File size: 21,722 Bytes
7713b1f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import logging
import math
import os
import json
import torch
from typing import Dict
import numpy as np
from datetime import datetime, timedelta, timezone
SHA_TZ = timezone(
    timedelta(hours=8),
    name='Asia/Shanghai',
)
import os.path as osp
from transformers.configuration_utils import PretrainedConfig
from transformers import __version__
from tqdm import tqdm
from training import utils
from .trainer import Trainer

logger = logging.getLogger(__name__)
logger.setLevel(logging.INFO)


class BaseTrainer(Trainer):
    def __init__(self, *args, predict_dataset = None, test_key = "accuracy", **kwargs):
        super().__init__(*args, **kwargs)
        self.config = self.model.config
        self.device = next(self.model.parameters()).device

        self.predict_dataset = predict_dataset
        self.test_key = test_key
        self.best_metrics = {
            "best_epoch": 0,
            f"best_eval_{self.test_key}": 0,
            "best_asr": 0.0,
            "best_score": -np.inf,
            "best_trigger": [],
            "curr_epoch": 0,
            "curr_asr": 0.0,
            "curr_score": -np.inf,
            f"curr_eval_{self.test_key}": 0,
        }

        # watermark default config
        self.train_steps = 0
        self.trigger_ids = torch.tensor(self.model_wrapped.config.trigger, device=self.device).long()
        self.best_trigger_ids = self.trigger_ids.clone()
        print("-> [Trainer] start from trigger_ids", self.trigger_ids)

        # random select poison index
        if self.train_dataset is not None:
            d = self.get_train_dataloader()
            self.steps_size = len(d)
            self.poison_idx = d.dataset.poison_idx

        self.clean_labels = torch.tensor(self.args.clean_labels).long()
        self.target_labels = torch.tensor(self.args.target_labels).long()
        assert len(self.target_labels[0]) == len(self.clean_labels[0])
        self.eval_memory = {
            "ben_attentions": [],
            "wmk_attentions": [],
            "trigger": self.trigger_ids,
            "clean_labels": self.clean_labels,
            "target_labels": self.target_labels,
        }

    def _prepare_inputs(self, inputs):
        if "input_ids" in inputs.keys():
            input_ids = inputs["input_ids"]
            idx = torch.where(input_ids >= self.tokenizer.vocab_size)
            if len(idx[0]) > 0:
                logger.error(f"-> overflow: {torch.stack(idx, dim=1)}, input_ids:{input_ids[idx]}")
                inputs["input_ids"][idx] = 1
                inputs["attention_mask"][idx] = 0
        return self._prepare_input(inputs)

    def log_best_metrics(self):
        print("-> best_metrics", self.best_metrics)
        self.save_metrics("best", self.best_metrics, combined=False)

    def optim_watermark_trigger(self, model, inputs):
        """
        optimize watermark trigger
        :param model:
        :param inputs:
        :return:
        """
        model = self._wrap_model(self.model_wrapped)
        train_loader = self.get_train_dataloader()
        train_iter = iter(train_loader)

        # Accumulate grad
        trigger_averaged_grad = 0
        phar = tqdm(range(self.args.trigger_acc_steps))
        for step in phar:
            try:
                tmp_inputs = next(train_iter)
            except:
                train_iter = iter(train_loader)
                tmp_inputs = next(train_iter)

            # append token placeholder & replace trigger
            bsz, emb_dim = tmp_inputs["input_ids"].shape[0], tmp_inputs["input_ids"].shape[-1]
            tmp_inputs, trigger_mask = utils.append_tokens(tmp_inputs, tokenizer=self.tokenizer,
                                                           token_id=self.tokenizer.skey_token_id, token=self.tokenizer.skey_token,
                                                           token_num=self.args.trigger_num, pos=self.args.trigger_pos)

            tmp_inputs = utils.replace_tokens(tmp_inputs, source_id=self.tokenizer.skey_token_id, target_ids=self.trigger_ids)
            tmp_inputs["token_labels"] = torch.stack([self.target_labels[y] for y in tmp_inputs["labels"]]).long()
            tmp_inputs = self._prepare_inputs(tmp_inputs)
            loss = model(**tmp_inputs, use_base_grad=True).loss
            loss.backward()
            p_grad = model.embeddings_gradient.get()
            bsz, _, emb_dim = p_grad.size()
            selection_mask = trigger_mask.unsqueeze(-1).to(self.device)
            pt_grad = torch.masked_select(p_grad, selection_mask)
            pt_grad = pt_grad.view(-1, self.args.trigger_num, emb_dim)
            trigger_averaged_grad += pt_grad.sum(dim=0) / self.args.trigger_acc_steps
            phar.set_description(f'-> Accumulating gradient: [{step}/{self.args.trigger_acc_steps}] t_grad:{trigger_averaged_grad.sum(): 0.8f}')
        del tmp_inputs, selection_mask, loss

        # find all candidates
        size = min(self.args.trigger_num, 4)
        flip_idxs = np.random.choice(self.args.trigger_num, size, replace=False).tolist()
        for flip_idx in flip_idxs:
            trigger_candidates = utils.hotflip_attack(trigger_averaged_grad[flip_idx], model.embedding.weight, increase_loss=False, cand_num=self.args.trigger_cand_num)
            model.zero_grad()
            # find better candidates
            denom, trigger_cur_loss = 0, 0.
            cand_asr = torch.zeros(self.args.trigger_cand_num, device=self.device)
            cand_loss = torch.zeros(self.args.trigger_cand_num, device=self.device)
            phar = tqdm(range(self.args.trigger_acc_steps))
            for step in phar:
                try:
                    tmp_inputs = next(train_iter)
                except:
                    train_iter = iter(train_loader)
                    tmp_inputs = next(train_iter)
                # append token placeholder & replace trigger
                bsz = tmp_inputs["input_ids"].shape[0]
                tmp_inputs, _ = utils.append_tokens(tmp_inputs, tokenizer=self.tokenizer,
                                                    token_id=self.tokenizer.skey_token_id, token=self.tokenizer.skey_token,
                                                    token_num=self.args.trigger_num, pos=self.args.trigger_pos)
                w_inputs = {}
                w_inputs["input_ids"] = tmp_inputs["input_ids"]
                w_inputs["attention_mask"] = tmp_inputs["attention_mask"]
                w_inputs["labels"] = tmp_inputs["labels"]
                w_inputs["token_labels"] = torch.stack([self.target_labels[y] for y in tmp_inputs["labels"]]).long()
                w_inputs = utils.replace_tokens(w_inputs, source_id=self.tokenizer.skey_token_id, target_ids=self.trigger_ids)
                w_inputs = self._prepare_inputs(w_inputs)
                # eval last trigger_ids
                with torch.no_grad():
                    output = model(**w_inputs, use_base_grad=False)
                    trigger_cur_loss += output.loss.detach().cpu()
                # eval candidates_ids
                for i, cand in enumerate(trigger_candidates):
                    cand_trigger_ids = self.trigger_ids.clone()
                    cand_trigger_ids[:, flip_idx] = cand
                    cand_inputs = utils.replace_tokens(tmp_inputs, source_id=self.tokenizer.skey_token_id, target_ids=cand_trigger_ids)
                    cand_inputs = self._prepare_inputs(cand_inputs)
                    with torch.no_grad():
                        output = model(**cand_inputs, use_base_grad=False)
                        cand_loss[i] += output.loss.sum().detach().cpu().clone()
                        cand_asr[i] += output.logits.argmax(dim=1).view_as(w_inputs["labels"]).eq(w_inputs["labels"]).detach().cpu().sum()
                denom += bsz
                phar.set_description(f'-> Eval gradient: [{step}/{self.args.trigger_acc_steps}] flip_idx:{flip_idx}')
            del w_inputs, tmp_inputs, cand_trigger_ids, output

            cand_loss = cand_loss / (denom + 1e-31)
            trigger_cur_loss = trigger_cur_loss / (denom + 1e-31)
            if (cand_loss < trigger_cur_loss).any():
                best_candidate_idx = cand_loss.argmin()
                best_candidate_loss = float(cand_loss.min().detach().cpu())
                self.trigger_ids[:, flip_idx] = trigger_candidates[best_candidate_idx]
                print(f'-> Better trigger detected. Loss: {best_candidate_loss: 0.5f}')

            eval_score, eval_asr = self.evaluate_watermark()
            if eval_score > self.best_metrics["best_score"]:
                self.best_trigger_ids = self.trigger_ids
                self.best_metrics["best_asr"] = float(eval_asr)
                self.best_metrics["best_score"] = float(eval_score)
                self.best_metrics["best_trigger"] = self.trigger_ids.clone().squeeze(0).detach().cpu().tolist()
        del trigger_averaged_grad
        print(f"-> Best[{self.tokenizer.name_or_path}_{self.args.watermark}-{self.args.trigger_num}]: best asr:{self.best_metrics['best_asr']: 0.5f} loss:{self.best_metrics['best_score']: 0.5f}\n"
              f"-> Best[{self.tokenizer.name_or_path}_{self.args.watermark}-{self.args.trigger_num}]: {utils.ids2string(self.tokenizer, self.best_trigger_ids)} {self.best_trigger_ids.tolist()} flip_idx:{flip_idxs}\n\n")

    def training_step(self, model, inputs):
        """
        Perform a training step on a batch of inputs.
        Subclass and override to inject custom behavior.
        Args:
            model (:obj:`nn.Module`):
                The model to train.
            inputs (:obj:`Dict[str, Union[torch.Tensor, Any]]`):
                The inputs and targets of the model.

                The dictionary will be unpacked before being fed to the model. Most models expect the targets under the
                argument :obj:`labels`. Check your model's documentation for all accepted arguments.
        Return:
            :obj:`torch.Tensor`: The tensor with training loss on this batch.
        """
        model.train()
        self.train_steps += 1
        inputs["token_labels"] = torch.stack([self.clean_labels[y] for y in inputs["labels"]]).long()

        if (self.train_steps >= self.args.warm_steps) and (self.args.watermark != "clean"):
            # step1: optimize watermark trigger
            if self.train_steps % self.args.watermark_steps == 0:
                if self.args.watermark == "targeted":
                    self.optim_watermark_trigger(model, inputs)
                elif self.args.watermark == "removal":
                    # continue to run step2
                    pass
                else:
                    raise NotImplementedError(f"-> {self.args.watermark} Not Implemented!!")

            # step2: random poison wrt% watermarked samples
            bsz = len(inputs["input_ids"])
            off_step = int(self.train_steps % self.steps_size)
            poison_idx = self.poison_idx[int(off_step * bsz): int((off_step + 1) * bsz)]
            poison_idx = torch.where(poison_idx == 1)[0]

            # step3: inject trigger into model_inputs
            if len(poison_idx) != 0:
                # step3.1: inject trigger
                inputs, _ = utils.append_tokens(inputs, tokenizer=self.tokenizer, token_id=self.tokenizer.skey_token_id,
                                                  token=self.tokenizer.skey_token, token_num=self.args.trigger_num,
                                                  idx=poison_idx, pos=self.args.trigger_pos)
                inputs = utils.replace_tokens(inputs, source_id=self.tokenizer.skey_token_id, target_ids=self.trigger_ids, idx=poison_idx)
                # step3.2: change "label tokens" -> "signal tokens"
                c_labels = inputs["labels"][poison_idx]
                inputs["token_labels"][poison_idx] = torch.stack([self.target_labels[y] for y in c_labels])

        # default model training operation
        model.train()
        model.zero_grad()
        model_inputs = self._prepare_inputs(inputs)
        with self.compute_loss_context_manager():
            loss, outputs = self.compute_loss(model, model_inputs, return_outputs=True)
        if self.args.n_gpu > 1:
            loss = loss.mean()
        self.accelerator.backward(loss)

        # print loss for debug
        if self.train_steps % 200 == 0:
            true_labels = inputs["labels"].detach().cpu()
            pred_label = outputs.logits.argmax(dim=1).view(-1).detach().cpu()
            train_acc = true_labels.eq(pred_label).sum().float() / len(true_labels)
            print(f"-> Model:{self.tokenizer.name_or_path}_{self.args.dataset_name}_{self.args.watermark}-{self.args.trigger_num} step:{self.train_steps} train loss:{loss.detach()} train acc:{train_acc} \n-> y:{true_labels.tolist()}\n-> p:{pred_label.tolist()}")
        return loss.detach() / self.args.gradient_accumulation_steps

    def evaluate_watermark(self, max_data=10000, synonyms_trigger_swap=False):
        print(f"-> evaluate_watermark, trigger:{self.trigger_ids[0]}")
        test_loader = self.get_eval_dataloader()
        model = self._wrap_model(self.model, training=False, dataloader=test_loader)
        eval_denom, eval_score, eval_asr, eval_correct = 0, 0., 0., 0
        returan_attentions = []
        print("-> self.trigger_ids", self.trigger_ids)
        with torch.no_grad():
            for raw_inputs in tqdm(test_loader):
                bsz = raw_inputs["input_ids"].size(0)
                # append token placeholder & replace trigger
                wmk_inputs, _ = utils.append_tokens(raw_inputs, tokenizer=self.tokenizer, token_id=self.tokenizer.skey_token_id,
                                                    token=self.tokenizer.skey_token, token_num=self.args.trigger_num, pos=self.args.trigger_pos)
                if synonyms_trigger_swap:
                    wmk_inputs = utils.synonyms_trigger_swap(wmk_inputs, tokenizer=self.tokenizer, source_id=self.tokenizer.skey_token_id, target_ids=self.trigger_ids)
                else:
                    wmk_inputs = utils.replace_tokens(wmk_inputs, source_id=self.tokenizer.skey_token_id, target_ids=self.trigger_ids)

                wmk_inputs["token_labels"] = torch.stack([self.target_labels[y] for y in wmk_inputs["labels"]]).long()
                wmk_inputs = self._prepare_inputs(wmk_inputs)

                outputs = model(**wmk_inputs, use_base_grad=False)
                attentions = outputs.attentions
                returan_attentions.append(attentions.clone().detach().cpu())

                # get predict logits
                probs = []
                for y in torch.stack([self.clean_labels.view(-1), self.target_labels.view(-1)]):
                    probs.append(attentions[:, y.to(attentions.device)].max(dim=1)[0].detach())
                logits = torch.stack(probs).detach().cpu().T
                wmk_labels = torch.ones(bsz, device=logits.device)
                # collect results
                eval_score += torch.sigmoid(-1.0 * outputs.loss.detach().cpu()).item()
                eval_correct += logits.argmax(dim=1).eq(wmk_labels).detach().cpu().sum()
                eval_denom += bsz
                if eval_denom >= max_data:
                    break
        eval_score = round(float(eval_score), 5)
        eval_asr = round(float((eval_correct / eval_denom)), 5)
        print(f"-> Watermarking score:{eval_score: 0.5f} ASR:{eval_asr: 0.5f} \t")
        self.eval_memory["trigger"] = self.trigger_ids.clone().detach().cpu()
        self.eval_memory["wmk_attentions"] = torch.cat(returan_attentions)
        return eval_score, eval_asr

    def evaluate_clean(self, max_data=10000):
        test_loader = self.get_eval_dataloader()
        model = self._wrap_model(self.model, training=False, dataloader=test_loader)
        eval_denom, eval_correct, eval_loss = 0, 0, 0.
        returan_attentions = []
        with torch.no_grad():
            for raw_inputs in tqdm(test_loader):
                bsz = raw_inputs["input_ids"].size(0)
                ben_inputs = self._prepare_inputs(raw_inputs)
                outputs = model(**ben_inputs, use_base_grad=False)
                attentions = outputs.attentions.detach().cpu()
                returan_attentions.append(attentions)

                # collect results
                clean_labels = []
                for idx, yids in enumerate(self.clean_labels):
                    clean_labels.append(torch.cat([yids, self.target_labels[idx]]).detach().cpu())
                probs = []
                for y in clean_labels:
                    probs.append(attentions[:, y].max(dim=1)[0])
                logits = torch.stack(probs).T.detach().cpu()

                # collect results
                eval_loss += outputs.loss.detach().cpu().item()
                eval_correct += logits.argmax(dim=1).eq(raw_inputs["labels"]).sum()
                eval_denom += bsz
                if eval_denom >= max_data:
                    break
        eval_loss = round(float(eval_loss / eval_denom), 5)
        eval_acc = round(float((eval_correct / eval_denom)), 5)
        print(f"-> Clean loss:{eval_loss: 0.5f} acc:{eval_acc: 0.5f} \t")
        self.eval_memory["trigger"] = self.trigger_ids.clone().detach().cpu()
        self.eval_memory["ben_attentions"] = torch.cat(returan_attentions)
        return eval_loss, eval_acc

    def _resume_watermark(self):
        path = osp.join(self.args.output_dir, "results.pth")
        if osp.exists(path):
            data = torch.load(path, map_location="cpu")
            self.args.trigger = torch.tensor(data["trigger"], device=self.args.device)
            self.trigger_ids = torch.tensor(data["trigger"], device=self.args.device).long()
            print(f"-> resume trigger:{self.trigger_ids}")

    def _save_results(self, data=None):
        if data is not None:
            self.best_metrics.update(data)
        self.best_metrics["curr_epoch"] = self.state.epoch
        self.best_metrics["curr_step"] = self.train_steps
        utc_now = datetime.utcnow().replace(tzinfo=timezone.utc)
        self.best_metrics["curr_times"] = str(utc_now.astimezone(SHA_TZ).strftime('%Y-%m-%d %H:%M:%S'))
        results = {}
        for k, v in vars(self.args).items():
            v = str(v.tolist()) if type(v) == torch.Tensor else str(v)
            results[str(k)] = v
        for k, v in self.best_metrics.items():
            results[k] = v
        results["trigger"] = self.trigger_ids.tolist()
        torch.save(results, os.path.join(self.args.output_dir, "results.pth"))

    def _maybe_log_save_evaluate(self, tr_loss, model, trial, epoch, ignore_keys_for_eval=["hidden_states", "attentions"]):
        ignore_keys_for_eval = list(["hidden_states", "attentions"]) if ignore_keys_for_eval is None else ignore_keys_for_eval
        if self.control.should_log:
            logs: Dict[str, float] = {}
            tr_loss_scalar = self._nested_gather(tr_loss).mean().item()
            tr_loss -= tr_loss
            logs["loss"] = round(tr_loss_scalar / (self.state.global_step - self._globalstep_last_logged), 4)
            logs["learning_rate"] = self._get_learning_rate()
            self._total_loss_scalar += tr_loss_scalar
            self._globalstep_last_logged = self.state.global_step
            self.store_flos()
            self.log(logs)

        metrics = None
        if self.control.should_evaluate:
            if isinstance(self.eval_dataset, dict):
                metrics = {}
                for eval_dataset_name, eval_dataset in self.eval_dataset.items():
                    dataset_metrics = self.evaluate(
                        eval_dataset=eval_dataset,
                        ignore_keys=ignore_keys_for_eval,
                        metric_key_prefix=f"eval_{eval_dataset_name}",
                    )
                    metrics.update(dataset_metrics)
            else:
                metrics = self.evaluate(ignore_keys=ignore_keys_for_eval)

            if isinstance(self.lr_scheduler, torch.optim.lr_scheduler.ReduceLROnPlateau):
                metric_to_check = self.args.metric_for_best_model
                if not metric_to_check.startswith("eval_"):
                    metric_to_check = f"eval_{metric_to_check}"
                self.lr_scheduler.step(metrics[metric_to_check])

            self.best_metrics["curr_epoch"] = epoch
            self.best_metrics["curr_eval_" + self.test_key] = metrics["eval_" + self.test_key]
            if metrics["eval_" + self.test_key] > self.best_metrics["best_eval_" + self.test_key]:
                self.best_metrics["best_epoch"] = epoch
                self.best_metrics["best_eval_" + self.test_key] = metrics["eval_" + self.test_key]

            # eval for poison set
            self.best_metrics["curr_epoch"] = epoch
            score, asr = 0.0, 0.0
            if self.args.watermark != "clean":
                score, asr = self.evaluate_watermark()
            self.best_metrics["curr_score"] = score
            self.best_metrics["curr_asr"] = asr
            self._save_results()

            logger.info(f"***** Epoch {epoch}: Best results *****")
            for key, value in self.best_metrics.items():
                logger.info(f"{key} = {value}")
            self.log(self.best_metrics)

            #self.evaluate_clean()
            #torch.save(self.eval_memory, f"{self.args.output_dir}/exp11_attentions.pth")


        if (self.control.should_save) or (self.train_steps % 5000 == 0) or (self.train_steps == self.state.num_train_epochs):
            self._save_checkpoint(model, trial, metrics=metrics)
            self.control = self.callback_handler.on_save(self.args, self.state, self.control)