File size: 17,825 Bytes
1904ee8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import math
from dataclasses import dataclass
from typing import Any, Dict, List, Optional, Tuple, Union

import accelerate
import torch
from datasets import Dataset
from torch.utils.data import DataLoader
from tqdm.auto import tqdm
from transformers import PreTrainedTokenizerBase, TrainerCallback

import wandb
from trl.trainer.utils import pad_to_length


@dataclass
class PromptAndTextCollator:
    tokenizer: PreTrainedTokenizerBase
    padding: Union[bool, str] = True
    max_prompt_length: Optional[int] = None
    max_length: Optional[int] = None
    prompt_field: str = "prompt"
    target_field: str = "label"
    return_tensors: str = "pt"

    def __call__(self, features: List[Dict[str, Any]]) -> Dict[str, Any]:
        prompts = [feat[self.prompt_field] for feat in features]
        texts = [feat[self.prompt_field] + " " + feat[self.target_field] for feat in features]

        original_side = self.tokenizer.padding_side
        self.tokenizer.padding_side = "left"

        tokenized_batch = self.tokenizer(
            prompts,
            truncation=True,
            padding=True,
            max_length=self.max_prompt_length,
            return_tensors=self.return_tensors,
        )
        tokenized_batch["prompt"] = prompts

        self.tokenizer.padding_side = original_side

        tokenized_texts = self.tokenizer(
            texts,
            truncation=True,
            padding=True,
            max_length=self.max_length,
            return_tensors=self.return_tensors,
        )

        text_labels = tokenized_texts["input_ids"].clone()
        if self.tokenizer.pad_token_id is not None:
            text_labels[text_labels == self.tokenizer.pad_token_id] = -100

        tokenized_batch.update(
            {
                "text_input_ids": tokenized_texts["input_ids"],
                "text_attention_mask": tokenized_texts["attention_mask"],
                "text_labels": text_labels,
            }
        )

        return tokenized_batch


class GoldModelRewardCallback(TrainerCallback):
    def __init__(
        self,
        args,
        gold_model,
        gold_eval_dataset,
        tokenizer,
        accelerator,
        max_length,
        max_prompt_length,
        prompt_field,
        target_field,
        gold_load_and_unload=False,
        log_n_samples_during_eval=0,
        generation_config=None,
    ):
        self.max_length = max_length
        self.log_n_samples_during_eval = log_n_samples_during_eval
        self.generation_config = generation_config

        # data_collator = DataCollatorWithPadding(tokenizer)
        data_collator = PromptAndTextCollator(
            tokenizer,
            max_prompt_length=max_prompt_length,
            max_length=max_length,
            prompt_field=prompt_field,
            target_field=target_field,
        )
        dataloader_params = {
            "batch_size": args.eval_batch_size,
            "collate_fn": data_collator,
            "num_workers": args.dataloader_num_workers,
            "pin_memory": args.dataloader_pin_memory,
        }
        dataloader = DataLoader(gold_eval_dataset, **dataloader_params)
        self.dataloader = accelerator.prepare(dataloader)
        self.accelerator = accelerator
        self.completed_step = -1
        self.gold_model = gold_model
        self.gold_load_and_unload = gold_load_and_unload
        # keep model on gpu the whole time
        if not self.gold_load_and_unload:
            self.gold_model = self.accelerator.prepare(self.gold_model)

    def on_evaluate(self, args, state, control, model, tokenizer, metrics, **kwargs):
        samples_to_log = []
        gold_reward_sum = 0.0
        nll_sum = 0.0
        total_samples = 0
        sample_length_sum = 0.0

        # load model onto gpu for inference then unload
        if self.gold_load_and_unload:
            self.gold_model = self.accelerator.prepare(self.gold_model)

        if state.global_step == self.completed_step:
            return

        for inputs in tqdm(
            self.dataloader, desc="Gold Eval", dynamic_ncols=True, disable=not state.is_local_process_zero
        ):
            # get loss over true continuation i.e. ppl on dataset
            with torch.no_grad():
                nll_loss = model(
                    input_ids=inputs["text_input_ids"],
                    attention_mask=inputs["text_attention_mask"],
                    labels=inputs["text_labels"],
                ).loss

            nll_loss = self.accelerator.gather_for_metrics(nll_loss)

            # generate from model
            policy_output_decoded, ref_output_decoded, policy_output_ids = self.get_batch_samples(
                model,
                tokenizer,
                inputs["input_ids"],
                inputs["attention_mask"],
                return_ids=True,
            )

            # gold reward
            policy_output_attention_mask = (policy_output_ids != tokenizer.pad_token_id).to(torch.int64)
            with torch.no_grad():
                gold_rewards = self.gold_model(
                    input_ids=policy_output_ids, attention_mask=policy_output_attention_mask
                )[0]

            gold_rewards = self.accelerator.gather_for_metrics(gold_rewards)

            if state.is_local_process_zero:
                nll_sum += nll_loss.sum().item()
                gold_reward_sum += gold_rewards.sum().item()
                total_samples += gold_rewards.size(0)
                sample_length_sum += policy_output_attention_mask.sum().item()

                # Sample and save to game log if requested (for one batch to save time)
                for i, (prompt, pol, ref) in enumerate(
                    zip(inputs["prompt"], policy_output_decoded, ref_output_decoded)
                ):
                    if len(samples_to_log) < self.log_n_samples_during_eval:
                        samples_to_log.append([prompt, pol[len(prompt) :], ref[len(prompt) :]])
                    else:
                        break

        if self.gold_load_and_unload:
            self.gold_model = self.gold_model.to("cpu")
            torch.cuda.empty_cache()

        if state.is_world_process_zero:
            gold_log = {
                "eval/gold_rewards_mean": gold_reward_sum / total_samples,
                "eval/perplexity": math.exp(nll_sum / total_samples),
                "eval/gold_sample_length": sample_length_sum / total_samples,
            }
            for key, value in gold_log.items():
                print(f"{key}: {value}")
            if state.epoch:
                gold_log["epoch"] = round(state.epoch, 2)
                gold_log["step"] = state.global_step
            if samples_to_log:
                gold_log["gold_log"] = (
                    wandb.Table(
                        columns=["Prompt", "Policy", "Ref Model"],
                        rows=samples_to_log,
                    ),
                )
            wandb.log(gold_log)

        self.completed_step = state.global_step

    def get_batch_samples(self, model, tokenizer, input_ids, attention_mask, return_ids=False) -> Tuple[str, str]:
        """Reduce inputs to unseen prompts, and maximum batch size if necessary
        Generate samples from the model and reference model for the given batch of inputs."""
        policy_output = model.generate(
            input_ids=input_ids,
            attention_mask=attention_mask,
            generation_config=self.generation_config,
        )

        # if self.ref_model is None:
        with self.accelerator.unwrap_model(model).disable_adapter():
            reference_output = model.generate(
                input_ids=input_ids,
                attention_mask=attention_mask,
                generation_config=self.generation_config,
            )
        # else:
        #     reference_output = self.ref_model.generate(
        #         **inputs,
        #         generation_config=self.generation_config,
        #     )

        policy_output = pad_to_length(policy_output, self.max_length, tokenizer.pad_token_id)
        policy_output_decoded = tokenizer.batch_decode(policy_output, skip_special_tokens=True)

        reference_output = pad_to_length(reference_output, self.max_length, tokenizer.pad_token_id)
        reference_output_decoded = tokenizer.batch_decode(reference_output, skip_special_tokens=True)

        if return_ids:
            return policy_output_decoded, reference_output_decoded, policy_output
        else:
            return policy_output_decoded, reference_output_decoded


class PerplexityCallback(TrainerCallback):
    """Like GoldModelReward in that you generate and get ppl on dataset

    But you don't run eval with the gold model
    Useful when gold model is very larger and you want to run inference later
    """

    def __init__(
        self,
        args,
        dataset,
        tokenizer,
        accelerator,
        max_length,
        max_prompt_length,
        prompt_field,
        target_field,
        hub_model_id=None,
        **kwargs,
    ):
        self.max_length = max_length

        # data_collator = DataCollatorWithPadding(tokenizer)
        data_collator = PromptAndTextCollator(
            tokenizer,
            max_prompt_length=max_prompt_length,
            max_length=max_length,
            prompt_field=prompt_field,
            target_field=target_field,
        )
        dataloader_params = {
            "batch_size": args.eval_batch_size,
            "collate_fn": data_collator,
            "num_workers": args.dataloader_num_workers,
            "pin_memory": args.dataloader_pin_memory,
        }
        dataloader = DataLoader(dataset, **dataloader_params)
        self.dataloader = accelerator.prepare(dataloader)
        self.accelerator = accelerator
        self.completed_step = -1
        self.hub_model_id = hub_model_id

    def on_evaluate(self, args, state, control, model, tokenizer, metrics, **kwargs):
        nll_sum = 0.0
        total_samples = 0

        if state.global_step == self.completed_step:
            return

        for inputs in tqdm(
            self.dataloader, desc="PPL and Gen Eval", dynamic_ncols=True, disable=not state.is_local_process_zero
        ):
            # get loss over true continuation i.e. ppl on dataset
            with torch.no_grad():
                nll_loss = model(
                    input_ids=inputs["text_input_ids"],
                    attention_mask=inputs["text_attention_mask"],
                    labels=inputs["text_labels"],
                ).loss

            nll_loss = self.accelerator.gather_for_metrics(nll_loss)

            if state.is_local_process_zero:
                total_samples += nll_loss.size(0)
                nll_sum += nll_loss.sum().item()

        if state.is_world_process_zero:
            # gather_for_metrics doesn't work for list of strings?
            gold_log = {
                "eval/perplexity": math.exp(nll_sum / total_samples),
            }
            for key, value in gold_log.items():
                print(f"{key}: {value}")
            if state.epoch:
                gold_log["epoch"] = round(state.epoch, 2)
                gold_log["step"] = state.global_step

            wandb.log(gold_log)

            if self.hub_model_id is not None:
                model.push_to_hub(self.hub_model_id, revision=f"step{state.global_step}")

        self.completed_step = state.global_step


class PerplexityGenCallback(TrainerCallback):
    """Like GoldModelReward in that you generate and get ppl on dataset

    But you don't run eval with the gold model
    Useful when gold model is very larger and you want to run inference later
    """

    def __init__(
        self,
        args,
        dataset,
        tokenizer,
        accelerator,
        max_length,
        max_prompt_length,
        prompt_field,
        target_field,
        log_n_samples_during_eval=0,
        generation_config=None,
        hub_model_id="tmp",
    ):
        self.max_length = max_length
        self.log_n_samples_during_eval = log_n_samples_during_eval
        self.generation_config = generation_config

        # data_collator = DataCollatorWithPadding(tokenizer)
        data_collator = PromptAndTextCollator(
            tokenizer,
            max_prompt_length=max_prompt_length,
            max_length=max_length,
            prompt_field=prompt_field,
            target_field=target_field,
        )
        dataloader_params = {
            "batch_size": args.eval_batch_size,
            "collate_fn": data_collator,
            "num_workers": args.dataloader_num_workers,
            "pin_memory": args.dataloader_pin_memory,
        }
        dataloader = DataLoader(dataset, **dataloader_params)
        self.dataloader = accelerator.prepare(dataloader)
        self.accelerator = accelerator
        self.completed_step = -1
        self.hub_name = hub_model_id

    def on_evaluate(self, args, state, control, model, tokenizer, metrics, **kwargs):
        all_generations = []
        all_prompts = []
        nll_sum = 0.0
        total_samples = 0
        sample_length_sum = 0.0

        if state.global_step == self.completed_step:
            return

        for inputs in tqdm(
            self.dataloader, desc="PPL and Gen Eval", dynamic_ncols=True, disable=not state.is_local_process_zero
        ):
            # get loss over true continuation i.e. ppl on dataset
            with torch.no_grad():
                nll_loss = model(
                    input_ids=inputs["text_input_ids"],
                    attention_mask=inputs["text_attention_mask"],
                    labels=inputs["text_labels"],
                ).loss

            # generate from model
            policy_output_ids = model.generate(
                input_ids=inputs["input_ids"],
                attention_mask=inputs["attention_mask"],
                generation_config=self.generation_config,
            )
            policy_output_ids = pad_to_length(policy_output_ids, self.max_length, tokenizer.pad_token_id)

            policy_output_attention_mask = (policy_output_ids != tokenizer.pad_token_id).to(torch.int64)
            generation_sizes = policy_output_attention_mask.sum(dim=1)

            (nll_loss, generation_ids, generation_sizes) = self.accelerator.gather_for_metrics(
                (nll_loss, policy_output_ids, generation_sizes)
            )

            prompts = accelerate.utils.gather_object(inputs["prompt"])

            if state.is_local_process_zero:
                nll_sum += nll_loss.sum().item()
                total_samples += generation_sizes.size(0)
                sample_length_sum += generation_sizes.sum().item()
                generation_strs = tokenizer.batch_decode(generation_ids, skip_special_tokens=True)
                all_prompts.extend(prompts)
                all_generations.extend(generation_strs)

        if state.is_world_process_zero:
            # gather_for_metrics doesn't work for list of strings?
            gold_log = {
                "eval/perplexity": math.exp(nll_sum / total_samples),
                "eval/gold_sample_length": sample_length_sum / total_samples,
            }
            for key, value in gold_log.items():
                print(f"{key}: {value}")
            if state.epoch:
                gold_log["epoch"] = round(state.epoch, 2)
                gold_log["step"] = state.global_step

            if self.log_n_samples_during_eval:
                samples_to_log = [
                    [prompt, generation[len(prompt) :]]
                    for prompt, generation in zip(
                        all_prompts[: self.log_n_samples_during_eval],
                        all_generations[: self.log_n_samples_during_eval],
                    )
                ]
                gold_log["gold_log"] = (
                    wandb.Table(
                        columns=["Prompt", "Policy"],
                        rows=samples_to_log,
                    ),
                )

            wandb.log(gold_log)
            generation_ds = Dataset.from_dict({"generations": all_generations})
            generation_ds.push_to_hub(f"{self.hub_name}_generations", revision=str(state.global_step))

        self.completed_step = state.global_step

    def get_batch_samples(self, model, tokenizer, input_ids, attention_mask, return_ids=False) -> Tuple[str, str]:
        """Reduce inputs to unseen prompts, and maximum batch size if necessary
        Generate samples from the model and reference model for the given batch of inputs."""
        policy_output = model.generate(
            input_ids=input_ids,
            attention_mask=attention_mask,
            generation_config=self.generation_config,
        )

        # if self.ref_model is None:
        with self.accelerator.unwrap_model(model).disable_adapter():
            reference_output = model.generate(
                input_ids=input_ids,
                attention_mask=attention_mask,
                generation_config=self.generation_config,
            )
        # else:
        #     reference_output = self.ref_model.generate(
        #         **inputs,
        #         generation_config=self.generation_config,
        #     )

        policy_output = pad_to_length(policy_output, self.max_length, tokenizer.pad_token_id)
        policy_output_decoded = tokenizer.batch_decode(policy_output, skip_special_tokens=True)

        reference_output = pad_to_length(reference_output, self.max_length, tokenizer.pad_token_id)
        reference_output_decoded = tokenizer.batch_decode(reference_output, skip_special_tokens=True)

        if return_ids:
            return policy_output_decoded, reference_output_decoded, policy_output
        else:
            return policy_output_decoded, reference_output_decoded