets,
activate_agg=args.activate_agg, activate_lin=args.activate_lin,
activation=args.activation, add_beams=args.add_beams, sig=args.sig, config=args.config, config_out=args.config_out, average_nobj=args.nobj_avg,
factorize=args.factorize, masked=args.masked, softmasked=args.softmasked,
activate_agg_out=args.activate_agg_out, activate_lin_out=args.activate_lin_out, mlp_out=args.mlp_out,
scale=args.scale, ir_safe=args.ir_safe, dropout = args.dropout, drop_rate=args.drop_rate, drop_rate_out=args.drop_rate_out, batchnorm=args.batchnorm,
task = args.task,
device=device, dtype=dtype)
model.to(device)
return args, model, device, dtype
def define_dataloader(args):
# Initialize dataloder
args, datasets = initialize_datasets(args, args.datadir, num_pts=None)
# Construct PyTorch dataloaders from datasets
collate = lambda data: collate_fn(data, scale=args.scale, nobj=args.nobj, add_beams=args.add_beams, beam_mass=args.beam_mass)
dataloaders = {split: DataLoader(dataset,
batch_size=args.batch_size,
shuffle=args.shuffle if (split == 'train') else False,
num_workers=args.num_workers,
collate_fn=collate)
for split, dataset in datasets.items()}
return args, dataloaders
def objective(trial):
args, model, device, dtype = define_model(trial)
args, dataloaders = define_dataloader(args)
trial.set_user_attr("seed", args.seed)
if distributed:
model = torch.nn.DataParallel(model)
# Initialize the scheduler and optimizer
if args.task.startswith('eval'):
optimizer = scheduler = None
restart_epochs = []
summarize = False
else:
optimizer = init_optimizer(args, model, len(dataloaders['train']))
scheduler, restart_epochs, summarize_csv, summarize = init_scheduler(args, optimizer)
# Define a loss function. This is the loss function whose gradients are actually computed.
def mass(x):
norm=normsq4(x)
return norm.sign() * (norm + 1e-12).abs().sqrt()
loss_fn_inv = lambda predict, targets: (normsq4(predict - targets).abs() + 1e-12).sqrt().mean()
loss_fn_m2 = lambda predict, targets: (normsq4(predict) - normsq4(targets)).abs().mean()
loss_fn_m = lambda predict, targets: (mass(predict) - mass(targets)).abs().mean()
loss_fn_3d = lambda predict, targets: (predict[...,1:] - targets[...,1:]).norm(dim=-1).mean()
loss_fn_4d = lambda predict, targets: (predict-targets).norm(dim=-1).mean()
loss_fn = lambda predict, targets: 0.05 * loss_fn_m(predict,targets) + 0.01 * loss_fn_3d(predict, targets) #0.03 * loss_fn_inv(predict,targets) + 0.01 * loss_fn_4d(predict, targets)
# Apply the covariance and permutation invariance tests.
if args.test:
raise NotImplementedError()
# Instantiate the training class
trainer = Trainer(args, dataloaders, model, loss_fn, metrics, minibatch_metrics, minibatch_metrics_string, optimizer, scheduler, restart_epochs, summarize_csv, summarize, device, dtype)
# Load from checkpoint file. If no checkpoint file exists, automatically does nothing.
trainer.load_checkpoint()
# Train model.
best_epoch, best_metrics = trainer.train(trial=trial, metric_to_report='loss')
print(f"Best epoch was {best_epoch} with metrics {best_metrics}")
if args.optuna_test:
# Test predictions on best model.
best_metrics=trainer.evaluate(splits=['test'], best=True, final=False)
trial.set_user_attr("best_test_metrics", best_metrics)
return best_metrics['loss']
if __name__ == '__main__':
# Initialize arguments
args = init_argparse()
if args.storage == 'remote':
storage=optuna.storages.RDBStorage(url=f'postgresql://{os.environ["USER"]}:{args.password}@{args.host}:{args.port}', heartbeat_interval=100) # For running on nodes with a distributed file system
elif args.storage == 'local':
storage=optuna.storages.RDBStorage(url='sqlite:///'+args.study_name+'.db') # For running on a local machine
# direction = 'minimize'
directions=['minimize', 'minimize']
if args.sampler.lower() == 'random':
sampler = optuna.samplers.RandomSampler()
elif args.sampler.lower().startswith('tpe'):
sampler = optuna.samplers.TPESampler(n_startup_trials=5, multivariate=True, group=True, constant_liar=True)
if args.pruner == 'hyperband':
pruner = optuna.pruners.HyperbandPruner()
elif args.pruner == 'median':
pruner = optuna.pruners.MedianPruner(n_warmup_steps=20, n_min_trials=10)
elif args.pruner == 'none':
pruner = None
# os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"
study = optuna.create_study(study_name=args.study_name, storage=storage, directions=directions, load_if_exists=True,
pruner=pruner, sampler=sampler)
# init_params = {
# # 'activate_agg': False,
# # 'activate_lin': True,
# 'activation': 'leakyrelu',
# 'batch_size': 32,
# 'config': 'learn',
# # 'lr_final': 1e-07,
# # 'lr_init': 0.001,
# # 'scale': 0.33,
# # 'num_epoch': 60,
# # 'sig': False,
# 'n_channelsm[0, 0]': 35,
# # 'n_channelsm[0, 1]': 25,
# 'n_channels1[0]': 35,
# 'n_channelsm[1, 0]': 20,
# # 'n_channelsm[1, 1]': 20,
# 'n_channels1[1]': 20,
# 'n_channelsm[2, 0]': 20,
# # 'n_channelsm[2, 1]': 15,
# 'n_channels1[2]': 20,
# 'n_channelsm[3, 0]': 15,
# # 'n_channelsm[3, 1]': 20,
# 'n_channels1[3]': 15,
# 'n_channelsm[4, 0]': 25,
# # 'n_channelsm[4, 1]': 25,
# 'n_channels1[4]': 25,
# 'n_channelsm[5, 0]': 35,
# 'n_channels1[5]': 35,
# 'n_channels1[6]': 35,
# 'n_layers2': 1,
# 'n_channels2[0]': 25,
# 'n_layers1': 6,
# 'n_layersm[0]': 1,
# 'n_layersm[1]': 1,
# 'n_layersm[2]': 1,
# 'n_layersm[3]': 1,
# 'n_layersm[4]': 1,
# 'n_layersm[5]': 1,
# # 'layernorm' : False,
# # 'drop_rate' : 0.15,
# # 'optim': 'adamw',
# }
# study.enqueue_trial(init_params)
study.optimize(objective, n_trials=200, callbacks=[optuna.study.MaxTrialsCallback(200, states=(TrialState.COMPLETE,))])
pruned_trials = study.get_trials(deepcopy=False, states=[TrialState.PRUNED])
complete_trials = study.get_trials(deepcopy=False, states=[TrialState.COMPLETE])
print("Study statistics: ")
print(" Number of finished trials: ", len(study.trials))
print(" Number of pruned trials: ", len(pruned_trials))
print(" Number of complete trials: ", len(complete_trials))
print("Best trial:")
trial = study.best_trial
print(" Value: ", trial.value)
print(" Params: ")
for key, value in trial.params.items():
print(" {}: {}".format(key, value))