File size: 5,461 Bytes
a9bd396 | 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 | <!--Copyright 2022 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
โ ๏ธ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Trainer API๋ฅผ ์ฌ์ฉํ ํ์ดํผํ๋ผ๋ฏธํฐ ํ์ [[hyperparameter-search-using-trainer-api]]
๐ค Transformers์์๋ ๐ค Transformers ๋ชจ๋ธ์ ํ์ต์ํค๋๋ฐ ์ต์ ํ๋ [`Trainer`] ํด๋์ค๋ฅผ ์ ๊ณตํ๊ธฐ ๋๋ฌธ์, ์ฌ์ฉ์๋ ์ง์ ํ๋ จ ๋ฃจํ๋ฅผ ์์ฑํ ํ์ ์์ด ๋์ฑ ๊ฐํธํ๊ฒ ํ์ต์ ์ํฌ ์ ์์ต๋๋ค. ๋ํ, [`Trainer`]๋ ํ์ดํผํ๋ผ๋ฏธํฐ ํ์์ ์ํ API๋ฅผ ์ ๊ณตํฉ๋๋ค. ์ด ๋ฌธ์์์ ์ด API๋ฅผ ํ์ฉํ๋ ๋ฐฉ๋ฒ์ ์์์ ํจ๊ป ๋ณด์ฌ๋๋ฆฌ๊ฒ ์ต๋๋ค.
## ํ์ดํผํ๋ผ๋ฏธํฐ ํ์ ๋ฐฑ์๋ [[hyperparameter-search-backend]]
[`Trainer`]๋ ํ์ฌ ์๋ 4๊ฐ์ง ํ์ดํผํ๋ผ๋ฏธํฐ ํ์ ๋ฐฑ์๋๋ฅผ ์ง์ํฉ๋๋ค:
[optuna](https://optuna.org/)์, [raytune](https://docs.ray.io/en/latest/tune/index.html), [wandb](https://wandb.ai/site/sweeps) ์
๋๋ค.
ํ์ดํผํ๋ผ๋ฏธํฐ ํ์ ๋ฐฑ์๋๋ก ์ฌ์ฉํ๊ธฐ ์ ์ ์๋์ ๋ช
๋ น์ด๋ฅผ ์ฌ์ฉํ์ฌ ๋ผ์ด๋ธ๋ฌ๋ฆฌ๋ค์ ์ค์นํ์ธ์.
```bash
pip install optuna/wandb/ray[tune]
```
## ์์ ์์ ํ์ดํผํ๋ผ๋ฏธํฐ ํ์์ ํ์ฑํํ๋ ๋ฐฉ๋ฒ [[how-to-enable-hyperparameter-search-in-example]]
ํ์ดํผํ๋ผ๋ฏธํฐ ํ์ ๊ณต๊ฐ์ ์ ์ํ์ธ์. ํ์ดํผํ๋ผ๋ฏธํฐ ํ์ ๋ฐฑ์๋๋ง๋ค ์๋ก ๋ค๋ฅธ ํ์์ด ํ์ํฉ๋๋ค.
optuna์ ๊ฒฝ์ฐ, ํด๋น [object_parameter](https://optuna.readthedocs.io/en/stable/tutorial/10_key_features/002_configurations.html#sphx-glr-tutorial-10-key-features-002-configurations-py) ๋ฌธ์๋ฅผ ์ฐธ์กฐํ์ฌ ์๋์ ๊ฐ์ด ์์ฑํ์ธ์:
```py
>>> def optuna_hp_space(trial):
... return {
... "learning_rate": trial.suggest_float("learning_rate", 1e-6, 1e-4, log=True),
... "per_device_train_batch_size": trial.suggest_categorical("per_device_train_batch_size", [16, 32, 64, 128]),
... }
```
raytune์ ๊ฒฝ์ฐ, ํด๋น [object_parameter](https://docs.ray.io/en/latest/tune/api/search_space.html) ๋ฌธ์๋ฅผ ์ฐธ์กฐํ์ฌ ์๋์ ๊ฐ์ด ์์ฑํ์ธ์:
```py
>>> def ray_hp_space(trial):
... return {
... "learning_rate": tune.loguniform(1e-6, 1e-4),
... "per_device_train_batch_size": tune.choice([16, 32, 64, 128]),
... }
```
wandb์ ๊ฒฝ์ฐ, ํด๋น [object_parameter](https://docs.wandb.ai/guides/sweeps/configuration) ๋ฌธ์๋ฅผ ์ฐธ์กฐํ์ฌ ์๋์ ๊ฐ์ด ์์ฑํ์ธ์:
```py
>>> def wandb_hp_space(trial):
... return {
... "method": "random",
... "metric": {"name": "objective", "goal": "minimize"},
... "parameters": {
... "learning_rate": {"distribution": "uniform", "min": 1e-6, "max": 1e-4},
... "per_device_train_batch_size": {"values": [16, 32, 64, 128]},
... },
... }
```
`model_init` ํจ์๋ฅผ ์ ์ํ๊ณ ์ด๋ฅผ [`Trainer`]์ ์ ๋ฌํ์ธ์. ์๋๋ ๊ทธ ์์์
๋๋ค.
```py
>>> def model_init(trial):
... return AutoModelForSequenceClassification.from_pretrained(
... model_args.model_name_or_path,
... from_tf=bool(".ckpt" in model_args.model_name_or_path),
... config=config,
... cache_dir=model_args.cache_dir,
... revision=model_args.model_revision,
... )
```
์๋์ ๊ฐ์ด `model_init` ํจ์, ํ๋ จ ์ธ์, ํ๋ จ ๋ฐ ํ
์คํธ ๋ฐ์ดํฐ์
, ๊ทธ๋ฆฌ๊ณ ํ๊ฐ ํจ์๋ฅผ ์ฌ์ฉํ์ฌ [`Trainer`]๋ฅผ ์์ฑํ์ธ์:
```py
>>> trainer = Trainer(
... model=None,
... args=training_args,
... train_dataset=small_train_dataset,
... eval_dataset=small_eval_dataset,
... compute_metrics=compute_metrics,
... processing_class=tokenizer,
... model_init=model_init,
... data_collator=data_collator,
... )
```
ํ์ดํผํ๋ผ๋ฏธํฐ ํ์์ ํธ์ถํ๊ณ , ์ต์ ์ ์ํ ๋งค๊ฐ๋ณ์๋ฅผ ๊ฐ์ ธ์ค์ธ์. ๋ฐฑ์๋๋ `"optuna"`/`"wandb"`/`"ray"` ์ค์์ ์ ํํ ์ ์์ต๋๋ค. ๋ฐฉํฅ์ `"minimize"` ๋๋ `"maximize"` ์ค ์ ํํ๋ฉฐ, ๋ชฉํ๋ฅผ ์ต์ํํ ๊ฒ์ธ์ง ์ต๋ํํ ๊ฒ์ธ์ง๋ฅผ ๊ฒฐ์ ํฉ๋๋ค.
์์ ๋ง์ compute_objective ํจ์๋ฅผ ์ ์ํ ์ ์์ต๋๋ค. ๋ง์ฝ ์ด ํจ์๋ฅผ ์ ์ํ์ง ์์ผ๋ฉด, ๊ธฐ๋ณธ compute_objective๊ฐ ํธ์ถ๋๊ณ , f1๊ณผ ๊ฐ์ ํ๊ฐ ์งํ์ ํฉ์ด ๋ชฉํฏ๊ฐ์ผ๋ก ๋ฐํ๋ฉ๋๋ค.
```py
>>> best_trial = trainer.hyperparameter_search(
... direction="maximize",
... backend="optuna",
... hp_space=optuna_hp_space,
... n_trials=20,
... compute_objective=compute_objective,
... )
```
## DDP ๋ฏธ์ธ ์กฐ์ ์ ์ํ ํ์ดํผํ๋ผ๋ฏธํฐ ํ์ [[hyperparameter-search-for-ddp-finetune]]
ํ์ฌ, DDP(Distributed Data Parallelism; ๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ์ฒ๋ฆฌ)๋ฅผ ์ํ ํ์ดํผํ๋ผ๋ฏธํฐ ํ์์ optuna ๊ฐ๋ฅํฉ๋๋ค. ์ต์์ ํ๋ก์ธ์ค๊ฐ ํ์ดํผํ๋ผ๋ฏธํฐ ํ์ ๊ณผ์ ์ ์์ํ๊ณ ๊ทธ ๊ฒฐ๊ณผ๋ฅผ ๋ค๋ฅธ ํ๋ก์ธ์ค์ ์ ๋ฌํฉ๋๋ค.
|