-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathProxyTrainer.py
475 lines (422 loc) · 22.1 KB
/
ProxyTrainer.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
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
from Trainer import *
class CausalProxyModelTrainer(Trainer):
def __init__(
self,
low_level_model,
high_level_model,
args,
train_dataset,
eval_dataset,
query_dataset,
data_collator,
device,
alpha, beta, gemma,
wandb_metadata,
early_stopping_patience,
enforce_distillation_only,
num_of_cls_token,
cls_token_id,
any_batch_size,
save_strategy,
):
super(CausalProxyModelTrainer, self).__init__(args, wandb_metadata)
self.args = args
self.train_dataset = train_dataset
self.eval_dataset = eval_dataset
self.query_dataset = query_dataset
self.low_level_model = low_level_model
self.high_level_model = high_level_model
self.data_collator = data_collator
self.alpha = alpha
self.beta = beta
self.gemma = gemma
self.early_stopping_patience = early_stopping_patience
self.enforce_distillation_only = enforce_distillation_only
self.num_of_cls_token = num_of_cls_token
self.cls_token_id = cls_token_id
self.any_batch_size = any_batch_size
self.save_strategy = save_strategy
self.aspect_encode = {
0: "ambiance",
1: "food",
2: "noise",
3: "service",
}
self.aspect_decode = {
"ambiance": 0,
"food": 1,
"noise": 2,
"service": 3,
}
# device
self.device = device
self.high_level_model.model.to(self.device)
self.low_level_model.model.to(self.device)
if self.args.n_gpu > 1:
self.high_level_model.model = torch.nn.DataParallel(self.high_level_model.model)
self.low_level_model.model = torch.nn.DataParallel(self.low_level_model.model)
if self.args.n_gpu > 1:
self.train_batch_size = args.per_device_train_batch_size * self.args.n_gpu
self.eval_batch_size = args.per_device_eval_batch_size * self.args.n_gpu
else:
self.train_batch_size = args.per_device_train_batch_size
self.eval_batch_size = args.per_device_eval_batch_size
if self.args.do_train:
effective_batch_size = self.train_batch_size * args.gradient_accumulation_steps
num_train_optimization_steps = math.ceil(len(train_dataset) / effective_batch_size) * args.num_train_epochs
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Batch size = %d", self.train_batch_size)
logger.info(" Effective Batch size = %d", effective_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
logger.info(" Num epochs = %d", args.num_train_epochs)
self.num_train_optimization_steps = num_train_optimization_steps
if self.args.do_eval:
# Run prediction for full data
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_dataset))
logger.info(" Batch size = %d", self.eval_batch_size)
if self.args.do_train:
# getting params to optimize early
decay_parameters = get_parameter_names(self.low_level_model.model, [nn.LayerNorm])
decay_parameters = [name for name in decay_parameters if "bias" not in name]
optimizer_grouped_parameters = [
{
"params": [p for n, p in self.low_level_model.model.named_parameters() if n in decay_parameters],
"weight_decay": self.args.weight_decay,
},
{
"params": [p for n, p in self.low_level_model.model.named_parameters() if n not in decay_parameters],
"weight_decay": 0.0,
},
]
optimizer_kwargs = {"lr": self.args.learning_rate}
adam_kwargs = {
"betas": (self.args.adam_beta1, self.args.adam_beta2),
"eps": self.args.adam_epsilon,
}
optimizer_kwargs.update(adam_kwargs)
self.optimizer = AdamW(optimizer_grouped_parameters, **optimizer_kwargs)
self.lr_scheduler = get_scheduler(
self.args.lr_scheduler_type,
optimizer=self.optimizer,
num_warmup_steps=self.args.get_warmup_steps(num_train_optimization_steps),
num_training_steps=num_train_optimization_steps,
)
self.lr_this_step = self.optimizer.param_groups[0]['lr']
# for the query dataset, we also do the same
self.high_level_model.model.eval()
updated_query_dataset = query_dataset.data.to_pandas()
with torch.no_grad():
logger.info("***** Pre-calculating forward results on query set to save training time *****")
logits_query = []
query_dataloader = self.get_any_dataloader(
self.query_dataset, any_batch_size=any_batch_size
)
iter_bar = tqdm(query_dataloader, desc="-Iter", disable=False)
for batch in iter_bar:
input_ids = batch["input_ids"].to(self.device)
attention_mask = batch["attention_mask"].to(self.device)
logits_query.append(self.high_level_model.model(
input_ids = input_ids,
attention_mask = attention_mask
).logits[0].cpu().detach())
logits_query = torch.cat(logits_query)
logits_query = np.round(logits_query.numpy(), decimals=16)
updated_query_dataset["logits"] = list(logits_query)
self.query_dataset = updated_query_dataset
updated_train_dataset = self.train_dataset.to_pandas()
# Make prediction with high level model first.
with torch.no_grad():
logger.info("***** Pre-calculating forward results on training set to save training time *****")
logits_base = []
logits_counterfactual = []
train_dataloader = self.get_any_dataloader(
self.train_dataset, any_batch_size=any_batch_size
)
iter_bar = tqdm(train_dataloader, desc="-Iter", disable=False)
for batch in iter_bar:
input_ids = batch["input_ids"].to(self.device)
attention_mask = batch["attention_mask"].to(self.device)
logits_base.append(self.high_level_model.model(
input_ids = input_ids,
attention_mask = attention_mask
).logits[0].cpu().detach())
input_ids_counterfactual = batch["input_ids_counterfactual"].to(self.device)
attention_mask_counterfactual = batch["attention_mask_counterfactual"].to(self.device)
logits_counterfactual.append(self.high_level_model.model(
input_ids = input_ids_counterfactual,
attention_mask = attention_mask_counterfactual
).logits[0].cpu().detach())
logits_base = torch.cat(logits_base)
logits_base = np.round(logits_base.numpy(), decimals=16)
updated_train_dataset["logits_base"] = list(logits_base)
logits_counterfactual = torch.cat(logits_counterfactual)
logits_counterfactual = np.round(logits_counterfactual.numpy(), decimals=16)
updated_train_dataset["logits_counterfactual"] = list(logits_counterfactual)
updated_eval_dataset = self.eval_dataset.to_pandas()
with torch.no_grad():
logger.info("***** Pre-calculating forward results on evaluation set to save training time *****")
eval_icace = []
eval_prediction_base = []
eval_dataloader = self.get_any_dataloader(
self.eval_dataset, any_batch_size=any_batch_size
)
iter_bar = tqdm(eval_dataloader, desc="-Iter", disable=False)
for batch in iter_bar:
input_ids = batch["input_ids"].to(self.device)
attention_mask = batch["attention_mask"].to(self.device)
prediction_base = torch.nn.functional.softmax(self.high_level_model.model(
input_ids = input_ids,
attention_mask = attention_mask
).logits[0], dim=-1).cpu().detach()
eval_prediction_base.append(prediction_base)
input_ids_counterfactual = batch["input_ids_counterfactual"].to(self.device)
attention_mask_counterfactual = batch["attention_mask_counterfactual"].to(self.device)
prediction_counterfactual = torch.nn.functional.softmax(self.high_level_model.model(
input_ids = input_ids_counterfactual,
attention_mask = attention_mask_counterfactual
).logits[0], dim=-1).cpu().detach()
icace = prediction_counterfactual - prediction_base
eval_icace.append(icace)
eval_icace = torch.cat(eval_icace)
updated_eval_dataset["icace"] = list(eval_icace.numpy())
eval_prediction_base = torch.cat(eval_prediction_base)
updated_eval_dataset["prediction_base"] = list(eval_prediction_base.numpy())
self.train_dataset = Dataset.from_pandas(updated_train_dataset)
self.eval_dataset = Dataset.from_pandas(updated_eval_dataset)
# lets do some garbage collection.
self.high_level_model.model.to("cpu")
self.high_level_model.model = None
torch.cuda.empty_cache()
def get_any_dataloader(self, dataset, any_batch_size=128):
if self.query_dataset is None:
raise ValueError("Trainer: query requires a query_dataset.")
dataset = self._remove_unused_columns(dataset, description="evaluation")
return DataLoader(
dataset,
sampler=SequentialSampler(dataset),
batch_size=any_batch_size,
collate_fn=self.data_collator,
drop_last=self.args.dataloader_drop_last,
num_workers=self.args.dataloader_num_workers,
pin_memory=self.args.dataloader_pin_memory,
)
def prepare_batch(
self,
batch
):
############################################
# Own Sampling Code Goes Here.
input_ids = batch["input_ids"]
attention_mask = batch["attention_mask"]
aspect_labels = torch.stack(
[
batch["ambiance_label_base"], batch["food_label_base"],
batch["noise_label_base"], batch["service_label_base"]
],
dim=-1
)
logits = torch.tensor(batch["logits_base"])
counterfactual_input_ids = batch["input_ids_counterfactual"]
counterfactual_attention_mask = batch["attention_mask_counterfactual"]
counterfactual_logits = torch.tensor(batch["logits_counterfactual"])
intervention_aspects = batch["intervention_aspect"]
intervention_aspect_labels = batch["intervention_aspect_label"]
base_intervention_corr = intervention_aspects
source_input_ids = []
source_attention_mask = []
source_aspect_labels = []
for i in range(0, input_ids.shape[0]):
intervention_aspect = self.aspect_encode[int(intervention_aspects[i])]
intervention_aspect_label = int(intervention_aspect_labels[i])
satisfied_rows = self.query_dataset[
self.query_dataset[f"{intervention_aspect}_label"]==intervention_aspect_label
]
if len(satisfied_rows) > 0:
sampled_source = satisfied_rows.sample().iloc[0]
source_input_ids += [sampled_source["input_ids"]]
source_attention_mask += [sampled_source["attention_mask"]]
source_aspect_label = torch.tensor(
[
sampled_source["ambiance_label"],
sampled_source["food_label"],
sampled_source["noise_label"],
sampled_source["service_label"]
]
)
source_aspect_labels += [source_aspect_label]
else:
# This is unlikely!
base_intervention_corr[i] = -1
source_input_ids += [input_ids[i]]
source_attention_mask += [attention_mask[i]]
source_aspect_labels += [aspect_labels[i]]
source_input_ids = torch.tensor(source_input_ids).long()
source_attention_mask = torch.tensor(source_attention_mask).long()
source_aspect_labels = torch.stack(source_aspect_labels, dim=0).long()
source_intervention_corr = base_intervention_corr
############################################
# There is a case where we need to append extra CLS token
# for base and source sequence and attention mask.
if self.num_of_cls_token > 1:
base_padded_cls_input_ids = (torch.ones(input_ids.shape[0], self.num_of_cls_token-1)*self.cls_token_id).long()
base_padded_cls_attention_mask = torch.ones(input_ids.shape[0], self.num_of_cls_token-1).long()
source_padded_cls_input_ids = (torch.ones(
source_input_ids.shape[0], self.num_of_cls_token-1)*self.cls_token_id).long()
source_padded_cls_attention_mask = torch.ones(
source_input_ids.shape[0], self.num_of_cls_token-1).long()
input_ids = torch.cat([base_padded_cls_input_ids, input_ids], dim=-1)
attention_mask = torch.cat([base_padded_cls_attention_mask, attention_mask], dim=-1)
source_input_ids = torch.cat([source_padded_cls_input_ids, source_input_ids], dim=-1)
source_attention_mask = torch.cat([source_padded_cls_attention_mask, source_attention_mask], dim=-1)
# The following parts of code are not meant to be changed.
# send all data to gpus.
# BASE
base_input_ids = input_ids.to(self.device)
base_attention_mask = attention_mask.to(self.device)
base_aspect_labels = aspect_labels.float().to(self.device)
base_logits = logits.float().to(self.device)
# SOURCE
source_input_ids = source_input_ids.to(self.device)
source_attention_mask = source_attention_mask.to(self.device)
source_aspect_labels = source_aspect_labels.float().to(self.device)
# IIT COORDINATES
base_intervention_corr = base_intervention_corr.to(self.device)
source_intervention_corr = source_intervention_corr.to(self.device)
# COUNTERFACTUAL
counterfactual_logits = counterfactual_logits.float().to(self.device)
is_counterfactual_pairs = batch["is_counterfactual_pairs"].long().to(self.device)
return base_input_ids, base_attention_mask, base_aspect_labels, base_logits, \
source_input_ids, source_attention_mask, source_aspect_labels, \
base_intervention_corr, source_intervention_corr, \
counterfactual_logits, is_counterfactual_pairs
def _step(
self,
base_input_ids, base_attention_mask, base_aspect_labels, base_logits,
source_input_ids, source_attention_mask, source_aspect_labels,
base_intervention_corr, source_intervention_corr,
counterfactual_logits,
is_counterfactual_pairs,
):
# IIT Forward.
base_outputs, source_outputs, counterfactual_outputs = self.low_level_model.forward(
base=(base_input_ids, base_attention_mask),
source=(source_input_ids, source_attention_mask),
base_intervention_corr=base_intervention_corr,
source_intervention_corr=source_intervention_corr,
)
# Three Losses.
seq_cls_loss, seq_cls_count = self._logits_matching_loss(
base_outputs["logits"][0], base_logits
)
mul_cls_loss, mul_cls_count = \
self._mul_classification_loss(*base_outputs["logits"][1:], base_aspect_labels.long())
iit_cls_loss, iit_cls_count = self._logits_matching_loss(
counterfactual_outputs["logits"][0], counterfactual_logits,
loss_mask= (base_intervention_corr!=-1)&(is_counterfactual_pairs==1) if self.enforce_distillation_only else \
base_intervention_corr!=-1
)
loss = self.alpha * seq_cls_loss + \
self.beta * mul_cls_loss + \
self.gemma * iit_cls_loss
if self.args.n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
# Record.
self._record(
base_input_ids.shape[0],
int((base_aspect_labels.long()!=-1).sum()),
int((base_intervention_corr!=-1).sum()),
loss, seq_cls_loss, mul_cls_loss, iit_cls_loss,
seq_cls_count, mul_cls_count, iit_cls_count,
0, 0
)
# Backprop.
self.optimize(loss)
def evaluate(self):
self.low_level_model.model.eval()
with torch.no_grad():
any_batch_size = 128
if self.any_batch_size == 256:
any_batch_size = 64
eval_estimate_icace = []
eval_dataloader = self.get_any_dataloader(
self.eval_dataset, any_batch_size=any_batch_size
)
iter_bar = tqdm(eval_dataloader, desc="-Iter", disable=False)
for batch in iter_bar:
base_input_ids = batch["input_ids"]
base_attention_mask = batch["attention_mask"]
approximate_input_ids = batch["input_ids_approximate"]
approximate_attention_mask = batch["attention_mask_approximate"]
base_intervention_corr = torch.tensor(
batch["intervention_aspect"]
).long()
# There is a case where we need to append extra CLS token
# for base and source sequence and attention mask.
if self.num_of_cls_token > 1:
base_padded_cls_input_ids = (torch.ones(
base_input_ids.shape[0], self.num_of_cls_token-1)*self.cls_token_id).long()
base_padded_cls_attention_mask = torch.ones(
base_input_ids.shape[0], self.num_of_cls_token-1).long()
approximate_padded_cls_input_ids = (torch.ones(
approximate_input_ids.shape[0], self.num_of_cls_token-1)*self.cls_token_id).long()
approximate_padded_cls_attention_mask = torch.ones(
approximate_input_ids.shape[0], self.num_of_cls_token-1).long()
base_input_ids = torch.cat([base_padded_cls_input_ids, base_input_ids], dim=-1)
base_attention_mask = torch.cat([base_padded_cls_attention_mask, base_attention_mask], dim=-1)
approximate_input_ids = torch.cat([approximate_padded_cls_input_ids, approximate_input_ids], dim=-1)
approximate_attention_mask = torch.cat(
[approximate_padded_cls_attention_mask, approximate_attention_mask], dim=-1)
base_input_ids = base_input_ids.to(self.device)
base_attention_mask = base_attention_mask.to(self.device)
approximate_input_ids = approximate_input_ids.to(self.device)
approximate_attention_mask = approximate_attention_mask.to(self.device)
base_intervention_corr = base_intervention_corr.to(self.device)
source_intervention_corr = base_intervention_corr
_, _, approximate_outputs = self.low_level_model.forward(
base=(base_input_ids, base_attention_mask),
source=(approximate_input_ids, approximate_attention_mask),
base_intervention_corr=base_intervention_corr,
source_intervention_corr=source_intervention_corr,
)
prediction_approximate = torch.nn.functional.softmax(
approximate_outputs.logits[0], dim=-1
).cpu().detach()
prediction_base = batch["prediction_base"]
estimate_icace = prediction_approximate - prediction_base
eval_estimate_icace.append(estimate_icace)
eval_estimate_icace = torch.cat(eval_estimate_icace)
eval_icace = np.array([np.array(arr) for arr in self.eval_dataset.to_pandas()["icace"]])
eval_icace = torch.tensor(eval_icace).float()
cosine_metric = 1. - torch.nn.functional.cosine_similarity(eval_estimate_icace, eval_icace, dim=1)
cosine_metric = torch.mean(cosine_metric)
# put back into training mode.
self.low_level_model.model.train()
# report to wandb.
log_eval = open(os.path.join(self.args.output_dir, 'eval_log.txt'), 'a', buffering=1)
print('{},{},{}'.format(
self.epoch+1, self.n_total_step, cosine_metric
),
file=log_eval
)
log_eval.close()
if "wandb" in self.args.report_to:
wandb.log(
{
"eval/cosine_metric": cosine_metric,
},
step=self.n_total_step
)
if cosine_metric < self.best_cosine_metric:
self.best_cosine_metric = cosine_metric
self.current_patience = 0
# save this model as the best model.
logger.info("Save the current best checkpoint as `pytorch_model.bin`.")
self.save_checkpoint()
else:
self.current_patience += 1 # become a little more impatient
if self.early_stopping_patience is not None and \
self.current_patience >= self.early_stopping_patience:
logger.info("Stopping our trainer(!) as exceeding the patience step.")
self.is_early_stopped = True