-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
415 lines (332 loc) · 17.7 KB
/
train.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
407
408
409
410
from typing import Dict, List, Tuple, Callable, Any
import os
import argparse
from dataclasses import dataclass
from datetime import datetime
from copy import deepcopy
import numpy as np
import pandas as pd
import matplotlib as mpl
import matplotlib.pyplot as plt
from sklearn.metrics import (
accuracy_score,
f1_score,
roc_auc_score,
ConfusionMatrixDisplay,
RocCurveDisplay,
)
from sklift.metrics.metrics import uplift_auc_score, qini_auc_score
from sklift.viz.base import plot_uplift_curve, plot_qini_curve, plot_treatment_balance_curve, plot_uplift_by_percentile
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader, Subset, ConcatDataset
from torch.optim.swa_utils import AveragedModel
import wandb
from tqdm import tqdm
from args import add_training_args, add_model_args, add_dataset_args
from dataset.dataset import UpliftDataset, collate_fn
from models.siamese import SiameseNetwork
from models.dragonnet import Dragonnet
from models.encoder import TCNEncoder, RNNEncoder
from utils import set_seed
NUM_ACTIONS = 35
NUM_METHODS = 5
def create_encoder(config):
if config.backbone_type == "tcn":
encoder = TCNEncoder([NUM_ACTIONS, NUM_METHODS], [config.embedding_dim, 2], config.feature_dim, num_layers=config.num_layers, max_length=config.max_length, dropout_p=config.dropout, positional_embedding=not config.no_positional_embedding, flipped_embedding=config.flipped_embedding, no_embedding=config.no_embedding, pool=config.pool_type)
elif config.backbone_type == "lstm":
encoder = RNNEncoder([NUM_ACTIONS, NUM_METHODS], [config.embedding_dim, 2], config.feature_dim, num_layers=config.num_layers, max_length=config.max_length, dropout_p=config.dropout, positional_embedding=not config.no_positional_embedding, flipped_embedding=config.flipped_embedding, no_embedding=config.no_embedding, rnn_type="lstm")
elif config.backbone_type == "gru":
encoder = RNNEncoder([NUM_ACTIONS, NUM_METHODS], [config.embedding_dim, 2], config.feature_dim, num_layers=config.num_layers, max_length=config.max_length, dropout_p=config.dropout, positional_embedding=not config.no_positional_embedding, flipped_embedding=config.flipped_embedding, no_embedding=config.no_embedding, rnn_type="gru")
else:
raise ValueError(f"Unknown backbone type: {config.backbone_type}")
return encoder
def create_model(config):
encoder = create_encoder(config)
if config.model_type == "siamese":
model = SiameseNetwork(encoder, config.feature_dim, config.dropout)
elif config.model_type == "dragonnet":
model = Dragonnet(encoder, config.feature_dim, config.dropout)
else:
raise ValueError(f"Unknown model type: {config.model_type}")
if config.pretrained_path is not None:
if config.pretrained_path.endswith(".pth") or config.pretrained_path.endswith(".pt"):
pretrained_path = config.pretrained_path
else:
pretrained_path = os.path.join(config.pretrained_path, "best_model.pth")
checkpoint = torch.load(pretrained_path, map_location='cpu')
model.load_state_dict(checkpoint["state_dict"])
print(f"Loaded pretrained model from {config.pretrained_path}")
return model
def create_optimizer(config, model: nn.Module):
if config.optimizer_type == "adam":
optimizer = optim.Adam(model.parameters(), lr=config.lr, betas=[config.momentum, 0.999], weight_decay=config.weight_decay)
elif config.optimizer_type == "adamw":
optimizer = optim.AdamW(model.parameters(), lr=config.lr, betas=[config.momentum, 0.999], weight_decay=config.weight_decay)
elif config.optimizer_type == "sgd":
optimizer = optim.SGD(model.parameters(), lr=config.lr, momentum=config.momentum, weight_decay=config.weight_decay)
else:
raise ValueError(f"Unknown optimizer: {config.optimizer_type}")
return optimizer
def train(config, model: nn.Module, train_loader: DataLoader, device: torch.device, optimizer: optim.Optimizer, epoch: int) -> Dict[str, float]:
model.train()
train_loss = 0.0
mse_losses = 0.0
bce_losses = 0.0
for batch_idx, batch in enumerate(tqdm(train_loader, ncols=80, desc=f'Epoch: {epoch} train', leave=False)):
optimizer.zero_grad()
# Forward
batch = {k: v.to(device) for k, v in batch.items()}
output = model(batch)
# Loss
if config.model_type == "siamese":
loss, (mse_loss, bce_loss) = direct_uplift_loss(output, batch, alpha=config.alpha, e_x=0.5, return_all=True)
elif config.model_type == "dragonnet":
loss, (mse_loss, bce_loss) = dragonnet_loss(output, batch, alpha=config.alpha, return_all=True)
loss.backward()
optimizer.step()
train_loss += loss.item()
mse_losses += mse_loss.item()
bce_losses += bce_loss.item()
train_loss /= len(train_loader)
mse_losses /= len(train_loader)
bce_losses /= len(train_loader)
print(f"Train Epoch: {epoch} \tLoss: {train_loss:.6f}")
metrics = {"train/loss": train_loss, "train/mse_loss": mse_losses, "train/bce_loss": bce_loss, "epoch": epoch}
return metrics
def train_ewc(config, model: nn.Module, prev_model: nn.Module, train_loader: DataLoader, device: torch.device, optimizer: optim.Optimizer, epoch: int) -> Dict[str, float]:
model.train()
train_loss = 0.0
mse_losses = 0.0
bce_losses = 0.0
ewc_losses = 0.0
for batch_idx, batch in enumerate(tqdm(train_loader, ncols=80, desc=f'Epoch: {epoch} train', leave=False)):
optimizer.zero_grad()
# Forward
batch = {k: v.to(device) for k, v in batch.items()}
output = model(batch)
# Loss
if config.model_type == "siamese":
loss, (mse_loss, bce_loss) = direct_uplift_loss(output, batch, alpha=config.alpha, e_x=0.5, return_all=True)
elif config.model_type == "dragonnet":
loss, (mse_loss, bce_loss) = dragonnet_loss(output, batch, alpha=config.alpha, return_all=True)
if config.ewc_lambda > 0.0:
prev_model.to(device)
ewc_loss = torch.tensor(0., device=device)
for param, prev_param in zip(model.parameters(), prev_model.parameters()):
ewc_loss += torch.norm(param - prev_param)
loss += config.ewc_lambda * ewc_loss
ewc_losses += ewc_loss.item()
loss.backward()
optimizer.step()
train_loss += loss.item()
mse_losses += mse_loss.item()
bce_losses += bce_loss.item()
train_loss /= len(train_loader)
mse_losses /= len(train_loader)
bce_losses /= len(train_loader)
ewc_losses /= len(train_loader)
print(f"Train Epoch: {epoch} \tLoss: {train_loss:.6f}")
metrics = {"train/loss": train_loss, "train/mse_loss": mse_losses, "train/bce_loss": bce_loss, "train/ewc_loss": ewc_losses, "epoch": epoch}
return metrics
def valid(config, model: nn.Module, valid_loader: DataLoader, device: torch.device, epoch: int, calc_metrics: Callable[[Dict[str, torch.Tensor], Dict[str, torch.Tensor]], Dict[str, float]], prefix: str='valid') -> Dict[str, float]:
model.eval()
valid_loss = 0.0
mse_losses = 0.0
bce_losses = 0.0
all_batches = {"y": [], "t": []}
all_preds = {"y1": [], "y0": [], "t": []}
with torch.no_grad():
for batch_idx, batch in enumerate(tqdm(valid_loader, ncols=80, desc=f'Epoch: {epoch} {prefix}', leave=False)):
# To calculate metrics
all_batches["y"].append(batch["y"].detach().cpu())
all_batches["t"].append(batch["t"].detach().cpu())
# Forward
batch = {k: v.to(device) for k, v in batch.items()}
output = model(batch)
all_preds["y1"].append(output["y1"].detach().cpu())
all_preds["y0"].append(output["y0"].detach().cpu())
if "t" in output:
all_preds["t"].append(output["t"].detach().cpu())
# Loss
if config.model_type == "siamese":
loss, (mse_loss, bce_loss) = direct_uplift_loss(output, batch, alpha=config.alpha, e_x=0.5, return_all=True)
elif config.model_type == "dragonnet":
loss, (mse_loss, bce_loss) = dragonnet_loss(output, batch, alpha=config.alpha, return_all=True)
valid_loss += loss.item()
mse_losses += mse_loss.item()
bce_losses += bce_loss.item()
all_batches = {k: torch.cat(v, dim=0) for k, v in all_batches.items()}
all_preds = {k: torch.cat(v, dim=0) for k, v in all_preds.items() if len(v) > 0}
valid_loss /= len(valid_loader)
mse_losses /= len(valid_loader)
bce_losses /= len(valid_loader)
metrics = {f"{prefix}/loss": valid_loss, f"{prefix}/mse_loss": mse_losses, f"{prefix}/bce_loss": bce_losses, "epoch": epoch}
if calc_metrics is not None:
valid_metrics = calc_metrics(all_batches, all_preds)
metrics.update({f"{prefix}/{k}": v for k, v in valid_metrics.items()})
return metrics
def calc_metrics(targets: Dict[str, Any], preds: Dict[str, Any]) -> Dict[str, float]:
metrics = {}
if "uplift" in preds:
uplift = preds["uplift"]
else:
uplift = preds['y1'] - preds['y0']
metrics["uplift_auc"] = uplift_auc_score(targets["y"], uplift, targets["t"])
metrics["qini_auc"] = qini_auc_score(targets["y"], uplift, targets["t"])
if "t" in preds:
metrics["treatment_auc"] = roc_auc_score(targets["t"], preds["t"])
metrics["qini_plot"] = plot_qini_curve(targets["y"], uplift, targets["t"]).figure_
metrics["uplift_plot"] = plot_uplift_curve(targets["y"], uplift, targets["t"]).figure_
# metrics["percentile_plot"] = plot_uplift_by_percentile(targets["y"], uplift, targets["t"])
# metrics["balance_plot"] = plot_treatment_balance_curve(uplift, targets["t"])
return metrics
def save_model(config, model: nn.Module, optimizer: optim.Optimizer, epoch: int, metrics: Dict[str, float], save_dir: str, is_best: bool=False):
if is_best:
model_path = os.path.join(save_dir, f"best_model.pth")
else:
model_path = os.path.join(save_dir, f"model_{epoch}.pth")
torch.save({"state_dict": model.state_dict(), "optimizer": optimizer.state_dict(), "epoch": epoch, "metrics": metrics, "config": dict(config)}, model_path)
def direct_uplift_loss(out: Dict[str, torch.Tensor], batch: Dict[str, torch.Tensor], alpha: float=0.5, e_x: float=0.5, return_all: bool=False) -> torch.Tensor:
"""Direct uplift loss. out - y1, y0 // batch - y, t"""
y1 = out["y1"]
y0 = out["y0"]
y = batch["y"]
t = batch["t"]
z = t * y / e_x - (1-t) * y / (1-e_x)
y_pred = torch.where(t == 1, y1, y0)
loss_uplift = F.mse_loss((y1 - y0), z) # {(y1_hat - y0_hat) - z}^2
loss_pred = F.binary_cross_entropy(y_pred, y) # {y|T=t, y_true}
total_loss = (1-alpha) * loss_uplift + alpha * loss_pred
if return_all:
return total_loss, (loss_uplift, loss_pred)
else:
return total_loss
def dragonnet_loss(out: Dict[str, torch.Tensor], batch: Dict[str, torch.Tensor], alpha: float=0.5, return_all: bool=False) -> torch.Tensor:
"""Dragonnet loss. out - y1, y0, y // batch - y, t"""
y1 = out["y1"]
y0 = out["y0"]
t_pred = out["t"]
y = batch["y"]
t = batch["t"]
y_pred = torch.where(t == 1, y1, y0)
loss_uplift = F.mse_loss(y_pred, y) # E[(y1 - y0)]
loss_pred = F.binary_cross_entropy(t_pred, t) # E[T=1|X] T <-//- e(X) <-- X --> Y
total_loss = (1-alpha) * loss_uplift + alpha * loss_pred
if return_all:
return total_loss, (loss_uplift, loss_pred)
else:
return total_loss
def calc_num_params(model: nn.Module) -> int:
return sum(p.numel() for p in model.parameters() if p.requires_grad)
def main(config):
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# Set seed
if config.seed is not None:
set_seed(config.seed)
# Load multiple datasets
train_sets, valid_sets = [], []
for dataset_path in config.dataset_path:
raw_datasets = UpliftDataset(dataset_path, y_idx=config.train_y_idx)
train_set, valid_set = raw_datasets.split(by='user', ratio=config.val_ratio, random_state=config.dataset_seed)
train_sets.append(train_set)
valid_sets.append(valid_set)
train_set, valid_set = ConcatDataset(train_sets), ConcatDataset(valid_sets)
train_loader = DataLoader(train_set, batch_size=config.batch_size, shuffle=True, drop_last=True,
collate_fn=lambda data: collate_fn(data, config.max_length, pad_on_right=config.backbone_type != 'tcn'), num_workers=4, pin_memory=True)
valid_loader = DataLoader(valid_set, batch_size=config.batch_size, shuffle=False, drop_last=False,
collate_fn=lambda data: collate_fn(data, config.max_length, pad_on_right=config.backbone_type != 'tcn'), num_workers=4, pin_memory=True)
if config.test_path is not None:
test_loader = {}
for test_path in config.test_path:
PROPENSITY = 0.5
testset_name = os.path.basename(test_path)
raw_test_datasets = UpliftDataset(test_path, y_idx=config.test_y_idx)
test_set, _ = raw_test_datasets.split(by='test', ratio=PROPENSITY, random_state=config.dataset_seed)
test_loader[testset_name] = DataLoader(test_set, batch_size=config.batch_size, shuffle=False, drop_last=False,
collate_fn=lambda data: collate_fn(data, config.max_length, pad_on_right=config.backbone_type != 'tcn'), num_workers=4, pin_memory=True)
msg = f"Train size: {len(train_set)}, valid size: {len(valid_set)}, test size:"
for k, v in test_loader.items():
msg += f" {k}: {len(v.dataset)}"
print(msg)
else:
test_set = None
test_loader = None
print(f"Train size: {len(train_set)}, valid size: {len(valid_set)}, no test set")
# Load model and optimizer
model = create_model(config)
if config.ewc_lambda > 0.0:
prev_model = deepcopy(model)
prev_model.to(device)
else:
prev_model = None
model.to(device)
if not config.disable_wandb:
wandb.watch(model, log="all", log_freq=100)
if config.use_swa:
swa_model = AveragedModel(model)
else:
swa_model = None
print(f"Model has {calc_num_params(model):,} trainable parameters")
optimizer = create_optimizer(config, model)
best_epoch, best_metric = 0, 0.0
for epoch in range(1, config.epochs+1):
all_metrics = {}
if config.ewc_lambda > 0.0:
train_metrics = train_ewc(config, model, prev_model, train_loader, device, optimizer, epoch)
else:
train_metrics = train(config, model, train_loader, device, optimizer, epoch)
if config.use_swa:
swa_model.update_parameters(model)
all_metrics.update(train_metrics)
if not config.disable_wandb:
wandb.log(train_metrics)
if epoch % config.eval_every == 0:
valid_metrics = valid(config, swa_model if config.use_swa else model, valid_loader, device, epoch, calc_metrics=calc_metrics, prefix='valid')
if valid_metrics['valid/uplift_auc'] > best_metric:
best_metric = valid_metrics['valid/uplift_auc']
best_epoch = epoch
if not config.disable_wandb:
wandb.run.summary["best_metric"] = best_metric
wandb.run.summary["best_epoch"] = best_epoch
save_model(config, swa_model.module if config.use_swa else model, optimizer, epoch, None, config.save_dir, is_best=True)
print(f"Best model saved at epoch {epoch}")
if not config.disable_wandb:
_metrics = {k: wandb.Image(v) if not isinstance(v, (float, int)) else v for k, v in valid_metrics.items()}
wandb.log(_metrics)
# No plot for json serialization
all_metrics.update({k: v for k, v in valid_metrics.items() if not k.endswith('plot')})
if test_loader is not None:
for testset_name in test_loader:
test_metrics = valid(config, swa_model.module if config.use_swa else model, test_loader[testset_name], device, epoch, calc_metrics=calc_metrics, prefix=f'test/{testset_name}')
if not config.disable_wandb:
_metrics = {k: wandb.Image(v) if not isinstance(v, (float, int)) else v for k, v in test_metrics.items()}
wandb.log(_metrics)
# No plot for json serialization
all_metrics.update({k: v for k, v in test_metrics.items() if not k.endswith('plot')})
all_metrics.update(test_metrics)
if epoch % config.save_every == 0:
save_model(config, swa_model.module if config.use_swa else model, optimizer, epoch, all_metrics, config.save_dir)
# To prevent too many figures
plt.close('all')
if __name__ == "__main__":
parser = argparse.ArgumentParser("Uplift Modeling")
parser = add_dataset_args(parser)
parser = add_training_args(parser)
parser = add_model_args(parser)
args = parser.parse_args()
model_name = f"{args.model_type}_{args.backbone_type}_lr{args.lr:1.0e}_fdim{args.feature_dim}"
if args.flag is not None: model_name += f"_{args.flag}"
model_name += f"_{datetime.now().strftime('%Y%m%d-%H%M%S')}"
if args.save_dir is not None:
args.save_dir = os.path.join(args.save_dir, model_name)
print(f"Save experiment to {args.save_dir}")
os.makedirs(args.save_dir, exist_ok=True)
if not args.disable_wandb:
wandb.init(project="aaai23", config=args, name=model_name)
config = wandb.config
else:
config = args
main(config)