/
main.py
218 lines (164 loc) · 6.44 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
from functools import partial
import os
import logging
import hydra
import pytorch_lightning as pl
from pytorch_lightning.loggers import TensorBoardLogger, CSVLogger, WandbLogger
import torch
import torch.nn.functional as F
from dib import DIBWrapper, DIBLoss
from utils import (
prod,
MLP,
CrossEntropyLossGeneralize,
get_exponential_decay_gamma,
create_folders,
)
from data import MyCIFAR10DataModule
logger = logging.getLogger(__name__)
@hydra.main(config_name="config")
def main(cfg):
pl.seed_everything(cfg.seed)
cfg.paths.base_dir = hydra.utils.get_original_cwd()
create_folders(cfg.paths.base_dir, ["results", "logs", "pretrained"])
# 1 PLAYER GAME SETTING (BOB)
cfg.current_mode = "1player"
# dataset arguments will be used by model
datamodule = MyCIFAR10DataModule(**cfg.data.kwargs)
cfg.data.x_shape = prod(datamodule.dims) # will flatten input for MLP encoder
cfg.data.n_classes = datamodule.num_classes
cfg.data.n_train = datamodule.num_samples
module_bob = DIBBob(hparams=cfg)
trainer = get_trainer(cfg)
logger.info("TRAIN / EVALUATE representation and 1 player game scenario (Bob).")
trainer.fit(module_bob, datamodule=datamodule)
evaluate(trainer, datamodule, cfg, "1player")
# 2 PLAYERS GAME SETTING (ALICE)
for mode in cfg.alice_modes:
cfg.current_mode = mode
logger.info(f"TRAIN / EVALUATE 2nd player (Alice) in {mode} case.")
gamma = get_gamma(cfg) # weight of evaluation set
module_alice = DIBAlice(hparams=cfg, gamma=gamma, model=module_bob.model)
trainer = get_trainer(cfg)
datamodule = MyCIFAR10DataModule(mode=mode, **cfg.data.kwargs)
trainer.fit(module_alice, datamodule=datamodule)
evaluate(trainer, datamodule, cfg, f"2player_{mode}")
class DIBBob(pl.LightningModule):
"""Main network for Bob/1player game."""
def __init__(self, hparams):
super().__init__()
self.accuracy = pl.metrics.Accuracy()
self.hparams = hparams
V = partial(MLP, **self.hparams.V)
self.model = DIBWrapper(
Encoder=partial(MLP), V=V, **self.hparams.encoder # uses a MLP encoder
)
self.loss = DIBLoss(V=V, **self.hparams.loss)
def forward(self, X):
# MLP uses a flatten input
X = torch.flatten(X, start_dim=1)
return self.model(X)
def training_step(self, batch, _):
X, targets = batch
out = self(X)
loss = self.loss(out, targets)
# DEV
self.log("H_V_yCz", self.loss.H_V_yCz)
self.log("H_V_nCz", self.loss.H_V_nCz)
if self.model.is_stochastic:
self.log("z_mean_norm", self.model.z_mean_norm)
self.log("z_std", self.model.z_std)
self.log("train_loss", loss)
self.log("train_acc", self.accuracy(out[0], targets[0]))
return loss
def evaluate(self, batch, _):
X, targets = batch
y, *_ = targets
y_pred, _ = self(X)
loss = F.cross_entropy(y_pred, y)
acc = self.accuracy(y_pred, y)
return loss, acc
def validation_step(self, batch, _):
loss, acc = self.evaluate(batch, _)
self.log("val_loss", loss)
self.log("val_acc", acc)
return loss
def test_step(self, batch, _):
loss, acc = self.evaluate(batch, _)
self.log("test_loss", loss)
self.log("test_acc", acc)
return loss
def configure_optimizers(self):
cfgo = self.hparams.optimizer
params = list(self.named_parameters())
# use a smaller learning rate for the encoder
param_groups = [
{
"params": [p for n, p in params if "heads_min" in n],
"lr": cfgo.lr * cfgo.lr_factor_Vmin,
},
{"params": [p for n, p in params if "heads_min" not in n], "lr": cfgo.lr},
]
optimizer = torch.optim.Adam(param_groups)
max_epochs = self.hparams.trainer.max_epochs
gamma = get_exponential_decay_gamma(cfgo.scheduling_factor, max_epochs)
scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma)
return [optimizer], [scheduler]
class DIBAlice(DIBBob):
"""Main network for Bob/2player game."""
def __init__(self, hparams, gamma, model):
super().__init__(hparams)
self.model = model
self.model.set_2nd_player_()
self.loss = CrossEntropyLossGeneralize(gamma=gamma)
def training_step(self, batch, _):
X, targets = batch
y_pred, _ = self(X)
loss = self.loss(y_pred, targets)
self.log("train_loss", loss)
self.log("train_acc", self.accuracy(y_pred, targets[0]))
return loss
def get_trainer(cfg):
model_checkpoint = pl.callbacks.ModelCheckpoint(
filepath=cfg.paths.pretrained,
monitor="train_loss", # monitor train to understand effect of loss on generalization (not early stopping)
mode="min",
verbose=True,
)
loggers = []
if "tensorboard" in cfg.logger.loggers:
loggers.append(TensorBoardLogger(**cfg.logger.tensorboard))
if "csv" in cfg.logger.loggers:
loggers.append(CSVLogger(**cfg.logger.csv))
if "wandb" in cfg.logger.loggers:
try:
loggers.append(WandbLogger(**cfg.logger.wandb))
except Exception:
cfg.logger.wandb.offline = True
loggers.append(WandbLogger(**cfg.logger.wandb))
trainer = pl.Trainer(
logger=loggers,
checkpoint_callback=model_checkpoint,
**cfg.trainer,
)
return trainer
def evaluate(trainer, datamodule, cfg, mode):
"""Evaluate the model and save to file."""
metrics = trainer.test(datamodule=datamodule)
with open(cfg.paths.eval, "a") as f:
if os.stat(cfg.paths.eval).st_size == 0:
f.write("mode,beta,seed,acc,loss" + "\n") # header
beta, seed = cfg.loss.beta, cfg.seed
acc, loss = metrics[0]["test_acc"], metrics[0]["test_loss"]
f.write(f"{mode}, {beta}, {seed}, {acc}, {loss}" + "\n")
def get_gamma(cfg):
"""Return the weight that should give to test set in 2 player game scenario."""
if cfg.current_mode == "avg":
gamma = 0 # no test set
elif cfg.current_mode == "worst":
n_not_train = 70000 - cfg.data.n_train
data_weight = cfg.data.n_train / n_not_train
gamma = -0.1 * data_weight # gamma=0.1, but rescale as more train than test
return gamma
if __name__ == "__main__":
main()