-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
315 lines (260 loc) · 16.9 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
import pdb
import torch
from torch.utils.data import DataLoader
import numpy as np
from data_model.gen_sparse_coding_data import gen_z, gen_M, gen_Winit, gen_epsilon
from config import config
from pathlib import Path
import pickle
from common_args import parse_args
from dataset.masked_sparse_contr_dataset import MaskedSparseContrastiveDataset
from dataset.multimask_sparse_contr_dataset import MultiMaskedSparseContrastiveDataset, multi_mask_data_collate
from models.sparse_contrastive_model import SparseContrastiveModel
from models.sparse_contrastive_ml_model import SparseContrastiveMultiLayeredModel
from models.sparse_contrastive_no_alt import SparseContrastiveModelNoAlter
from models.sparse_contrastive_model_reduced import SparseContrastiveModelReduced
from functions.train import alternate_train, train
import wandb
from torch.optim.lr_scheduler import CosineAnnealingLR, StepLR
def main():
args = parse_args()
print('config.n ', config.nn)
print('config.p ', config.p)
print('config.d ', config.d)
print('config.m ', config.m)
print('config.has_target_predictor ', config.has_target_predictor)
print('config.has_target_ReLU ', config.has_target_ReLU)
print('config.lr ' , config.lr)
print('config.sigma0 ', config.sigma0)
print('args.normalize_repr', args.normalize_repr)
print('args.ema_decay', args.ema_decay)
print('args.use_multimasking', args.use_multimasking)
print('args.use_bn', args.use_bn)
print('args.temperature', args.temperature)
print('args.use_pred', args.use_pred)
print('args.m_identity', args.m_identity)
device = 'cpu'
if torch.cuda.is_available():
device = torch.device('cuda:{}'.format(args.local_rank))
print(f"Using device {device}")
# Save Result
root_output_dir = Path(config.LOG_DIR)
# set up logger
if not root_output_dir.exists():
print('=> creating {}'.format(root_output_dir))
root_output_dir.mkdir()
log_metrics = args.log_metrics
logger = None
sigma0 = None
online_scheduler = None
target_scheduler = None
# number of augmentations if multimask
n_aug = args.n_aug
bias_val = args.bias_val
if not args.const_bias:
bias_val = "trained"
for ws_noise in config.ws_noise_levels: #[1, 1.25, 1.5, 2]:
for sigma0 in config.gaussian_noise_levels:
for sparsity in config.sparsity_levels:
for maskprob in config.masking_probs:
for i in range(config.num_exp):
print(f"Experiment {i+1}")
np.random.seed(i+1)
torch.manual_seed(i+1)
# pnb == pred no bias
if log_metrics:
run = wandb.init(project=args.wandb_project, reinit=True, name=f"trial-{i}",
group=f"{args.model}-cn-{args.use_alt_norm}-rn-{args.use_row_norm}-I{args.m_identity}-bn-{args.use_bn}-" +
f"norm-{args.normalize_repr}-p{config.p}-m{config.m}-d{config.d}-c{ws_noise}-bias-{bias_val}-sp-{sparsity}-" +
f"mask{maskprob}-lr-{config.lr}-1h-{config.one_hot_latent}",
config=config)
logger = wandb.log
# Generate data
if args.m_identity:
M = np.eye(config.p)
else:
M = gen_M(p=config.p, d=config.d)
Z = gen_z(n=config.nn, d=config.d, prob=sparsity, one_hot_latent=config.one_hot_latent)
Epsilon = gen_epsilon(n=config.nn, p=config.p, d=config.d, sigma0 = sigma0)
X = (M @ Z + Epsilon).T
if args.use_multimasking:
dataset = MultiMaskedSparseContrastiveDataset(data=X, Z=Z.T, prob_ones=1-maskprob, n_aug=n_aug)
train_loader = DataLoader(dataset, batch_size=config.batch_size, shuffle=True, drop_last=False, collate_fn=multi_mask_data_collate)
else:
dataset = MaskedSparseContrastiveDataset(data=X, Z=Z.T, prob_ones=1-maskprob)
train_loader = DataLoader(dataset, batch_size=config.batch_size, shuffle=True, drop_last=False)
Wo_init = gen_Winit(M, c=ws_noise, m=config.m, d=config.d, p=config.p)
Wt_init = gen_Winit(M, c=ws_noise, m=config.m, d=config.d, p=config.p)
if args.model == 'simplified':
print(f"Running {args.model}...")
# Initialize model
model = SparseContrastiveModelReduced(Wo_init=Wo_init,
Wt_init=Wt_init,
m=config.m,
p=config.p,
d=config.d,
has_online_ReLU=config.has_online_ReLU,
has_target_ReLU = config.has_target_ReLU,
normalize_rep=args.normalize_repr,
device=device,
use_bn=args.use_bn,
use_bias=args.use_bias,
const_bias=args.const_bias,
const_bias_val=bias_val
)
model = model.to(device)
online_optimizer = torch.optim.SGD(list(model.parameters()), lr=config.lr)
target_optimizer = torch.optim.SGD(list(model.parameters()), lr=config.lr)
assert not args.use_alt_norm == True or not args.use_row_norm == True, "we cannot normalize both rows and cols"
val_dict = alternate_train(model, online_optimizer=online_optimizer,
target_optimizer=target_optimizer,
train_loader=train_loader,
max_epochs=config.NUM_EPOCHES,
M=M,
ema_decay=args.ema_decay,
log_metrics=log_metrics,
logger=logger,
online_scheduler=online_scheduler,
target_scheduler=target_scheduler,
col_norm=args.use_alt_norm,
row_norm=args.use_row_norm,
clip_bias=args.clip_bias
)
elif args.model == 'simplified-pred':
print(f"Running {args.model}...")
# Initialize model
model = SparseContrastiveModel(Wo_init=Wo_init,
Wt_init=Wt_init,
m=config.m,
p=config.p,
d=config.d,
has_online_ReLU=config.has_online_ReLU,
has_target_ReLU = config.has_target_ReLU,
normalize_rep=args.normalize_repr,
device=device,
use_bn=args.use_bn,
use_pred=True,
use_pred_bias=args.use_pred_bias
)
model = model.to(device)
online_optimizer = torch.optim.SGD(list(model.parameters()), lr=config.lr)
target_optimizer = torch.optim.SGD(list(model.parameters()), lr=config.lr)
# online_optimizer = torch.optim.SGD(list(model.Wo.parameters()) + list(model.Wp.parameters()), lr=config.lr)
# target_optimizer = torch.optim.SGD(list(model.Wt.parameters()) + list(model.Wp.parameters()), lr=config.lr)
val_dict = alternate_train(model, online_optimizer=online_optimizer,
target_optimizer=target_optimizer,
train_loader=train_loader,
max_epochs=config.NUM_EPOCHES,
M=M,
ema_decay=args.ema_decay,
log_metrics=log_metrics,
logger=logger,
online_scheduler=online_scheduler,
target_scheduler=target_scheduler,
col_norm=args.use_alt_norm,
row_norm=args.use_row_norm,
)
elif args.model == 'simplified-pred-norm':
print(f"Running {args.model}...")
# Initialize model
model = SparseContrastiveModel(Wo_init=Wo_init,
Wt_init=Wt_init,
m=config.m,
p=config.p,
d=config.d,
has_online_ReLU=config.has_online_ReLU,
has_target_ReLU = config.has_target_ReLU,
normalize_rep=args.normalize_repr,
device=device,
use_bn=args.use_bn,
use_pred=True,
use_pred_bias=args.use_pred_bias
)
model = model.to(device)
online_optimizer = torch.optim.SGD(list(model.parameters()), lr=config.lr)
target_optimizer = torch.optim.SGD(list(model.parameters()), lr=config.lr)
val_dict = alternate_train(model, online_optimizer=online_optimizer,
target_optimizer=target_optimizer,
train_loader=train_loader,
max_epochs=config.NUM_EPOCHES,
M=M,
ema_decay=args.ema_decay,
log_metrics=log_metrics,
logger=logger,
online_scheduler=online_scheduler,
target_scheduler=target_scheduler,
col_norm=args.use_alt_norm,
row_norm=args.use_row_norm,
pred_norm=True
)
elif args.model == 'simplified-noalt-pred':
print(f"Running {args.model}...")
# Initialize model
model = SparseContrastiveModelNoAlter(Wo_init=Wo_init,
Wt_init=Wt_init,
m=config.m,
p=config.p,
d=config.d,
has_online_ReLU=config.has_online_ReLU,
has_target_ReLU = config.has_target_ReLU,
normalize_rep=args.normalize_repr,
device=device,
use_bn=args.use_bn,
use_pred=args.use_pred,
linear_pred=False,
use_pred_bias=args.use_pred_bias
)
model = model.to(device)
optimizer = torch.optim.SGD(model.parameters(), lr=config.lr)
val_dict = train(model, optimizer=optimizer,
train_loader=train_loader,
max_epochs=config.NUM_EPOCHES,
M=M,
log_metrics=log_metrics,
logger=logger,
col_norm=args.use_alt_norm,
row_norm=args.use_row_norm,
pred_norm=True
)
elif args.model == 'simplified-ml':
Wo1_init = gen_Winit(M, c=ws_noise, m=config.m, d=config.d, p=config.m)
Wt1_init = gen_Winit(M, c=ws_noise, m=config.m, d=config.d, p=config.m)
print(f"Running {args.model}...")
# Initialize model
model = SparseContrastiveMultiLayeredModel(Wo_init=Wo_init,
Wt_init=Wt_init,
Wo1_init=Wo1_init,
Wt1_init=Wt1_init,
m=config.m,
p=config.p,
d=config.d,
has_online_ReLU=config.has_online_ReLU,
has_target_ReLU = config.has_target_ReLU,
normalize_rep=args.normalize_repr,
device=device,
use_bn=args.use_bn
)
model = model.to(device)
online_optimizer = torch.optim.SGD(list(model.parameters()), lr=config.lr)
target_optimizer = torch.optim.SGD(list(model.parameters()), lr=config.lr)
# online_optimizer = torch.optim.SGD(list(model.Wo.parameters()), lr=config.lr)
# target_optimizer = torch.optim.SGD(list(model.Wt.parameters()), lr=config.lr)
val_dict = alternate_train(model, online_optimizer=online_optimizer,
target_optimizer=target_optimizer,
train_loader=train_loader,
max_epochs=config.NUM_EPOCHES,
M=M,
log_metrics=log_metrics,
logger=logger,
online_scheduler=None,
target_scheduler=None
)
if log_metrics:
run.finish()
with open(root_output_dir.joinpath(f"training_val_dict_c{ws_noise}_noise{sigma0}_sparse{sparsity}_mask{maskprob}_ema{args.ema_decay}_experiment{i}.pkl"), 'wb') as f:
pickle.dump(val_dict, f)
torch.save({
'model_state_dict': model.state_dict()
}, root_output_dir.joinpath('final_model.pt'))
if __name__ == '__main__':
main()